PLoS Medicine

doi: 10.1371/journal.pmed.1001882pmid: 26418866

What Is VOI and How Does it Work? The VOI approach consists of a set of analytic tools that can be used to assess the value of acquiring additional evidence to inform a clinical (or public health) decision [8,9]. VOI quantifies the net benefit from the improvement of population health expected from additional research against the cost of implementation. Within this framework, the value of a study is the extent to which it reduces uncertainty on a particular topic, thus potentially reducing the errors in decision-making that would have been made, had less definitive evidence been used instead. Fig 1 schematically describes the VOI approach. The starting point is a “statistical model” that estimates all relevant unknown quantities. For example, if decision-making is about implementation of a cancer screening program, relevant parameters would include sensitivity and specificity of the screening test, cancer prevalence, and health benefits (relative to mortality, morbidity, and quality of life) and costs associated with the clinical pathway with and without the proposed program. These quantities are linked by complex relationships and informed by composite sources of evidence, from published studies or resources directly available to the researcher. Evidence is synthesized using a probabilistic model, often developed within a Bayesian framework [10]. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. Schematic diagram illustrating the different steps of the VOI approach. https://doi.org/10.1371/journal.pmed.1001882.g001 The results of the statistical model are fed into the “economic model,” which builds suitable population summaries for benefits and costs. In the example of cancer screening, we would compute the overall cost for both clinical pathways (with and without screening) by multiplying the average cost per person associated with the resources utilized (e.g., screening, diagnostic procedures, and treatments) by the expected number of users. Quality-Adjusted Life Years (QALYs), which combine changes in quantity and quality of life associated with an intervention, are often used as an average measure of clinical benefits. In our example, we would compute overall benefits of the program by multiplying QALYs gained through a successful diagnosis by the number of successful diagnoses, but we may also need to compute QALYs lost as a result of false-positive findings. The output of the economic model provides the basis for the “decision analysis,” which applies a set of rules to determine the best course of action given current evidence. Any intervention (screening or treatment) is considered cost-effective if its cost-per-QALY does not exceed a pre-specified willingness-to-pay threshold, for example, £20,000–30,000 in the UK [11]. All parameters in the statistical model are subject to uncertainty. “Uncertainty analysis” (technically a probabilistic sensitivity analysis) assesses how much the limited knowledge about the parameters can impact the results of the decision analysis and quantifies the expected economic return of obtaining new evidence before committing to a decision. In the VOI approach, this is the crucial step informing “research prioritization.” VOI analysis may, for example, point to additional research on test sensitivity as the highest priority to reduce errors in the decision of whether to implement a screening program. VOI analysis can be applied both to identify the crucial parameters within the appraisal of a single intervention and to compare different interventions. This kind of comparison is particularly important when allocating research funding across different topics and medical areas to maximize value for money and reduce waste. Challenges and Opportunities The comprehensiveness of VOI comes at the price of a complex analytic framework based on a number of modelling choices and assumptions that may influence the results. An important choice is the perspective from which costs are to be considered. For example, costs related to productivity losses are a crucial aspect when taking a societal perspective and dealing with diseases affecting a working-age population. Productivity costs will often show large uncertainty, and a societal perspective may well point to the need to acquire this information as a priority [12]. In contrast, data on productivity costs are irrelevant from a health system perspective, in which the only costs considered are those related to health care. The results of VOI analyses also depend on the choice of health metrics. Although QALYs are far from perfect and have been criticized for relying on strong theoretical assumptions about consistency in people’s preferences [13], they represent a convenient common currency that allows comparison of health gains between different interventions and across different diseases. VOI has classically been used for prioritization of research on health care screening and treatment interventions [14–18], but it can be equally useful in other areas, for example, risk prediction research. We are investigating its application within the Ageing Lungs in European Cohorts study (http://www.alecstudy.org), a European Commission–funded project aimed at developing a predictive risk score for chronic obstructive pulmonary disease (COPD). Although COPD is well known to be a smoking-related disease [19], there is evidence that other factors may contribute to a substantial proportion of its worldwide burden [20]. However, in addressing COPD in those who are not exposed to tobacco smoke, it is difficult to decide which other lifestyle, environmental, clinical, or genetic potential predictors should receive research priority. VOI can address this issue by evaluating the relative value for money of different research strategies according to their expected contribution in reducing the uncertainty in COPD prediction and thus minimizing prediction errors. VOI has been used in combination with other methods at different stages of the prioritization process [21,22], but the flexible Bayesian approach [10] also allows incorporation, directly within the VOI modelling framework, of the information on which other methods are based. Such information could include disease burden and variations in clinical practice, as well as experts’ opinions on parameters for which empirical evidence is limited or not available. Similarly, patients’ preferences could be formally incorporated in VOI analyses. For example, recognition of the importance of patients’ engagement in setting research agendas has motivated initiatives such as the “priority setting partnerships” created by the James Lind Alliance, in which patients and clinicians collaborate to identify research priorities (http://www.lindalliance.org). While VOI has been increasingly employed by regulatory agencies to inform decisions about adoption and reimbursement of treatments, its uptake by institutions prioritizing and commissioning research has been much more limited, although arguably the need to improve credibility and transparency of decisions is equally important for the latter [7]. Although VOI analyses are based on complex and computationally demanding modelling, statistical packages have been made freely available and fast methods recently developed to reduce the computational burden [23–25]. Dissemination outside the fields of health policy and health economics has also recently improved, and there are currently no real barriers to wider uptake of VOI in research prioritization.

doi: 10.1371/journal.pmed.1001871pmid: 26327350

Background Breast cancer is a leading malignancy affecting the female population worldwide. Most morbidity is caused by metastases that remain incurable to date. TGF-β1 has been identified as a key driving force behind metastatic breast cancer, with promising therapeutic implications. Methods and Findings Employing immunohistochemistry (IHC) analysis, we report, to our knowledge for the first time, that asporin is overexpressed in the stroma of most human breast cancers and is not expressed in normal breast tissue. In vitro, asporin is secreted by breast fibroblasts upon exposure to conditioned medium from some but not all human breast cancer cells. While hormone receptor (HR) positive cells cause strong asporin expression, triple-negative breast cancer (TNBC) cells suppress it. Further, our findings show that soluble IL-1β, secreted by TNBC cells, is responsible for inhibiting asporin in normal and cancer-associated fibroblasts. Using recombinant protein, as well as a synthetic peptide fragment, we demonstrate the ability of asporin to inhibit TGF-β1-mediated SMAD2 phosphorylation, epithelial to mesenchymal transition, and stemness in breast cancer cells. In two in vivo murine models of TNBC, we observed that tumors expressing asporin exhibit significantly reduced growth (2-fold; p = 0.01) and metastatic properties (3-fold; p = 0.045). A retrospective IHC study performed on human breast carcinoma (n = 180) demonstrates that asporin expression is lowest in TNBC and HER2+ tumors, while HR+ tumors have significantly higher asporin expression (4-fold; p = 0.001). Assessment of asporin expression and patient outcome (n = 60; 10-y follow-up) shows that low protein levels in the primary breast lesion significantly delineate patients with bad outcome regardless of the tumor HR status (area under the curve = 0.87; 95% CI 0.78–0.96; p = 0.0001). Survival analysis, based on gene expression (n = 375; 25-y follow-up), confirmed that low asporin levels are associated with a reduced likelihood of survival (hazard ratio = 0.58; 95% CI 0.37–0.91; p = 0.017). Although these data highlight the potential of asporin to serve as a prognostic marker, confirmation of the clinical value would require a prospective study on a much larger patient cohort. Conclusions Our data show that asporin is a stroma-derived inhibitor of TGF-β1 and a tumor suppressor in breast cancer. High asporin expression is significantly associated with less aggressive tumors, stratifying patients according to the clinical outcome. Future pre-clinical studies should consider options for increasing asporin expression in TNBC as a promising strategy for targeted therapy. Background Breast cancer is the most common cancer in women worldwide. Nearly 1.7 million new cases were diagnosed in 2012, and half a million women died from the disease. Breast cancer begins when cells in the breast that normally make milk (epithelial cells) acquire genetic changes that allow them to divide uncontrollably and to move around the body (metastasize). Uncontrolled cell division leads to the formation of a lump that can be detected by mammography (a breast X-ray) or by manual breast examination. Breast cancer is treated by surgical removal of the lump or, if the cancer has started to spread, by removal of the whole breast (mastectomy). After surgery, women often receive chemotherapy or radiotherapy to kill any remaining cancer cells, and women whose tumors express receptors for the female sex hormones estrogen and progesterone or for HER2, a growth factor receptor, are treated with drugs that block these receptors; estrogen, progesterone, and HER2 all control breast cell growth. Nowadays, the prognosis (outlook) for women living in high-income countries who develop breast cancer is generally good—nearly 90% of such women are still alive five years after diagnosis. Why Was This Study Done? The cells surrounding cancer cells—cancer-associated fibroblasts and other components of the stroma—support cancer growth and metastasis and are good targets for new cancer therapies. But, although there is mounting evidence that cancer cells actively adapt the stroma so that it produces various factors the tumor needs to grow and spread, very few molecules produced by the stroma that might serve as targets for drug development have been identified. Here, the researchers investigate whether a molecule called asporin might represent one such target. Asporin, which is highly expressed in the stroma of breast tumors, inhibits a growth factor called TGF-β1. TGF-β1 is involved in maintaining healthy joints, but is also a key molecule in the development of metastatic breast cancer. Most particularly, it modulates an important step in metastasis called the epithelial to mesenchymal transition and it regulates “stemness” in cancer cells. Stem cells are a special type of cell that can multiply indefinitely; tumor cells often look and behave very much like stem cells. What Did the Researchers Do and Find? Using a technique called immunohistochemistry, the researchers first showed that asporin is highly expressed in the stroma of most human breast cancers but not in normal breast tissue. Next, they showed that breast fibroblasts secrete asporin when exposed to conditioned medium from some human breast cancer cell lines (breast cancer cells adapted to grow continuously in the laboratory; conditioned medium is the solution in which cells have been grown). Specifically, conditioned medium from hormone receptor positive cells induced strong asporin expression by breast fibroblasts, whereas medium from breast cancer cells not expressing estrogen or progesterone receptors or HER2 receptors (triple-negative breast cancer cells) suppressed asporin expression. Other experiments showed that TGF-β1 secreted by breast cancer cells induces asporin expression in breast fibroblasts, and that asporin, in turn, inhibits TGF-β1-mediated induction of the epithelial to mesenchymal transition and stemness in breast cancer cells. Triple negative breast cancers appear to inhibit stromal expression of asporin at least in part via expression of the soluble signaling protein interleukin-1β. Notably, in mouse models of triple-negative breast cancer, tumors engineered to express asporin grew slower and metastasized less than tumors not expressing asporin. Finally, among women with breast cancer, asporin expression was low in triple-negative and HER2-positive tumors but significantly higher in hormone receptor positive tumors, and low asporin levels in primary breast lesions were associated with a reduced likelihood of survival independent of hormone receptor and HER2 expression. What Do These Findings Mean? These findings suggest that asporin is a stroma-derived inhibitor of TGF-β1 and a tumor suppressor in breast cancer. Importantly, they also provide preliminary evidence that high asporin expression is associated with less aggressive tumors (hormone receptor positive tumors), whereas low asporin expression is associated with more aggressive tumors (triple negative tumors and HER2-positive tumors). Thus, asporin expression might provide a new prognostic marker for breast cancer. However, before asporin can be used as a biomarker to predict outcomes in women with breast cancer and to identify those women in need of more aggressive treatment, these findings need to be confirmed in large prospective clinical studies. If these findings are confirmed, methods for increasing asporin expression in the stromal tissues of triple negative breast cancer could be a promising strategy for targeted therapy for this group of breast cancers, which currently have a poor prognosis. Additional Information This list of resources contains links that can be accessed when viewing the PDF on a device or via the online version of the article at http://dx.doi.org/10.1371/journal.pmed.1001871. The US National Cancer Institute provides comprehensive information about cancer (in English and Spanish), including detailed information for patients and professionals about breast cancer and an online booklet for patients Cancer Research UK, a not-for-profit organization, provides information about cancer; its detailed information about breast cancer includes sections on tests for hormone receptors and HER2, on treatments that target hormone receptors and treatments that target HER2, and on triple negative breast cancer Breastcancer.org is a not-for-profit organization that provides up-to-date information about breast cancer (in English and Spanish), including information on hormone receptor status, HER2 status, and triple negative breast cancer The UK National Health Service Choices website has information and personal stories about breast cancer; the not-for-profit organization Healthtalk.org also provides personal stories about dealing with breast cancer Introduction The tumor stroma, and especially cancer-associated fibroblasts (CAFs), is emerging as a key element of cancer growth and metastasis. CAFs supply cancer cells with a plethora of growth factors, energy substrates, and immune suppressors [1–3]. In most studies to date, the CAFs and other stromal cells have been observed to support tumor growth. The reverse is naturally less evident, as tumors inhibited by the stroma do not necessarily develop. Indeed, the inability of malignant cells to properly activate the host fibroblasts and program them to serve their needs would probably result in tumor failure [4–7]. However, it is far from clear how cancer cells perform this very early reprogramming of the stroma, what the anti-tumor responses of the stromal cells to these initial events are, and why, sometimes, the battle is lost against the tumor. Our previous studies, aiming to recognize accessible tumor proteins in human renal carcinoma [8] and colon [9], pancreas [10], and breast [11] adenocarcinomas, have consistently identified an overexpression of several small leucine-rich proteoglycans (SLRPs). In the current study, we aimed to explore asporin, a member of the class I SLRP family [12], which is at present insufficiently researched in cancer. Asporin is a secreted extracellular matrix protein that contains 380 amino acids. It was first identified in human cartilage, and its overexpression has been associated with osteoarthritis pathogenesis [13]. In normal tissues, asporin is found in articular cartilage, periodontal ligaments, the aorta, and the uterus [13,14], with no known protein isoforms reported to date. Like other SLRP family members, asporin contains a highly conserved (putative) pro-peptide sequence, has a series of leucine-rich repeats that are flanked by two cysteine residues in the C-terminal region, and has four cysteine residues that form disulfide bonds in the N-terminal domain [12]. Despite this similarity to other members of the SLRP family, in contrast to decorin and biglycan, asporin cannot be considered a typical proteoglycan because it lacks the consensus sequence necessary for glycosaminoglycan binding. Moreover, unlike other proteoglycans, asporin contains an aspartic acid repeat in its N-terminal region, polymorphisms of which have been associated with osteoarthritis [13,15]. SLRPs have been shown to be involved in several signaling pathways in which they bind to either ligands or receptors—such as bone morphogenic protein-4 (BMP-4), Wnt-I-induced secreted protein-1 (WISP-1), platelet-derived growth factor (PDGF), tumor necrosis factor-alpha (TNFα), and transforming growth factor-β1 (TGF-β1)—in the extracellular compartment [14,15]. Thus far, in tumor, high expression of asporin protein has been confirmed in pancreas [10], breast [11], prostate [16], and, recently, scirrhous gastric cancers [17]. We were intrigued by the low expression of asporin in normal human tissues, its high expression in breast carcinoma, and, particularly, its previously reported interaction with TGF-β1 in the context of osteoarthritis [15,18]. Kou et al. [18], and previously Kizawa et al. [15], demonstrated that asporin was able to bind to TGF-β1 and inhibit its ability to induce cartilage matrix gene expression. This regulation was mediated by direct binding to TGF-β1 of amino acids 159–205 of the asporin protein [18]. The reversibility of this interaction and, more importantly, its relevance in vivo to diseases such as cancer remain yet to be defined and explored using appropriate animal models. TGF-β1 is a paramount cytokine that is a potent modulator of immune evasion [19], angiogenesis [20], invasion [21], epithelial to mesenchymal transition (EMT) [22], metastasis [23], and stem cell biology [24,25]. In malignant tumors, TGF-β1 is secreted by both cancer cells and CAFs and has a demonstrated, and intriguing, dual role (pro- and anti-tumor) [26], suggesting the intervention of a more complex regulatory mechanism to modulate the spatiotemporal activity of this cytokine. Complementing the classic “seed and soil” theory, mounting evidence shows that cancer cells actively adapt the stroma (“soil”), enabling the colonization of distant organs from the primary tumor site [1,3]. Preventing the stroma from reprogramming into a tumor-supportive environment is therefore key to a successful anti-cancer therapy. However, to date, there are only a few well-characterized stromal molecules that could serve as a basis for effective drug development [1]. The present study contributes to the field by exploring the function of a new soluble stromal protein in breast cancer growth and progression. Given its previously described TGF-β1-inhibiting function in normal chondrocytes, we hypothesized that asporin may assume an important multifaceted tumor-suppressor function in breast cancer. Methods Patient Samples The ethical committee of the University Hospital Liège approved the use of human material in the current study. All samples were obtained from the institutional biobank of the University Hospital Liège, Belgium. According to Belgian law, patients were informed that the residual material from surgical procedures could be used for research purposes, and consent is presumed as long as the patient does not oppose (opt out). CAFs were isolated from tumors of three individual breast cancer patients (all female, mean age 55 y; tumors: estrogen receptor [ER] positive/progesterone receptor [PR] positive/HER2 negative, grade 2, Ki67+ [40%–60%]). For breast cancer, two collections of paraffin-embedded material and one set of freshly sampled tumors with adjacent non-tumoral tissue were used. The analysis of asporin expression in different subtypes of breast ductal adenocarcinoma was conducted retrospectively on a series of 180 patients (45 per subgroup). The correlation of asporin with the patient outcome was examined using an additional set of 60 patients, who had an average follow-up of 10 y. In this cohort, 30 cases had developed distant metastases (referred to as poor outcome), whereas the remainder showed no evidence of disease progression (good outcome) following the removal of the primary tumor. Other than as mentioned above, the same inclusion criteria were used for both cohorts. The inclusion criteria were as follows: (i) tumor lesion of 0.5–50 mm diameter, (ii) tumor lesion confirmed to be a breast adenocarcinoma of grade 2 and 3 after histology analysis, (iii) patient had no treatment before surgery, and (iv) patient had no metastasis at the time of surgery. Pathological characteristics for both patient groups are outlined in S1 and S2 Tables. Immunohistochemistry Formalin-fixed paraffin-embedded tissue sections were prepared from primary breast cancer lesions (see “Patient Samples” above) and from xenografted tumors (see “In Vivo Study” below). Tissue samples were sliced from paraffin blocks (5-μm sections), deparaffinated three times in xylene for 5 min and hydrated in a methanol gradient (100%, 95%, 70%, and 50%). Blocking of unspecific peroxidase activity was performed for 30 min with 3% H2O2 and 90% methanol. Citrate buffer (10 mM [pH 6]) was used for antigen retrieval. The following antibodies were used: asporin (rabbit anti-ASPN, dilution 1:150, Sigma-Aldrich, catalog no. HPA008435), IL-1β (dilution 1:80, Santa Cruz Biotechnology, catalog no. sc-7884), Ki67 (Ventana Medical Systems, catalog no. 790–4286), and vimentin (Ventana Medical Systems, catalog no. 760–2512). The incubation with the primary antibody was performed overnight at 4°C. Following this, the slides were washed with PBS for 10 min. The biotinylated secondary antibody was incubated initially for 30 min and subsequently with the avidin biotin complex kit (Dako, catalog no. X0590) for an additional 30 min. 3,3′-diaminobenzidine tetrachlorhydrate dihydrate (DAB) with 5% H2O2 was used for detection. The slides were counter-stained with hematoxylin. The quantification of protein expression was performed by two independent observers (average values are reported) and according to previously published methodology [27] with minor modifications to the scoring scale. Briefly, each immunohistochemistry (IHC) slide was assessed for the intensity of the staining of the tumor stroma using the following scale: 0 = no staining, 1 = weak, 2 = moderate, and 3 = strong. The tissue was further evaluated for the extent of positivity (percent positive stroma in the tumor) using the following scale: 1 = 0%–25%, 2 = 25%–50%, 3 = 50%–75%, and 4 = 75%–100%. The values obtained by each of the two scales were multiplied to yield a composite value called the IHC score. Pictures of representative fields were taken under a Leica DMRB light microscope. The details on statistical analysis are outlined below. Isolation of Primary Fibroblasts NBFs were derived from mammary reduction specimens, whereas CAFs were collected from ductal adenocarcinoma tissue material. Fibrous areas of normal breast tissue and breast tumors were cut into small pieces and digested for 18 h at 37°C in Dulbecco’s Modified Eagle’s Medium (DMEM, Lonza) supplemented with 10% FBS, 100 U/ml streptomycin, 100 μg/ml penicillin, 2.5 μg/ml Fungizone (Gibco BRL, Life Technologies), 150 U/ml hyaluronidase (Sigma-Aldrich), and 200 U/ml collagenase type III (Gibco BRL). The digested tissue was centrifuged at 100g and plated in T25 tissue culture flasks with DMEM and 20% FBS. Cell Culture Human epithelial breast cells MCF-7, T47D, ZR751, SKBR3, BT-474, MDA-MB-231, BT-549, and MCF-10A were obtained from ATCC. MDA-MB-468 cells were a kind gift from Dr. Sebastiano Andò (Laboratory of General Pathology, Department of Pharmacy and Health and Nutritional Sciences, University of Calabria), and EpRAS murine breast cancer cells were a kind gift of Dr. Sabine Macho-Maschler (Department of Molecular Genetics, Faculty of Veterinary Medicine, University of Vienna). MCF-7 cells were maintained in Eagle’s Minimum Essential Medium (Lonza) supplemented with 10% FBS, 1% Non-Essential Amino acid Solution (Lonza), and 2.5 mM L-glutamine (Lonza). T47D, ZR751, MDA-MB-231, and MDA-MB-468 cells were maintained in DMEM (Lonza) supplemented with 10% FBS and 2.5 mM L-glutamine (Lonza). SKBR3 cells were maintained in McCoy’s medium (Lonza) supplemented with 10% FBS. BT-474 cells were maintained in RPMI 1640 (Lonza) supplemented with 10% FBS and 1 mM sodium pyruvate, and BT-549 cells were maintained in RPMI 1640 (Lonza) supplemented with 10% FBS and 1 μg/ml bovine insulin (Sigma-Aldrich). EpRAS cells were maintained in DMEM supplemented with 10% FBS and 1 mM sodium pyruvate (Lonza). MCF-10A cells were maintained in DMEM containing 5% horse serum, 2.5 g/l glucose, 20 ng/ml EGF, 100 ng/ml cholera toxin, 0.01 ng/ml insulin, and 0.5 μg/ml hydrocortisone (all from Sigma-Aldrich). Normal breast fibroblasts (NBFs) and CAFs were maintained in DMEM supplemented with 10% FBS and 2.5 mM L-glutamine. Sub-confluent cultures (70%–90% confluence) of low passages (until passage 8) were utilized for all experiments. Conditioned medium (CM) from breast cancer cell lines was obtained following 48 h of incubation of 80% confluent cells in serum-free medium. For starvation, DMEM was used for all the cell lines. Cancer cell CM was collected, centrifuged for 5 min at 150g at room temperature, and then added to NBF and CAF monolayers (both cell types were pre-starved in serum-free medium for 24 h) for an additional 48 h. Following this, the NBF and CAF monolayers were washed two times with PBS and then either lysed with RIPA buffer for Western blot analysis or used for RNA extraction. Western Blot Analysis and ELISA The tissues were obtained immediately after surgery from patients undergoing breast cancer resection or from primary tumors from mice. The samples were frozen in liquid nitrogen and crushed into powder. CM samples were concentrated 10-fold using Amicon Ultra centrifugal filters (Millipore, catalog no. UFC500324). Total proteins from tissues or cells were extracted using RIPA buffer (50 mM Tris-HCl [pH 7.5], 150 mM NaCl, 1% Triton X-100, 0.5% sodium deoxycholate, 0.2% sodium dodecyl sulfate, and protease/phosphatase inhibitor cocktail; Thermo Scientific, catalog no. 78440). The protein content was determined using the Pierce BCA Protein Assay Kit (Thermo Scientific, catalog no. 23225). Twenty micrograms of proteins or concentrated CM was supplemented with Laemmli buffer (0.1% 2-mercaptoethanol, 0.0005% bromophenol blue, 10% glycerol, 2% SDS in 63 mM Tris-HCl [pH 6.8]) and were separated on 10% polyacrylamide denaturing gel and transferred to nitrocellulose membranes. The following antibodies were used: anti-ASPN pAb (dilution 1:500, Sigma-Aldrich, St. Louis, MO, USA, catalog no. HPA008435), anti-SMAD2/3 mAb (dilution 1:1,000, Cell Signaling Technology, catalog no. 8685), anti-phospho-SMAD2/3 pAb (dilution 1:500, Cell Signaling Technology, catalog no. 8828), anti-E-cadherin mAb (dilution 1:1,000, BD, catalog no. 610181), anti-human vimentin mAb (dilution 1:1,000, Sigma-Aldrich, catalog no. V6389), anti-mouse vimentin mAb (dilution 1:1,000, Cell Signaling Technology, catalog no. 5741), and anti-HSC70 mAb (dilution 1:30,000, Santa Cruz Biotechnology, catalog no. sc-7298). For ELISA assay, serum-free CM from cell lines was collected after 48 h of incubation, clarified by centrifugation, and then activated and processed using the TGF-β1 ELISA kit (R&D Systems, catalog no. DB100B) following the manufacturer’s instructions. Data were normalized according to the number of cells. Gene Expression Analysis Total RNA was isolated with the High Pure RNA Isolation Kit (Roche, catalog no. 11828665001). One microgram of total RNA was reverse-transcribed using the Transcriptor First Strand cDNA Synthesis Kit (Roche, catalog no.04897030001) according to the manufacturer’s instructions. The cDNAs (100 ng) were mixed with primers (0.5 μM), human UPL-probe system (0.2 μM) (Roche, catalog no. 04683633001), and 2× FastStart Universal Probe Master mix (Roche, catalog no. 04914058001) and analyzed in triplicate. Quantitative real-time PCR (qRT-PCR) was performed using the LightCycler 480 system (Roche) and the corresponding manufacturer software. The following cycling conditions were used: 95°C for 10 min then 40 cycles of 95°C (15 s) and 60°C (1 min). Sequences of asporin primers were as follows: forward 5′-GGTGGATAACTTCTACTTTTAGGAGGA-3′ and reverse 5′-AAGAAGGGTTT-GGCAGAGC-3′ and UPL probe #72. The relative gene expression levels were normalized using 18S rRNA content (Life Technologies, catalog no. 4310893E). Treatment with rTGF-β1, rASPN, ASPN Peptide Fragment, IL-1β, and IL-1RA Cells were starved in serum-free medium for 16 h and then treated for 15 min with a recombinant active form of TGF-β1 (Roche, catalog no. 11412272001) and/or recombinant human asporin (a kind gift of Targetome) and a synthetic peptide fragment of asporin protein (amino acids 159–205, H-NQLSEIPLNLPKSLAELRIHENKVKKIQKDTFKGMNALHVLEMSAN-OH) (Bachem). TGF-β1, recombinant asporin, and asporin peptide were dissolved in PBS and used at different concentrations according to the experimental setup (for respective concentrations and treatment times see figure captions). When used together, asporin peptide and recombinant TGF-β1 were pre-incubated for 1 h at 37°C before being added to the cells. Analysis of the EMT was conducted using Ras-transformed mammary epithelial cells (EpRAS). The EMT induction was performed as previously described [28]. Briefly, 5 × 104 cells/well were plated in a six-well plate and were grown in the presence or absence of TGF-β1 and/or asporin peptide. Treatments were repeated every day, following medium change, and the cells were cultured for 10 d. During this period the cells were re-plated every 3 d at 5 × 104 cells/well. For IL-1β experiments, starved NBFs and CAFs were incubated with CM or serum-free medium supplemented with 5 ng/ml TGF-β1 in the presence or absence of 0.1–0.5 ng/ml IL-1β (Peprotech, catalog no. 200-01B) for 48 h. IL-1β activity was blocked by pretreating starved NBFs with 40 ng/ml IL-1RA (Peprotech, catalog no. 200-01RA) for 1 h, followed by the addition of MDA-MB-231 CM to the pretreated NBFs for 48 h. Migration Assay EpRAS mouse breast cancer cells were pretreated for 10 d with TGF-β1 and/or asporin peptide (P159–205), as described above. At the end of this period, 1 × 105 cells were suspended in serum-free medium (0.1% BSA, 1% penicillin/streptomycin) and seeded into the upper part of a Transwell filter (diameter 6.5 mm, pore size 8 μm; Costar, catalog no. 3422). The lower compartment was filled with DMEM containing 1% pen/strep and 10% FBS. Following 16 h of incubation at 37°C, migrating cells were fixed and stained with Diff-Quick kit (Reagena, catalog no. 102164). Pictures of each insert were taken at 5× magnification, and migrating cells were counted using ImageJ software (US National Institutes of Health). Quantification of Stem Cells EpRAS cells. On the tenth day of treatment with TGF-β1 and/or asporin peptide, the cells were harvested with trypsin, and 2.5 × 105 cells were suspended in 25 μl of PBS and labeled with 1/100 (0.01 mg/ml) anti-CD24 (biotinylated; BioLegend, catalog no. 101803) for 1 h at 4°C. Following a wash, the cells were further labeled with 1/100 (0.01 mg/ml) anti-CD44 (PE-labeled; eBioscience, catalog no. 12–0441) and 1/1,000 streptavidin-FITC (Invitrogen, catalog no. SA100-02) for 1 h at 4°C. Following this, the cells were washed two times, and 1/50 7-aminoactinomycin D (7-AAD; BD-Pharmingen, catalog no. 51-68981E) was added for 10 min. The cell suspensions were analyzed using a FACSAria flow cytometer (BD Biosciences). Stem cells were quantified by evaluating the percentage of 7-AADneg, CD44high/CD24low cell population. MDA-MB-468 xenografts. Tumors were removed from NOD-SCID mice 7 wk post-implantation, minced, and digested in a solution of hyaluronidase (300 μg/ml) (Sigma-Aldrich, catalog no. H-3506) and collagenase I (1.75 mg/ml) (Sigma-Aldrich, catalog no. C0130) in HSSB (Life Technologies, catalog no. 14025–050) and incubated for 2 h at 37°C. 5 × 105 isolated cells were assayed for ALDH activity using the Aldefluor kit (Stemcell Technologies, catalog no. 01700), according to the supplied protocol. Human CD44 R-PE conjugate (Life Technologies, catalog no. MHCD4404) and human CD24 PE-Alexa Fluor 610 conjugate (Life Technologies, catalog no. MHCD2422) were incubated (both at 1/100 dilution) with the cell suspension for 1 h at 4°C. Following this, the cells were washed two times and 1/50 7-AAD was added for a further 10 min. The cell suspension was analyzed using a FACSAria flow cytometer (BD Bioscences). Stem cells were quantified by evaluating the percentage of cell populations characterized by two separate signatures: (i) 7-AADneg, ALDH+ and (ii) 7-AADneg, CD44high/CD24low. Stable Clones The MDA-MB-468 cells and NBFs were modified to express luciferase and asporin (MDA-MB-468-aspn; NBF-aspn) or green fluorescent protein (MDA-MB-468-ctrl; NBF-ctrl). Briefly, pLenti6-IRES-Luciferase plasmid was generated by cloning IRES (Internal Ribosome Entry Site) and firefly (Photinus pyralis) luciferase sequences into a lentiviral plasmid using the pLenti6/V5 Directional TOPO Cloning Kit (Invitrogen) in order to allow the expression of the luciferase gene under the control of CMV promoter. The 1,152 bp of the Homo sapiens asporin (ASPN, transcript variant 1) cDNA (NM_017680.4) was synthetized by GenScript and then cloned into pLenti6-IRES-Luciferase to obtain the pLenti6-ASPN-IRES-Luciferase plasmid for dual ASPN and luciferase expression. Lenti-X 293T cells (Clontech, 632180) were co-transfected with pSPAX2 (Addgene, plasmid #12260), a VSV-G-encoding vector, along with pLenti6-IRES-Luciferase or pLenti6-ASPN-IRES-Luciferase plasmids. 48 h and 72 h post-transfection, viral supernatants were collected, filtrated, and concentrated 100× by ultracentrifugation. The lentiviral vectors were then titrated with qPCR Lentivirus Titration Kit (ABM, LV900). Finally, the MDA-MB-468 cells and NBFs were transduced with 30 viral vectors per cell. After 48 h, positively transduced cells were selected with 10 μg/ml blasticidin (Invivogen, ant-bl-1). The cell culture supernatants were checked for the absence of replication-competent lentivirus before employing cells in vivo. In Vivo Study All experimental procedures used in the current work were performed in accordance with the ARRIVE ethical guidelines [29] and were reviewed and approved by the Institutional Animal Care and Ethics Committee of the University of Liège (Belgium). The experimentation adhered to the Guide for the Care and Use of Laboratory Animals prepared by the Institute of Laboratory Animal Resources of the National Research Council and published by National Academies Press, as well as to European and local legislation. NOD-SCID mice were purchased from Janvier Labs and housed in the animal facility of the University of Liège under standard conditions (12 h light/dark cycle, lights on at 7 a.m.). They were acclimated to the room 1 wk before the beginning of the experiment. Food and water were provided ad libitum. Tumor development was monitored at weekly intervals using in vivo imaging and caliper volume measurement (primary experimental outcome). For in vivo imaging, an intra-peritoneal injection of luciferin (Promega, catalog no. E1605) was given to the mice, and the signal was accrued using a Xenogen IVIS 200 imaging system (Caliper Life Sciences). Tumor volumes were calculated by acquiring the length (L), width (W), and height (H) of the xenografts and employing the formula V = (L/2) × (W/2) × (H/2) × π × (4/3). For the follow-up (size-matched experiments), the tumors were surgically removed at the target volume of 200–250 mm3 (reached in 6 to 12 wk, depending on the experimental condition). For the surgical removal of the primary tumor, mice were anaesthetized using 75 mg/kg of ketamine (Ceva) and 10 mg/kg of xylazine (Rompun, Bayer). Lung metastases were quantified in necropsy material using either IHC (vimentin staining) or Alu-PCR (see below), depending on the required sensitivity (secondary experimental outcome). Quantification of IHC was performed in serial paraffin sections by evaluating two parameters: (i) the frequency of metastatic foci (each individual cell or group of cells was counted as one deposit) and (ii) the size of the metastatic deposits (grouping them into three categories: <10 cells, 10–20 cells, and >20 cells). Statistical analysis was performed as described below. Fibroblast/MDA-MB-468 co-injection xenografts. 5 × 105 luciferase-positive NBF-aspn or-ctrl cells were mixed with 5 × 105 luciferase-positive MDA-MB-468 cells, suspended in cell culture medium, mixed (1:1) with growth-factor-reduced Matrigel (BDBiosciences, catalog no. 356230), and inoculated subcutaneously into the flanks of 5-wk-old NOD-SCID mice. Mice were randomly allocated to one of the two groups. Following tumor removal the mice were monitored for metastasis development for an additional 2 wk. Twenty mice were used in total. MDA-MB-468 xenografts. 2 × 106 million luciferase-positive MDA-MB-468-aspn or-ctrl cells were suspended in cell culture medium, mixed (1:1) with growth-factor-reduced Matrigel (BD Biosciences), and inoculated subcutaneously into the flanks of 5-wk-old NOD-SCID mice. Mice were randomly allocated to one of the two groups. After the resection of the primary tumor, the follow-up for metastasis development was conducted for an additional 3 wk. Eighty mice were used in total. qRT-PCR for Human Alu Sequences Genomic DNA was extracted from 10 mg of lung tissue using the High Pure PCR Template Preparation Kit (Roche, catalog no. 11796828001), according to the manufacturer’s instructions. A standard curve was generated by serially diluting MDA-MB-468 cells in 10 mg of normal mouse lung tissue, followed by DNA extraction. Next, 20 ng of DNA was mixed with primers against human Alu sequences (forward 5′-CATGGTGAAACCCCGTCTCTA-3′ and reverse 5′-GCCTCAGCCTCCCGAGTAG-3′) or with primers for human/mouse GAPDH, which was used as a normalizator (forward 5′-CAGCGACACCC-ACTCCTCCACCTT-3′ and reverse 5′-CATGAGGTCCACCACCCTGTTGCT-3′) [30]. All primers were used at 0.5 μM final concentration. Finally, to the DNA/primer mix, 2× FastStart Universal SYBR Green Master mix (Roche, catalog no. 04913850001) was added. The Alu sequences were amplified using the LightCycler 480 system (Roche) and the following cycling conditions: 95°C for 10 min followed by 40 cycles of 95°C (15 s) and 60°C (1 min). Survival Analysis and Gene Expression Patterns Kaplan-Meier survival curves were plotted with publicly deposited gene expression data (EGA and TCGA) originating from 375 untreated breast cancer patients and using Kaplan-Meier Plotter, which integrates statistical analysis (http://kmplot.com/analysis) [31]. All settings were left at default values except the following ones: gene symbol (ASPN; 219087_at), survival (OS), auto select best cutoff (on), and include systemically untreated patients (on). Expression of asporin at the mRNA level was quantified in 1,280 tumor samples from six different molecular subtypes using GOBO and publicly deposited gene expression datasets (http://co.bmc.lu.se/gobo/gsa.pl) [32]. Asporin mRNA expression was also quantified in tumors of different grades (n = 1,411) using the same tool. Finally, GOBO was also used to estimate IL-1β expression levels in different breast cancer cell lines (http://co.bmc.lu.se/gobo/gsa_cellines.pl). All statistics concerning Kaplan-Meier Plotter and GOBO data analysis were reported as calculated by the respective software and are detailed elsewhere [31,32]. Statistical Analysis Unless otherwise indicated, statistical analysis was performed using a two-sided, unpaired Student’s t-test, assuming equal variances using GraphPad Prism (version 5.01, GraphPad Software). The t-test was used because data followed a normal distribution (Shapiro-Wilk test, threshold 0.05). For IHC evaluation, box plots were generated using SigmaPlot (version 11.0, Systat). Testing of statistical significance was performed using a Mann-Whitney U test because the data did not follow the normal distribution (Shapiro-Wilk test, threshold 0.05). The receiver operating characteristic (ROC) curve was generated using GraphPad Prism. Patient Samples The ethical committee of the University Hospital Liège approved the use of human material in the current study. All samples were obtained from the institutional biobank of the University Hospital Liège, Belgium. According to Belgian law, patients were informed that the residual material from surgical procedures could be used for research purposes, and consent is presumed as long as the patient does not oppose (opt out). CAFs were isolated from tumors of three individual breast cancer patients (all female, mean age 55 y; tumors: estrogen receptor [ER] positive/progesterone receptor [PR] positive/HER2 negative, grade 2, Ki67+ [40%–60%]). For breast cancer, two collections of paraffin-embedded material and one set of freshly sampled tumors with adjacent non-tumoral tissue were used. The analysis of asporin expression in different subtypes of breast ductal adenocarcinoma was conducted retrospectively on a series of 180 patients (45 per subgroup). The correlation of asporin with the patient outcome was examined using an additional set of 60 patients, who had an average follow-up of 10 y. In this cohort, 30 cases had developed distant metastases (referred to as poor outcome), whereas the remainder showed no evidence of disease progression (good outcome) following the removal of the primary tumor. Other than as mentioned above, the same inclusion criteria were used for both cohorts. The inclusion criteria were as follows: (i) tumor lesion of 0.5–50 mm diameter, (ii) tumor lesion confirmed to be a breast adenocarcinoma of grade 2 and 3 after histology analysis, (iii) patient had no treatment before surgery, and (iv) patient had no metastasis at the time of surgery. Pathological characteristics for both patient groups are outlined in S1 and S2 Tables. Immunohistochemistry Formalin-fixed paraffin-embedded tissue sections were prepared from primary breast cancer lesions (see “Patient Samples” above) and from xenografted tumors (see “In Vivo Study” below). Tissue samples were sliced from paraffin blocks (5-μm sections), deparaffinated three times in xylene for 5 min and hydrated in a methanol gradient (100%, 95%, 70%, and 50%). Blocking of unspecific peroxidase activity was performed for 30 min with 3% H2O2 and 90% methanol. Citrate buffer (10 mM [pH 6]) was used for antigen retrieval. The following antibodies were used: asporin (rabbit anti-ASPN, dilution 1:150, Sigma-Aldrich, catalog no. HPA008435), IL-1β (dilution 1:80, Santa Cruz Biotechnology, catalog no. sc-7884), Ki67 (Ventana Medical Systems, catalog no. 790–4286), and vimentin (Ventana Medical Systems, catalog no. 760–2512). The incubation with the primary antibody was performed overnight at 4°C. Following this, the slides were washed with PBS for 10 min. The biotinylated secondary antibody was incubated initially for 30 min and subsequently with the avidin biotin complex kit (Dako, catalog no. X0590) for an additional 30 min. 3,3′-diaminobenzidine tetrachlorhydrate dihydrate (DAB) with 5% H2O2 was used for detection. The slides were counter-stained with hematoxylin. The quantification of protein expression was performed by two independent observers (average values are reported) and according to previously published methodology [27] with minor modifications to the scoring scale. Briefly, each immunohistochemistry (IHC) slide was assessed for the intensity of the staining of the tumor stroma using the following scale: 0 = no staining, 1 = weak, 2 = moderate, and 3 = strong. The tissue was further evaluated for the extent of positivity (percent positive stroma in the tumor) using the following scale: 1 = 0%–25%, 2 = 25%–50%, 3 = 50%–75%, and 4 = 75%–100%. The values obtained by each of the two scales were multiplied to yield a composite value called the IHC score. Pictures of representative fields were taken under a Leica DMRB light microscope. The details on statistical analysis are outlined below. Isolation of Primary Fibroblasts NBFs were derived from mammary reduction specimens, whereas CAFs were collected from ductal adenocarcinoma tissue material. Fibrous areas of normal breast tissue and breast tumors were cut into small pieces and digested for 18 h at 37°C in Dulbecco’s Modified Eagle’s Medium (DMEM, Lonza) supplemented with 10% FBS, 100 U/ml streptomycin, 100 μg/ml penicillin, 2.5 μg/ml Fungizone (Gibco BRL, Life Technologies), 150 U/ml hyaluronidase (Sigma-Aldrich), and 200 U/ml collagenase type III (Gibco BRL). The digested tissue was centrifuged at 100g and plated in T25 tissue culture flasks with DMEM and 20% FBS. Cell Culture Human epithelial breast cells MCF-7, T47D, ZR751, SKBR3, BT-474, MDA-MB-231, BT-549, and MCF-10A were obtained from ATCC. MDA-MB-468 cells were a kind gift from Dr. Sebastiano Andò (Laboratory of General Pathology, Department of Pharmacy and Health and Nutritional Sciences, University of Calabria), and EpRAS murine breast cancer cells were a kind gift of Dr. Sabine Macho-Maschler (Department of Molecular Genetics, Faculty of Veterinary Medicine, University of Vienna). MCF-7 cells were maintained in Eagle’s Minimum Essential Medium (Lonza) supplemented with 10% FBS, 1% Non-Essential Amino acid Solution (Lonza), and 2.5 mM L-glutamine (Lonza). T47D, ZR751, MDA-MB-231, and MDA-MB-468 cells were maintained in DMEM (Lonza) supplemented with 10% FBS and 2.5 mM L-glutamine (Lonza). SKBR3 cells were maintained in McCoy’s medium (Lonza) supplemented with 10% FBS. BT-474 cells were maintained in RPMI 1640 (Lonza) supplemented with 10% FBS and 1 mM sodium pyruvate, and BT-549 cells were maintained in RPMI 1640 (Lonza) supplemented with 10% FBS and 1 μg/ml bovine insulin (Sigma-Aldrich). EpRAS cells were maintained in DMEM supplemented with 10% FBS and 1 mM sodium pyruvate (Lonza). MCF-10A cells were maintained in DMEM containing 5% horse serum, 2.5 g/l glucose, 20 ng/ml EGF, 100 ng/ml cholera toxin, 0.01 ng/ml insulin, and 0.5 μg/ml hydrocortisone (all from Sigma-Aldrich). Normal breast fibroblasts (NBFs) and CAFs were maintained in DMEM supplemented with 10% FBS and 2.5 mM L-glutamine. Sub-confluent cultures (70%–90% confluence) of low passages (until passage 8) were utilized for all experiments. Conditioned medium (CM) from breast cancer cell lines was obtained following 48 h of incubation of 80% confluent cells in serum-free medium. For starvation, DMEM was used for all the cell lines. Cancer cell CM was collected, centrifuged for 5 min at 150g at room temperature, and then added to NBF and CAF monolayers (both cell types were pre-starved in serum-free medium for 24 h) for an additional 48 h. Following this, the NBF and CAF monolayers were washed two times with PBS and then either lysed with RIPA buffer for Western blot analysis or used for RNA extraction. Western Blot Analysis and ELISA The tissues were obtained immediately after surgery from patients undergoing breast cancer resection or from primary tumors from mice. The samples were frozen in liquid nitrogen and crushed into powder. CM samples were concentrated 10-fold using Amicon Ultra centrifugal filters (Millipore, catalog no. UFC500324). Total proteins from tissues or cells were extracted using RIPA buffer (50 mM Tris-HCl [pH 7.5], 150 mM NaCl, 1% Triton X-100, 0.5% sodium deoxycholate, 0.2% sodium dodecyl sulfate, and protease/phosphatase inhibitor cocktail; Thermo Scientific, catalog no. 78440). The protein content was determined using the Pierce BCA Protein Assay Kit (Thermo Scientific, catalog no. 23225). Twenty micrograms of proteins or concentrated CM was supplemented with Laemmli buffer (0.1% 2-mercaptoethanol, 0.0005% bromophenol blue, 10% glycerol, 2% SDS in 63 mM Tris-HCl [pH 6.8]) and were separated on 10% polyacrylamide denaturing gel and transferred to nitrocellulose membranes. The following antibodies were used: anti-ASPN pAb (dilution 1:500, Sigma-Aldrich, St. Louis, MO, USA, catalog no. HPA008435), anti-SMAD2/3 mAb (dilution 1:1,000, Cell Signaling Technology, catalog no. 8685), anti-phospho-SMAD2/3 pAb (dilution 1:500, Cell Signaling Technology, catalog no. 8828), anti-E-cadherin mAb (dilution 1:1,000, BD, catalog no. 610181), anti-human vimentin mAb (dilution 1:1,000, Sigma-Aldrich, catalog no. V6389), anti-mouse vimentin mAb (dilution 1:1,000, Cell Signaling Technology, catalog no. 5741), and anti-HSC70 mAb (dilution 1:30,000, Santa Cruz Biotechnology, catalog no. sc-7298). For ELISA assay, serum-free CM from cell lines was collected after 48 h of incubation, clarified by centrifugation, and then activated and processed using the TGF-β1 ELISA kit (R&D Systems, catalog no. DB100B) following the manufacturer’s instructions. Data were normalized according to the number of cells. Gene Expression Analysis Total RNA was isolated with the High Pure RNA Isolation Kit (Roche, catalog no. 11828665001). One microgram of total RNA was reverse-transcribed using the Transcriptor First Strand cDNA Synthesis Kit (Roche, catalog no.04897030001) according to the manufacturer’s instructions. The cDNAs (100 ng) were mixed with primers (0.5 μM), human UPL-probe system (0.2 μM) (Roche, catalog no. 04683633001), and 2× FastStart Universal Probe Master mix (Roche, catalog no. 04914058001) and analyzed in triplicate. Quantitative real-time PCR (qRT-PCR) was performed using the LightCycler 480 system (Roche) and the corresponding manufacturer software. The following cycling conditions were used: 95°C for 10 min then 40 cycles of 95°C (15 s) and 60°C (1 min). Sequences of asporin primers were as follows: forward 5′-GGTGGATAACTTCTACTTTTAGGAGGA-3′ and reverse 5′-AAGAAGGGTTT-GGCAGAGC-3′ and UPL probe #72. The relative gene expression levels were normalized using 18S rRNA content (Life Technologies, catalog no. 4310893E). Treatment with rTGF-β1, rASPN, ASPN Peptide Fragment, IL-1β, and IL-1RA Cells were starved in serum-free medium for 16 h and then treated for 15 min with a recombinant active form of TGF-β1 (Roche, catalog no. 11412272001) and/or recombinant human asporin (a kind gift of Targetome) and a synthetic peptide fragment of asporin protein (amino acids 159–205, H-NQLSEIPLNLPKSLAELRIHENKVKKIQKDTFKGMNALHVLEMSAN-OH) (Bachem). TGF-β1, recombinant asporin, and asporin peptide were dissolved in PBS and used at different concentrations according to the experimental setup (for respective concentrations and treatment times see figure captions). When used together, asporin peptide and recombinant TGF-β1 were pre-incubated for 1 h at 37°C before being added to the cells. Analysis of the EMT was conducted using Ras-transformed mammary epithelial cells (EpRAS). The EMT induction was performed as previously described [28]. Briefly, 5 × 104 cells/well were plated in a six-well plate and were grown in the presence or absence of TGF-β1 and/or asporin peptide. Treatments were repeated every day, following medium change, and the cells were cultured for 10 d. During this period the cells were re-plated every 3 d at 5 × 104 cells/well. For IL-1β experiments, starved NBFs and CAFs were incubated with CM or serum-free medium supplemented with 5 ng/ml TGF-β1 in the presence or absence of 0.1–0.5 ng/ml IL-1β (Peprotech, catalog no. 200-01B) for 48 h. IL-1β activity was blocked by pretreating starved NBFs with 40 ng/ml IL-1RA (Peprotech, catalog no. 200-01RA) for 1 h, followed by the addition of MDA-MB-231 CM to the pretreated NBFs for 48 h. Migration Assay EpRAS mouse breast cancer cells were pretreated for 10 d with TGF-β1 and/or asporin peptide (P159–205), as described above. At the end of this period, 1 × 105 cells were suspended in serum-free medium (0.1% BSA, 1% penicillin/streptomycin) and seeded into the upper part of a Transwell filter (diameter 6.5 mm, pore size 8 μm; Costar, catalog no. 3422). The lower compartment was filled with DMEM containing 1% pen/strep and 10% FBS. Following 16 h of incubation at 37°C, migrating cells were fixed and stained with Diff-Quick kit (Reagena, catalog no. 102164). Pictures of each insert were taken at 5× magnification, and migrating cells were counted using ImageJ software (US National Institutes of Health). Quantification of Stem Cells EpRAS cells. On the tenth day of treatment with TGF-β1 and/or asporin peptide, the cells were harvested with trypsin, and 2.5 × 105 cells were suspended in 25 μl of PBS and labeled with 1/100 (0.01 mg/ml) anti-CD24 (biotinylated; BioLegend, catalog no. 101803) for 1 h at 4°C. Following a wash, the cells were further labeled with 1/100 (0.01 mg/ml) anti-CD44 (PE-labeled; eBioscience, catalog no. 12–0441) and 1/1,000 streptavidin-FITC (Invitrogen, catalog no. SA100-02) for 1 h at 4°C. Following this, the cells were washed two times, and 1/50 7-aminoactinomycin D (7-AAD; BD-Pharmingen, catalog no. 51-68981E) was added for 10 min. The cell suspensions were analyzed using a FACSAria flow cytometer (BD Biosciences). Stem cells were quantified by evaluating the percentage of 7-AADneg, CD44high/CD24low cell population. MDA-MB-468 xenografts. Tumors were removed from NOD-SCID mice 7 wk post-implantation, minced, and digested in a solution of hyaluronidase (300 μg/ml) (Sigma-Aldrich, catalog no. H-3506) and collagenase I (1.75 mg/ml) (Sigma-Aldrich, catalog no. C0130) in HSSB (Life Technologies, catalog no. 14025–050) and incubated for 2 h at 37°C. 5 × 105 isolated cells were assayed for ALDH activity using the Aldefluor kit (Stemcell Technologies, catalog no. 01700), according to the supplied protocol. Human CD44 R-PE conjugate (Life Technologies, catalog no. MHCD4404) and human CD24 PE-Alexa Fluor 610 conjugate (Life Technologies, catalog no. MHCD2422) were incubated (both at 1/100 dilution) with the cell suspension for 1 h at 4°C. Following this, the cells were washed two times and 1/50 7-AAD was added for a further 10 min. The cell suspension was analyzed using a FACSAria flow cytometer (BD Bioscences). Stem cells were quantified by evaluating the percentage of cell populations characterized by two separate signatures: (i) 7-AADneg, ALDH+ and (ii) 7-AADneg, CD44high/CD24low. EpRAS cells. On the tenth day of treatment with TGF-β1 and/or asporin peptide, the cells were harvested with trypsin, and 2.5 × 105 cells were suspended in 25 μl of PBS and labeled with 1/100 (0.01 mg/ml) anti-CD24 (biotinylated; BioLegend, catalog no. 101803) for 1 h at 4°C. Following a wash, the cells were further labeled with 1/100 (0.01 mg/ml) anti-CD44 (PE-labeled; eBioscience, catalog no. 12–0441) and 1/1,000 streptavidin-FITC (Invitrogen, catalog no. SA100-02) for 1 h at 4°C. Following this, the cells were washed two times, and 1/50 7-aminoactinomycin D (7-AAD; BD-Pharmingen, catalog no. 51-68981E) was added for 10 min. The cell suspensions were analyzed using a FACSAria flow cytometer (BD Biosciences). Stem cells were quantified by evaluating the percentage of 7-AADneg, CD44high/CD24low cell population. MDA-MB-468 xenografts. Tumors were removed from NOD-SCID mice 7 wk post-implantation, minced, and digested in a solution of hyaluronidase (300 μg/ml) (Sigma-Aldrich, catalog no. H-3506) and collagenase I (1.75 mg/ml) (Sigma-Aldrich, catalog no. C0130) in HSSB (Life Technologies, catalog no. 14025–050) and incubated for 2 h at 37°C. 5 × 105 isolated cells were assayed for ALDH activity using the Aldefluor kit (Stemcell Technologies, catalog no. 01700), according to the supplied protocol. Human CD44 R-PE conjugate (Life Technologies, catalog no. MHCD4404) and human CD24 PE-Alexa Fluor 610 conjugate (Life Technologies, catalog no. MHCD2422) were incubated (both at 1/100 dilution) with the cell suspension for 1 h at 4°C. Following this, the cells were washed two times and 1/50 7-AAD was added for a further 10 min. The cell suspension was analyzed using a FACSAria flow cytometer (BD Bioscences). Stem cells were quantified by evaluating the percentage of cell populations characterized by two separate signatures: (i) 7-AADneg, ALDH+ and (ii) 7-AADneg, CD44high/CD24low. Stable Clones The MDA-MB-468 cells and NBFs were modified to express luciferase and asporin (MDA-MB-468-aspn; NBF-aspn) or green fluorescent protein (MDA-MB-468-ctrl; NBF-ctrl). Briefly, pLenti6-IRES-Luciferase plasmid was generated by cloning IRES (Internal Ribosome Entry Site) and firefly (Photinus pyralis) luciferase sequences into a lentiviral plasmid using the pLenti6/V5 Directional TOPO Cloning Kit (Invitrogen) in order to allow the expression of the luciferase gene under the control of CMV promoter. The 1,152 bp of the Homo sapiens asporin (ASPN, transcript variant 1) cDNA (NM_017680.4) was synthetized by GenScript and then cloned into pLenti6-IRES-Luciferase to obtain the pLenti6-ASPN-IRES-Luciferase plasmid for dual ASPN and luciferase expression. Lenti-X 293T cells (Clontech, 632180) were co-transfected with pSPAX2 (Addgene, plasmid #12260), a VSV-G-encoding vector, along with pLenti6-IRES-Luciferase or pLenti6-ASPN-IRES-Luciferase plasmids. 48 h and 72 h post-transfection, viral supernatants were collected, filtrated, and concentrated 100× by ultracentrifugation. The lentiviral vectors were then titrated with qPCR Lentivirus Titration Kit (ABM, LV900). Finally, the MDA-MB-468 cells and NBFs were transduced with 30 viral vectors per cell. After 48 h, positively transduced cells were selected with 10 μg/ml blasticidin (Invivogen, ant-bl-1). The cell culture supernatants were checked for the absence of replication-competent lentivirus before employing cells in vivo. In Vivo Study All experimental procedures used in the current work were performed in accordance with the ARRIVE ethical guidelines [29] and were reviewed and approved by the Institutional Animal Care and Ethics Committee of the University of Liège (Belgium). The experimentation adhered to the Guide for the Care and Use of Laboratory Animals prepared by the Institute of Laboratory Animal Resources of the National Research Council and published by National Academies Press, as well as to European and local legislation. NOD-SCID mice were purchased from Janvier Labs and housed in the animal facility of the University of Liège under standard conditions (12 h light/dark cycle, lights on at 7 a.m.). They were acclimated to the room 1 wk before the beginning of the experiment. Food and water were provided ad libitum. Tumor development was monitored at weekly intervals using in vivo imaging and caliper volume measurement (primary experimental outcome). For in vivo imaging, an intra-peritoneal injection of luciferin (Promega, catalog no. E1605) was given to the mice, and the signal was accrued using a Xenogen IVIS 200 imaging system (Caliper Life Sciences). Tumor volumes were calculated by acquiring the length (L), width (W), and height (H) of the xenografts and employing the formula V = (L/2) × (W/2) × (H/2) × π × (4/3). For the follow-up (size-matched experiments), the tumors were surgically removed at the target volume of 200–250 mm3 (reached in 6 to 12 wk, depending on the experimental condition). For the surgical removal of the primary tumor, mice were anaesthetized using 75 mg/kg of ketamine (Ceva) and 10 mg/kg of xylazine (Rompun, Bayer). Lung metastases were quantified in necropsy material using either IHC (vimentin staining) or Alu-PCR (see below), depending on the required sensitivity (secondary experimental outcome). Quantification of IHC was performed in serial paraffin sections by evaluating two parameters: (i) the frequency of metastatic foci (each individual cell or group of cells was counted as one deposit) and (ii) the size of the metastatic deposits (grouping them into three categories: <10 cells, 10–20 cells, and >20 cells). Statistical analysis was performed as described below. Fibroblast/MDA-MB-468 co-injection xenografts. 5 × 105 luciferase-positive NBF-aspn or-ctrl cells were mixed with 5 × 105 luciferase-positive MDA-MB-468 cells, suspended in cell culture medium, mixed (1:1) with growth-factor-reduced Matrigel (BDBiosciences, catalog no. 356230), and inoculated subcutaneously into the flanks of 5-wk-old NOD-SCID mice. Mice were randomly allocated to one of the two groups. Following tumor removal the mice were monitored for metastasis development for an additional 2 wk. Twenty mice were used in total. MDA-MB-468 xenografts. 2 × 106 million luciferase-positive MDA-MB-468-aspn or-ctrl cells were suspended in cell culture medium, mixed (1:1) with growth-factor-reduced Matrigel (BD Biosciences), and inoculated subcutaneously into the flanks of 5-wk-old NOD-SCID mice. Mice were randomly allocated to one of the two groups. After the resection of the primary tumor, the follow-up for metastasis development was conducted for an additional 3 wk. Eighty mice were used in total. Fibroblast/MDA-MB-468 co-injection xenografts. 5 × 105 luciferase-positive NBF-aspn or-ctrl cells were mixed with 5 × 105 luciferase-positive MDA-MB-468 cells, suspended in cell culture medium, mixed (1:1) with growth-factor-reduced Matrigel (BDBiosciences, catalog no. 356230), and inoculated subcutaneously into the flanks of 5-wk-old NOD-SCID mice. Mice were randomly allocated to one of the two groups. Following tumor removal the mice were monitored for metastasis development for an additional 2 wk. Twenty mice were used in total. MDA-MB-468 xenografts. 2 × 106 million luciferase-positive MDA-MB-468-aspn or-ctrl cells were suspended in cell culture medium, mixed (1:1) with growth-factor-reduced Matrigel (BD Biosciences), and inoculated subcutaneously into the flanks of 5-wk-old NOD-SCID mice. Mice were randomly allocated to one of the two groups. After the resection of the primary tumor, the follow-up for metastasis development was conducted for an additional 3 wk. Eighty mice were used in total. qRT-PCR for Human Alu Sequences Genomic DNA was extracted from 10 mg of lung tissue using the High Pure PCR Template Preparation Kit (Roche, catalog no. 11796828001), according to the manufacturer’s instructions. A standard curve was generated by serially diluting MDA-MB-468 cells in 10 mg of normal mouse lung tissue, followed by DNA extraction. Next, 20 ng of DNA was mixed with primers against human Alu sequences (forward 5′-CATGGTGAAACCCCGTCTCTA-3′ and reverse 5′-GCCTCAGCCTCCCGAGTAG-3′) or with primers for human/mouse GAPDH, which was used as a normalizator (forward 5′-CAGCGACACCC-ACTCCTCCACCTT-3′ and reverse 5′-CATGAGGTCCACCACCCTGTTGCT-3′) [30]. All primers were used at 0.5 μM final concentration. Finally, to the DNA/primer mix, 2× FastStart Universal SYBR Green Master mix (Roche, catalog no. 04913850001) was added. The Alu sequences were amplified using the LightCycler 480 system (Roche) and the following cycling conditions: 95°C for 10 min followed by 40 cycles of 95°C (15 s) and 60°C (1 min). Survival Analysis and Gene Expression Patterns Kaplan-Meier survival curves were plotted with publicly deposited gene expression data (EGA and TCGA) originating from 375 untreated breast cancer patients and using Kaplan-Meier Plotter, which integrates statistical analysis (http://kmplot.com/analysis) [31]. All settings were left at default values except the following ones: gene symbol (ASPN; 219087_at), survival (OS), auto select best cutoff (on), and include systemically untreated patients (on). Expression of asporin at the mRNA level was quantified in 1,280 tumor samples from six different molecular subtypes using GOBO and publicly deposited gene expression datasets (http://co.bmc.lu.se/gobo/gsa.pl) [32]. Asporin mRNA expression was also quantified in tumors of different grades (n = 1,411) using the same tool. Finally, GOBO was also used to estimate IL-1β expression levels in different breast cancer cell lines (http://co.bmc.lu.se/gobo/gsa_cellines.pl). All statistics concerning Kaplan-Meier Plotter and GOBO data analysis were reported as calculated by the respective software and are detailed elsewhere [31,32]. Statistical Analysis Unless otherwise indicated, statistical analysis was performed using a two-sided, unpaired Student’s t-test, assuming equal variances using GraphPad Prism (version 5.01, GraphPad Software). The t-test was used because data followed a normal distribution (Shapiro-Wilk test, threshold 0.05). For IHC evaluation, box plots were generated using SigmaPlot (version 11.0, Systat). Testing of statistical significance was performed using a Mann-Whitney U test because the data did not follow the normal distribution (Shapiro-Wilk test, threshold 0.05). The receiver operating characteristic (ROC) curve was generated using GraphPad Prism. Results Asporin Has Low or No Expression in Most Normal Tissues and Is Overexpressed in Breast Cancer We analyzed a publicly available gene expression repository (BioGPS, Scripps Research Institute) and compared the gene expression profiles in normal tissues of asporin and two other well-studied members of the SLRP family, biglycan and decorin (Fig 1A). The analysis showed that both biglycan and decorin are expressed in many normal tissues, whereas asporin expression was very low or not detected in most normal tissues, except the uterus. Next, we analyzed asporin expression using IHC in breast ductal adenocarcinoma (n = 30) as well as in adjacent non-tumoral tissue and normal breast tissue from breast reduction surgery (n = 10) (Fig 1B). Strong asporin expression was detectable in the stroma of the cancer lesions, with epithelial cancer cells being negative for asporin expression. Adjacent non-tumoral tissue showed a moderate positivity in the extracellular matrix and no positivity in non-tumoral epithelial cells. Healthy breast tissue was negative for asporin expression. Western blot analysis on fresh tissue extracts from matched tumoral and adjacent non-tumoral parts of the resected breast specimens confirmed our IHC observations (Fig 1C). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. Asporin is overexpressed in breast cancer tissues. (A) Tissue-specific pattern of mRNA expression of asporin (ASPN), biglycan (BGN), and decorin (DCN). Source: BioGPS (http://biogps.org). The data are presented as mean ± standard deviation (SD). (B) Representative IHC staining of asporin expression in ductal carcinoma and adjacent non-tumoral breast tissue (left panel) and normal breast tissue obtained from patients undergoing mammary reduction surgery (right panel). Asporin is almost exclusively expressed in breast cancer lesions, while a very low signal is detectable in the adjacent non-tumoral regions. Normal breast tissues are negative. Images of representative fields were taken at 100× and 400× magnification. (C) Western blot analysis of asporin expression in tumoral breast cancer tissues (T) and the adjacent normal counterpart (AdN) of six ductal adenocarcinoma patients. Ponceau red staining was used as loading control. https://doi.org/10.1371/journal.pmed.1001871.g001 Breast Fibroblasts Secrete Asporin after Their Activation by Cancer Cells Owing to the observations made above, in which asporin was found deposited in the extracellular matrix of the tumor, we sought to investigate which cells are responsible for producing the protein. As demonstrated in Fig 2A, none of the human breast cancer cell lines tested showed detectable asporin expression levels (both protein and mRNA). NBFs isolated from the mammary tissue of healthy individuals responded to the CM of several breast cancer cell lines by expressing asporin. The results indicated that tumorigenic and highly metastatic triple-negative breast cancer (TNBC) cells of the basal-like subtype (e.g., MDA-MB-231 and MDA-MB-468) [33–35] did not induce asporin expression in NBFs. This was different in noninvasive luminal-like hormone receptor (HR) positive cell lines (e.g., T47D and MCF-7) [33–35], which activated very high asporin expression in NBFs. Both observations were confirmed at the protein and gene expression levels. Similar experiments with immortalized, non-transformed mammary epithelial cells (MCF-10A) demonstrated that such cells are unable to induce asporin expression in NBFs (Fig 2B). We next sought to examine whether CAFs would express asporin following their isolation from cancer tissue and whether they would react similarly to NBFs when exposed to the CM of breast cancer cells. We isolated CAFs from three breast cancer patients and validated these as pure CAF populations, negative for cytokeratins and overexpressing α-smooth muscle actin (Fig 2C). As shown in Fig 2D, CAFs isolated from HR+ tumors expressed high levels of asporin, which they were able to maintain in vitro (for several weeks) without the need to be in contact with cancer cells. However, CAFs challenged with the CM of TNBC cells (MDA-MB-231) responded by lowering asporin expression. The CM of HR+ cells (MCF-7) was unable to further increase asporin levels in CAFs. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. Asporin is produced by breast fibroblasts in response to conditioned medium from breast cancer cells. (A) Western blot of total cell extracts (upper panel) and qRT-PCR analysis for asporin expression (lower panel) in breast cancer cell lines and NBFs incubated for 48 h with CM collected from a panel of breast cancer cells. (B) Western blot of total cell extracts (upper panel) and qRT-PCR analysis of asporin expression (lower panel) in non-cancerous epithelial breast cell line MCF-10A cells and NBFs incubated for 48 h with CM collected from MCF-10A. Fibroblasts treated with MCF-7 CM were used as the positive control for asporin expression induction. (C) Validation of NBFs and CAFs isolated from patient material. MCF-7 and MDA-MB-231 cells were used as epithelial controls. (D) Western blot analysis of asporin expression in total cell extracts of CAFs obtained from three different patients and treated with the CM of breast cancer cell lines. (A and B): The data are presented as mean ± SD. All panels: HSC70 was used as loading control; Western blots show representative data of three independent experiments. https://doi.org/10.1371/journal.pmed.1001871.g002 TGF-β1 Induces Asporin Expression in Fibroblasts while IL-1β Suppresses It In order to further understand which soluble factors could modulate asporin expression in fibroblasts, we first tested the cytokine TGF-β1. As shown in Fig 3A, TGF-β1 induced asporin expression in NBFs, both under basal conditions and in the presence of CM from MCF-7 cells. In sharp contrast to this, the presence of CM from MDA-MB-231 cells strongly inhibited the ability of TGF-β1 to induce asporin in NBFs. Despite this, our measurements showed that MDA-MB-231 cells, including other TNBC cells, are the highest producers of TGF-β1 among different breast cancer cells (Fig 3B). These findings raised the question of the specific mechanism by which TNBC cells inhibit asporin expression while still producing large quantities of TGF-β1. To further clarify this, we used publicly deposited gene expression data (GEO datasets GSE56265 [two replicates for each cell line] and GSE41445 [three replicates]) comparing the mRNA expression of MDA-MB-231 and MCF-7 cells [36,37]. The summary of the results is shown in Fig 3C. We identified 1,337 genes that were uniquely expressed in MDA-MB-231 cells. Considering that the CM of TNBC cells is able to convey the suppression of asporin expression without the need for cell-to-cell contact, we hypothesized that the effect must be mediated through a soluble protein. We therefore focused only on the genes whose products are known to be soluble proteins. The analysis highlighted a strong cluster of interleukins that were expressed in MDA-MB-231 cells. We next sought to verify which of the observed interleukin genes would discriminate between HR+ (luminal) and TNBC (basal) cells. This analysis was performed using GOBO [32], which compares the profiles of 51 breast cancer cell lines representing different molecular subtypes [35]. The results indicated that IL-1β could be of interest because it is highly expressed in cell lines of the basal-b subtype, which includes MDA-MB-231 cells (Fig 3C, right panel). Encouraged by these in silico findings, we sought to verify the expression of IL-1β in patient material from different breast tumor subtypes and compare this with asporin expression. IHC analysis of ductal adenocarcinoma cases (n = 20) demonstrated that IL-1β expression was significantly increased in TNBC compared to HR+ tumors (Fig 3D). In contrast to this, asporin expression in serial sections of the same tissues followed an inverse trend, with the highest expression in HR+ and the lowest in TNBC tumors. Next we tested whether recombinant human IL-1β could suppress basal, MCF-7-induced, or TGF-β1-induced asporin expression in NBFs. Supplementing CM with IL-1β readily blocked asporin expression in NBFs and CAFs (Fig 3E). We also examined whether IL-1RA, a naturally produced IL-1β inhibitor, would overcome the inhibition of asporin expression in NBFs treated with the CM of MDA-MB-231 cells. Indeed, the pre-incubation of NBFs with IL-1RA blocked the suppressive effect of CM from MDA-MB-231 cells on asporin expression in NBFs (Fig 3F). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. Asporin expression is induced by TGF-β1 and suppressed by IL-1β. (A) Western blot analysis of asporin expression in NBFs treated for 24 h with TGF-β1 with or without CM of MCF-7 and MDA-MB-231 breast cancer cells. (B) ELISA quantification of TGF-β1 levels secreted by human breast cancer cell lines during 48 h. (C) Differential in silico analysis of MCF-7 and MDA-MB-231 gene expression identifies a cluster of interleukins that are uniquely expressed in MDA-MB-231 cells (left panel). Analysis of IL-1β mRNA expression using GOBO (http://co.bmc.lu.se/gobo/gsa.pl) across a panel of breast cancer cell lines [35] subdivided into three subtypes (right panel). (D) Representative IHC analysis (upper panel) of IL-1β and asporin expression in a cohort of breast cancer patients, subdivided into two main subtypes (n = 20). Scoring and statistics (lower panel) were performed as outlined in the Methods section. (E) Inhibition of basal, MCF-7 CM-induced, and TGF-β1-induced asporin expression by IL-1β treatment of NBFs and CAFs. (F) Reversion of MDA-MB-231 CM inhibitory effect on asporin expression using IL-1β natural inhibitor IL-1RA. (B and D): The data are presented as mean ± SD. All panels: Western blots show representative data of three independent experiments. https://doi.org/10.1371/journal.pmed.1001871.g003 Asporin Inhibits TGF-β1-Mediated SMAD2 Activation and Induction of the Epithelial to Mesenchymal Transition Previous studies in chondrocytes [18] identified that the peptide region of asporin (residues 159 to 205) is relevant for the interaction with TGF-β1. We therefore tested whether recombinant asporin and the synthetically produced peptide fragment were able to inhibit TGF-β1-mediated activation of SMAD2 in breast cancer cells (Fig 4A and 4B). SMAD2 was efficiently phosphorylated upon treatment of MDA-MB-468 cells with TGF-β1. The TGF-β1 activity was inhibited when the cytokine was pre-incubated for 1 h at 37°C with increasing doses of recombinant asporin (Fig 4A). Cells treated with recombinant asporin alone showed no modulation of SMAD2 phosphorylation. Analogously to the effects observed with the recombinant protein, the peptide fragment of asporin showed an inhibitory effect on TGF-β1-induced SMAD2 phosphorylation (Fig 4B). Further data showed that parallel treatment of cancer cells with TGF-β1 and asporin peptide, without prior pre-incubation, failed to inhibit SMAD2 phosphorylation. These results are in agreement with previous findings in normal chondrocytes showing that asporin directly binds to TGF-β1, rather than acting as a competitive inhibitor for TGF-β1 receptor [28,38]. To functionally test the ability of asporin to interfere with TGF-β1-induced processes, we employed EpRAS, a murine mammary cancer cell line. EpRAS cells have an established responsiveness to TGF-β1, especially with respect to EMT and migration [28,38]. Similarly to the effects observed in MDA-MB-468 cells, asporin peptide was able to block TGF-β1-mediated phosphorylation of SMAD2 in EpRAS cells (Fig 4C). EpRAS cells are known for undergoing EMT upon stimulation with TGF-β1. Apart from the phenotypic appearance, the EMT switch can be readily observed through the up-regulation of vimentin (VIM) (Fig 4D). TGF-β1-mediated induction of VIM was weaker when TGF-β1 was pre-incubated with asporin peptide (Fig 4D). We further sought to investigate the ability of asporin to interfere with TGF-β1-induced migration of EpRAS cells (Fig 4E). Treatment of EpRAS cells with TGF-β1 induced significant cell migration, while pre-incubation of TGF-β1 with the asporin peptide significantly curbed this effect. The EMT switch has been described as relevant for the acquisition of the cancer stem cell (CSC) phenotype [24]. Therefore, we tested whether TGF-β1-induced EMT would increase the CSC population in EpRAS cells and whether this effect could be inhibited by asporin peptide (Fig 4F). Indeed, TGF-β1 treatment induced an increase of CSCs in EpRAS cells, as evidenced by the established CD44high/CD24low breast cancer stemness signature [39]. The TGF-β1-induced increase of the CSC population was significantly inhibited when asporin peptide was pre-incubated with TGF-β1. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 4. Asporin binds to TGF-β1 and inhibits its downstream signaling and function. (A) Western blot analysis of phospho-SMAD2 (p-SMAD2) and SMAD2 total protein extracts from MDA-MB-468 breast cancer cells treated for 15 min with TGF-β1 and/or human recombinant asporin (Rec. ASPN). (B) Western blot analysis of p-SMAD2 in total protein extracts from MDA-MB-468 breast cancer cells treated with TGF-β1 and/or asporin peptide corresponding to the 159–205 amino acid region (ASPNpep.). (C) Western blot analysis of p-SMAD2 and SMAD2 in total protein extracts from EpRAS cells treated for 15 min with TGF-β1 (5 ng/ml) and/or asporin peptide. (D) EMT induction in EpRAS cells in the presence of TGF-β1 and/or asporin peptide. EMT was monitored both at the phenotype level (upper panel) and using Western blot evaluation of VIM expression in total protein extracts from EpRAS cells (lower panel). (A–D): HSC70 was used as loading control. (E) Transwell migration assay of EpRAS cells pretreated with TGF-β1 (5 ng/ml) and/or asporin peptide (10 μg/ml). (F) Quantification of the CSC population in EpRAS cells following TGF-β1 and/or asporin peptide treatment. (E and F): The data are presented as mean ± SD. All panels: statistical significance was calculated using the Student’s t-test (as described in the Methods section). Western blots show representative data of three independent experiments. https://doi.org/10.1371/journal.pmed.1001871.g004 Induction of Asporin Expression in Triple-Negative Breast Tumors Reduces Growth and Metastasis In Vivo In order to evaluate the impact of asporin expression on TNBC growth and progression, we co-injected MDA-MB-468 cells and NBFs stably overexpressing asporin (NBF-aspn) subcutaneously in NOD-SCID mice (Fig 5A). The control group consisted of mice xenografted with MDA-MB-468 cells with NBFs overexpressing GFP (NBF-ctrl). The growth of the tumors was followed weekly, and the results indicated that the control tumors reached volumes of ~200 mm3 in 6 wk, whereas the asporin-overexpressing tumors grew significantly slower (day 42 post-engraftment: 78.9 mm3 smaller than control; 95% CI 24.3–123.4; p = 0.007) and required 8 wk to reach the same volume (Fig 5B and 5C). Following the resection of the primary tumor, the ability of cancer cells to colonize the lungs was assessed in a follow-up experiment. Two weeks after the removal of the primary tumors, mice were sacrificed and the lungs were collected. Assessment of metastasis formation in the lungs was performed using the Alu-PCR technique. The results evidenced a 3-fold, significant (p = 0.002) increase in the number of cancer cells present in the lung tissue of the control mice compared to the mice with asporin-overexpressing NBFs (Fig 5D). Previously published studies using xenografts based on co-injection of fibroblasts and epithelial cancer cells showed that human fibroblasts are rapidly displaced by murine fibroblasts in vivo [40–42]. We thus sought to verify whether asporin expression remained constant in the tumor during the present experiments. Western blot analysis of tumor tissue extracts showed that asporin expression decreased starting from week 4, reaching low levels at week 6 (Fig 5E). This result suggested the ongoing replacement of human xenografted fibroblasts by murine counterparts. This diminishing asporin expression may underestimate asporin’s effect on tumor growth and metastasis in vivo. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 5. Co-injection of cancer cells and fibroblasts overexpressing asporin reduces primary breast cancer tumor growth and lung metastasis formation in vivo. (A) Western blot analysis of asporin expression in CM of MDA-MB-468 cells and in NBF stable clones used for subcutaneous injection in mice. Ponceau red is shown as loading control. (B) Bioluminescence imaging of control and asporin-expressing xenografts at day 28 after tumor engraftment. The color scale indicates the fluorescent intensity. (C) The volume (in cubic millimeters) of primary tumors measured weekly (from day 7 onwards). The data are presented as mean ± standard error of the mean (SEM) (n = 10 for each group). Statistical significance was calculated using Student’s t-test (**0.01 < p < 0.001; ***0.001 < p < 0.0001). (D) Human-specific Alu-PCR performed on genomic DNA isolated from dissected lungs was used to detect human cancer cells. The data are presented as mean ± SD. (E) Western blot analysis of asporin expression in mice primary tumors monitored for several weeks. HSC70 was used as loading control. https://doi.org/10.1371/journal.pmed.1001871.g005 Therefore, we next sought to engraft asporin-overexpressing cancer cells that would maintain constant asporin expression in the tumor. This was performed with stably transduced asporin-expressing MDA-MB-468 cells (Fig 6A). The control and asporin-expressing MDA-MB-468 cells were implanted subcutaneously in NOD-SCID mice. Primary tumor growth was monitored weekly. The results indicated that asporin-expressing tumors were significantly smaller, reaching up to 2-fold lower volumes at 7 wk post-engraftment (day 49 post-engraftment: 124.1 mm3 smaller than control; 95% CI 75.2–180.4; p = 0.001) (Fig 6B). Histological evaluation demonstrated invasive control tumors developing towards the muscle layers, whereas this was not observed in asporin-expressing counterparts (Fig 6C). Further analysis of asporin-expressing tumors evidenced extensive zones of tumor necrosis in the central areas (Fig 6C), as well as numerous cells with condensed chromatin. In the control conditions necrosis was less pronounced, whereas transparent chromatin staining and the presence of nucleoli further characterized tumor cells. The latter suggested a higher proliferation rate in control tumors. The assessment of tumor proliferation based on Ki67 staining showed stronger and more frequent nuclear positivity in the control tumors in comparison to the asporin-expressing counterparts (Fig 6C). IHC evaluation of asporin expression in the experimental tumors evidenced the expected asporin overexpression. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 6. Asporin reduces primary breast cancer tumor growth and lung metastasis formation in vivo. (A) Western blot analysis of asporin expression in the CM of MDA-MB-468 stable clones expressing asporin, used for subcutaneous injection in mice. Ponceau red is shown as loading control. (B) Bioluminescence imaging of control and asporin-expressing xenografts at day 28 after tumor engraftment (left panel). The color scale indicates the fluorescent intensity. The mean (± SEM) volume (in cubic millimeters) of primary tumors measured weekly (from day 14 onwards) for the time-matched cohort is also shown (n = 20 for each group) (right panel). Statistical significance was calculated using Student’s t-test (**0.01 < p < 0.001; ***0.001 < p < 0.0001). (C) Representative hematoxylin and eosin (H&E), asporin, and Ki67 IHC staining in MDA-MB-468 xenografts collected 7 wk post-engraftment. Control xenografts consistently displayed an invasion in the muscle layer (M). An extended necrotic (N) area was present in the peri-tumoral zone of MDA-MB-468-aspn mice tumors. (D) Quantification of the stem cell population in xenografted tumors expressing asporin using ALDH+ and CD44high/CD24low stemness markers (7 wk post-engraftment). (E) Post-operative follow-up of mice that had primary tumors removed at the same time (time-matched). (F) Mean (± SEM) volume (in cubic millimeters) of primary tumors measured weekly for the size-matched cohort (n = 20 for each group). (G) Post-operative follow-up of mice that had primary tumors removed at the same volume (size-matched). (E and G): IHC evaluation of vimentin in lung necropsies and quantification of metastatic deposits. All images of representative fields were taken at 40×, 100×, and 400× magnification. (D, E, and G): The data are presented as mean ± SD. Statistical significance was calculated using Student’s t-test. https://doi.org/10.1371/journal.pmed.1001871.g006 Considering that asporin blocks TGF-β1 activity and EMT, processes known to enrich stem cells, we hypothesized that the abundance of stem cells would be different in these two experimental conditions. The evaluation of tumor stemness, using two different and independent signatures, showed that asporin-expressing tumors had a significantly lower percentage of stem cells (Fig 6D). As CSCs are essential for tumor survival and metastasis, we sought to evaluate tumor dissemination in mice following tumor resection. For this purpose the animal experiments were divided into two separate cohorts: (i) time-matched and (ii) size-matched. For the time-matched cohort, the tumorectomy was performed at week 7. For the size-matched group, the tumorectomy was conducted at week 9 for control and at week 12 for asporin-expressing tumors (Fig 6F). In both instances the mice were allowed to recover and were observed for axial lymph node and lung metastases during an additional period of 3 wk. As indicated by the time-matched data, control mice developed overt lung metastases, whereas this was not observed in the asporin condition (Fig 6E). Control animals consistently developed frequent and large deposits in the lungs. Animals carrying asporin-expressing tumors also showed lung metastases; however, they were less frequent and of smaller size. The notion that asporin is interfering with the process of metastasis was further confirmed in the size-matched experiments. In this cohort the tumor growth was followed for a longer period of time (control mice 9 wk, asporin 12 wk), highlighting an overall 3-wk delay of tumor growth in asporin-expressing mice. The results quantifying metastases 3 wk post-tumorectomy were similar to those of the time-matched condition (Fig 6G). Collectively, the data obtained with both xenograft models suggested that asporin expression inhibits tumor growth as well as metastatic progression. High Asporin Levels Delineate Breast Cancer Patients with Good Clinical Outcome Considering the in vitro and in vivo data, we expanded our observations using IHC to 180 breast cancer patients, subdivided in four categories with 45 cases each: (i) ER−/PR−/HER2− (triple-negative), (ii) ER+/PR+/HER2+ (triple-positive), (iii) ER+/PR+/HER2− (HR+), and (iv) ER−/PR−/HER2+ (HER2+) (Fig 7A; S1 Table). In our cohort, patients of all subgroups had similar age and showed similar tumor size. Triple-negative and HER2+ cases had higher tumor grade (Bloom 3 versus 2) and a stronger percentage of proliferating cells (Ki67+ cells: ~43% versus ~17%) than the other two subgroups. The frequency of metastasis was highest in TNBC patients (22%), followed by HER2+ and triple-positive breast cancer patients, who displayed similar frequencies (~9%). The IHC results showed that HR+ tumors had a high asporin expression, which was significantly elevated (up to 4-fold) in comparison to TNBC and HER2+ tumors. The latter subgroups had low, and in some cases non-detectable, asporin expression. Intrigued by these findings we aimed to examine the validity of our observations in more individuals. To do so, we used GOBO and publicly deposited mRNA expression data from breast cancer patients [31]. Analysis of asporin mRNA expression in tumors from different molecular subtypes (Fig 7B) confirmed the results obtained with IHC analysis, demonstrating that asporin expression is high in luminal-A and low in basal-like subtypes (n = 1,280). Evaluation of asporin mRNA expression in tumors of different pathological grades showed that its expression is higher in grade 1 and lower in grade 3 tumors (n = 1,411). This gradual decrease of asporin expression with the grade of the tumor suggested a relationship between asporin expression and breast cancer progression. Thus, we sought to verify how asporin expression correlates with breast cancer patient outcome. We assessed asporin protein expression retrospectively using IHC and tissues from 60 breast cancer patients with over 10-y follow-up (Fig 7C; S2 Table). The patients were divided into two groups: (i) good outcome (signified by no metastatic disease in the follow-up period and following the resection of the primary tumor) and (ii) poor outcome (patients who developed metastases). The two groups showed no major difference in age, tumor size, tumor grade, and ER, PR, HER2, and Ki67 status. The IHC results evidenced significantly higher levels of asporin in patients with good outcome than in patients with poor outcome (2-fold; p = 0.001). The suitability of asporin as a biomarker candidate for predicting metastasis in breast cancer patients was evaluated using a ROC curve. As indicated in Fig 7C, the area under the curve was 0.87 (95% CI 0.78–0.96; p = 0.0001). These data warranted further examination in a larger cohort of patients, with longer survival follow-up. We therefore examined mRNA expression in breast cancers using Kaplan-Meier Plotter and publicly deposited data [32], where the corresponding patients had a post-operative follow-up of 25 y and had no adjuvant treatment (n = 375). The Kaplan-Meier survival curve obtained from these data confirmed the IHC results and demonstrated that low asporin mRNA expression is significantly associated with decreased overall survival (hazard ratio = 0.58; 95% CI 0.37–0.91; logrank p = 0.017) (Fig 7D). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 7. High asporin expression in human breast cancer matches with luminal-like tumor type and good patient outcome. (A) Representative IHC staining of asporin expression in human breast cancer tissues (upper panel). Box plots of asporin expression in 180 breast cancer patients with different status of HER2, ER, and PR are also shown (lower panel). The black line denotes the median expression, and the red line the mean expression. Significant differences in asporin expression were detected among all different subtypes of breast cancer. (B) Analysis of asporin mRNA expression in breast cancer tumors from different molecular subtypes (n = 1,280) and evaluation of asporin mRNA expression in breast cancer of different pathological grades (n = 1,411). (C) Representative IHC staining of asporin expression (upper panel), box plot showing the IHC score (middle panel) in breast cancer tissues from 60 patients with different outcomes, and ROC curve analysis of data obtained from 60 patients with different outcome (lower panel). Scoring and statistics were performed as outlined in the Methods section. (D) Kaplan-Meier survival curve based on asporin mRNA expression in untreated breast cancer with post-operative follow-up of 25 y (n = 375). Images in panels were taken at 100× magnification. All analyses outlined in (B) and (D) were performed using publicly deposited gene expression datasets [31,32] and according to procedures outlined in the Methods. AUC, area under the curve; HR, hazard ratio; SE, standard error. https://doi.org/10.1371/journal.pmed.1001871.g007 Asporin Has Low or No Expression in Most Normal Tissues and Is Overexpressed in Breast Cancer We analyzed a publicly available gene expression repository (BioGPS, Scripps Research Institute) and compared the gene expression profiles in normal tissues of asporin and two other well-studied members of the SLRP family, biglycan and decorin (Fig 1A). The analysis showed that both biglycan and decorin are expressed in many normal tissues, whereas asporin expression was very low or not detected in most normal tissues, except the uterus. Next, we analyzed asporin expression using IHC in breast ductal adenocarcinoma (n = 30) as well as in adjacent non-tumoral tissue and normal breast tissue from breast reduction surgery (n = 10) (Fig 1B). Strong asporin expression was detectable in the stroma of the cancer lesions, with epithelial cancer cells being negative for asporin expression. Adjacent non-tumoral tissue showed a moderate positivity in the extracellular matrix and no positivity in non-tumoral epithelial cells. Healthy breast tissue was negative for asporin expression. Western blot analysis on fresh tissue extracts from matched tumoral and adjacent non-tumoral parts of the resected breast specimens confirmed our IHC observations (Fig 1C). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. Asporin is overexpressed in breast cancer tissues. (A) Tissue-specific pattern of mRNA expression of asporin (ASPN), biglycan (BGN), and decorin (DCN). Source: BioGPS (http://biogps.org). The data are presented as mean ± standard deviation (SD). (B) Representative IHC staining of asporin expression in ductal carcinoma and adjacent non-tumoral breast tissue (left panel) and normal breast tissue obtained from patients undergoing mammary reduction surgery (right panel). Asporin is almost exclusively expressed in breast cancer lesions, while a very low signal is detectable in the adjacent non-tumoral regions. Normal breast tissues are negative. Images of representative fields were taken at 100× and 400× magnification. (C) Western blot analysis of asporin expression in tumoral breast cancer tissues (T) and the adjacent normal counterpart (AdN) of six ductal adenocarcinoma patients. Ponceau red staining was used as loading control. https://doi.org/10.1371/journal.pmed.1001871.g001 Breast Fibroblasts Secrete Asporin after Their Activation by Cancer Cells Owing to the observations made above, in which asporin was found deposited in the extracellular matrix of the tumor, we sought to investigate which cells are responsible for producing the protein. As demonstrated in Fig 2A, none of the human breast cancer cell lines tested showed detectable asporin expression levels (both protein and mRNA). NBFs isolated from the mammary tissue of healthy individuals responded to the CM of several breast cancer cell lines by expressing asporin. The results indicated that tumorigenic and highly metastatic triple-negative breast cancer (TNBC) cells of the basal-like subtype (e.g., MDA-MB-231 and MDA-MB-468) [33–35] did not induce asporin expression in NBFs. This was different in noninvasive luminal-like hormone receptor (HR) positive cell lines (e.g., T47D and MCF-7) [33–35], which activated very high asporin expression in NBFs. Both observations were confirmed at the protein and gene expression levels. Similar experiments with immortalized, non-transformed mammary epithelial cells (MCF-10A) demonstrated that such cells are unable to induce asporin expression in NBFs (Fig 2B). We next sought to examine whether CAFs would express asporin following their isolation from cancer tissue and whether they would react similarly to NBFs when exposed to the CM of breast cancer cells. We isolated CAFs from three breast cancer patients and validated these as pure CAF populations, negative for cytokeratins and overexpressing α-smooth muscle actin (Fig 2C). As shown in Fig 2D, CAFs isolated from HR+ tumors expressed high levels of asporin, which they were able to maintain in vitro (for several weeks) without the need to be in contact with cancer cells. However, CAFs challenged with the CM of TNBC cells (MDA-MB-231) responded by lowering asporin expression. The CM of HR+ cells (MCF-7) was unable to further increase asporin levels in CAFs. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. Asporin is produced by breast fibroblasts in response to conditioned medium from breast cancer cells. (A) Western blot of total cell extracts (upper panel) and qRT-PCR analysis for asporin expression (lower panel) in breast cancer cell lines and NBFs incubated for 48 h with CM collected from a panel of breast cancer cells. (B) Western blot of total cell extracts (upper panel) and qRT-PCR analysis of asporin expression (lower panel) in non-cancerous epithelial breast cell line MCF-10A cells and NBFs incubated for 48 h with CM collected from MCF-10A. Fibroblasts treated with MCF-7 CM were used as the positive control for asporin expression induction. (C) Validation of NBFs and CAFs isolated from patient material. MCF-7 and MDA-MB-231 cells were used as epithelial controls. (D) Western blot analysis of asporin expression in total cell extracts of CAFs obtained from three different patients and treated with the CM of breast cancer cell lines. (A and B): The data are presented as mean ± SD. All panels: HSC70 was used as loading control; Western blots show representative data of three independent experiments. https://doi.org/10.1371/journal.pmed.1001871.g002 TGF-β1 Induces Asporin Expression in Fibroblasts while IL-1β Suppresses It In order to further understand which soluble factors could modulate asporin expression in fibroblasts, we first tested the cytokine TGF-β1. As shown in Fig 3A, TGF-β1 induced asporin expression in NBFs, both under basal conditions and in the presence of CM from MCF-7 cells. In sharp contrast to this, the presence of CM from MDA-MB-231 cells strongly inhibited the ability of TGF-β1 to induce asporin in NBFs. Despite this, our measurements showed that MDA-MB-231 cells, including other TNBC cells, are the highest producers of TGF-β1 among different breast cancer cells (Fig 3B). These findings raised the question of the specific mechanism by which TNBC cells inhibit asporin expression while still producing large quantities of TGF-β1. To further clarify this, we used publicly deposited gene expression data (GEO datasets GSE56265 [two replicates for each cell line] and GSE41445 [three replicates]) comparing the mRNA expression of MDA-MB-231 and MCF-7 cells [36,37]. The summary of the results is shown in Fig 3C. We identified 1,337 genes that were uniquely expressed in MDA-MB-231 cells. Considering that the CM of TNBC cells is able to convey the suppression of asporin expression without the need for cell-to-cell contact, we hypothesized that the effect must be mediated through a soluble protein. We therefore focused only on the genes whose products are known to be soluble proteins. The analysis highlighted a strong cluster of interleukins that were expressed in MDA-MB-231 cells. We next sought to verify which of the observed interleukin genes would discriminate between HR+ (luminal) and TNBC (basal) cells. This analysis was performed using GOBO [32], which compares the profiles of 51 breast cancer cell lines representing different molecular subtypes [35]. The results indicated that IL-1β could be of interest because it is highly expressed in cell lines of the basal-b subtype, which includes MDA-MB-231 cells (Fig 3C, right panel). Encouraged by these in silico findings, we sought to verify the expression of IL-1β in patient material from different breast tumor subtypes and compare this with asporin expression. IHC analysis of ductal adenocarcinoma cases (n = 20) demonstrated that IL-1β expression was significantly increased in TNBC compared to HR+ tumors (Fig 3D). In contrast to this, asporin expression in serial sections of the same tissues followed an inverse trend, with the highest expression in HR+ and the lowest in TNBC tumors. Next we tested whether recombinant human IL-1β could suppress basal, MCF-7-induced, or TGF-β1-induced asporin expression in NBFs. Supplementing CM with IL-1β readily blocked asporin expression in NBFs and CAFs (Fig 3E). We also examined whether IL-1RA, a naturally produced IL-1β inhibitor, would overcome the inhibition of asporin expression in NBFs treated with the CM of MDA-MB-231 cells. Indeed, the pre-incubation of NBFs with IL-1RA blocked the suppressive effect of CM from MDA-MB-231 cells on asporin expression in NBFs (Fig 3F). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. Asporin expression is induced by TGF-β1 and suppressed by IL-1β. (A) Western blot analysis of asporin expression in NBFs treated for 24 h with TGF-β1 with or without CM of MCF-7 and MDA-MB-231 breast cancer cells. (B) ELISA quantification of TGF-β1 levels secreted by human breast cancer cell lines during 48 h. (C) Differential in silico analysis of MCF-7 and MDA-MB-231 gene expression identifies a cluster of interleukins that are uniquely expressed in MDA-MB-231 cells (left panel). Analysis of IL-1β mRNA expression using GOBO (http://co.bmc.lu.se/gobo/gsa.pl) across a panel of breast cancer cell lines [35] subdivided into three subtypes (right panel). (D) Representative IHC analysis (upper panel) of IL-1β and asporin expression in a cohort of breast cancer patients, subdivided into two main subtypes (n = 20). Scoring and statistics (lower panel) were performed as outlined in the Methods section. (E) Inhibition of basal, MCF-7 CM-induced, and TGF-β1-induced asporin expression by IL-1β treatment of NBFs and CAFs. (F) Reversion of MDA-MB-231 CM inhibitory effect on asporin expression using IL-1β natural inhibitor IL-1RA. (B and D): The data are presented as mean ± SD. All panels: Western blots show representative data of three independent experiments. https://doi.org/10.1371/journal.pmed.1001871.g003 Asporin Inhibits TGF-β1-Mediated SMAD2 Activation and Induction of the Epithelial to Mesenchymal Transition Previous studies in chondrocytes [18] identified that the peptide region of asporin (residues 159 to 205) is relevant for the interaction with TGF-β1. We therefore tested whether recombinant asporin and the synthetically produced peptide fragment were able to inhibit TGF-β1-mediated activation of SMAD2 in breast cancer cells (Fig 4A and 4B). SMAD2 was efficiently phosphorylated upon treatment of MDA-MB-468 cells with TGF-β1. The TGF-β1 activity was inhibited when the cytokine was pre-incubated for 1 h at 37°C with increasing doses of recombinant asporin (Fig 4A). Cells treated with recombinant asporin alone showed no modulation of SMAD2 phosphorylation. Analogously to the effects observed with the recombinant protein, the peptide fragment of asporin showed an inhibitory effect on TGF-β1-induced SMAD2 phosphorylation (Fig 4B). Further data showed that parallel treatment of cancer cells with TGF-β1 and asporin peptide, without prior pre-incubation, failed to inhibit SMAD2 phosphorylation. These results are in agreement with previous findings in normal chondrocytes showing that asporin directly binds to TGF-β1, rather than acting as a competitive inhibitor for TGF-β1 receptor [28,38]. To functionally test the ability of asporin to interfere with TGF-β1-induced processes, we employed EpRAS, a murine mammary cancer cell line. EpRAS cells have an established responsiveness to TGF-β1, especially with respect to EMT and migration [28,38]. Similarly to the effects observed in MDA-MB-468 cells, asporin peptide was able to block TGF-β1-mediated phosphorylation of SMAD2 in EpRAS cells (Fig 4C). EpRAS cells are known for undergoing EMT upon stimulation with TGF-β1. Apart from the phenotypic appearance, the EMT switch can be readily observed through the up-regulation of vimentin (VIM) (Fig 4D). TGF-β1-mediated induction of VIM was weaker when TGF-β1 was pre-incubated with asporin peptide (Fig 4D). We further sought to investigate the ability of asporin to interfere with TGF-β1-induced migration of EpRAS cells (Fig 4E). Treatment of EpRAS cells with TGF-β1 induced significant cell migration, while pre-incubation of TGF-β1 with the asporin peptide significantly curbed this effect. The EMT switch has been described as relevant for the acquisition of the cancer stem cell (CSC) phenotype [24]. Therefore, we tested whether TGF-β1-induced EMT would increase the CSC population in EpRAS cells and whether this effect could be inhibited by asporin peptide (Fig 4F). Indeed, TGF-β1 treatment induced an increase of CSCs in EpRAS cells, as evidenced by the established CD44high/CD24low breast cancer stemness signature [39]. The TGF-β1-induced increase of the CSC population was significantly inhibited when asporin peptide was pre-incubated with TGF-β1. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 4. Asporin binds to TGF-β1 and inhibits its downstream signaling and function. (A) Western blot analysis of phospho-SMAD2 (p-SMAD2) and SMAD2 total protein extracts from MDA-MB-468 breast cancer cells treated for 15 min with TGF-β1 and/or human recombinant asporin (Rec. ASPN). (B) Western blot analysis of p-SMAD2 in total protein extracts from MDA-MB-468 breast cancer cells treated with TGF-β1 and/or asporin peptide corresponding to the 159–205 amino acid region (ASPNpep.). (C) Western blot analysis of p-SMAD2 and SMAD2 in total protein extracts from EpRAS cells treated for 15 min with TGF-β1 (5 ng/ml) and/or asporin peptide. (D) EMT induction in EpRAS cells in the presence of TGF-β1 and/or asporin peptide. EMT was monitored both at the phenotype level (upper panel) and using Western blot evaluation of VIM expression in total protein extracts from EpRAS cells (lower panel). (A–D): HSC70 was used as loading control. (E) Transwell migration assay of EpRAS cells pretreated with TGF-β1 (5 ng/ml) and/or asporin peptide (10 μg/ml). (F) Quantification of the CSC population in EpRAS cells following TGF-β1 and/or asporin peptide treatment. (E and F): The data are presented as mean ± SD. All panels: statistical significance was calculated using the Student’s t-test (as described in the Methods section). Western blots show representative data of three independent experiments. https://doi.org/10.1371/journal.pmed.1001871.g004 Induction of Asporin Expression in Triple-Negative Breast Tumors Reduces Growth and Metastasis In Vivo In order to evaluate the impact of asporin expression on TNBC growth and progression, we co-injected MDA-MB-468 cells and NBFs stably overexpressing asporin (NBF-aspn) subcutaneously in NOD-SCID mice (Fig 5A). The control group consisted of mice xenografted with MDA-MB-468 cells with NBFs overexpressing GFP (NBF-ctrl). The growth of the tumors was followed weekly, and the results indicated that the control tumors reached volumes of ~200 mm3 in 6 wk, whereas the asporin-overexpressing tumors grew significantly slower (day 42 post-engraftment: 78.9 mm3 smaller than control; 95% CI 24.3–123.4; p = 0.007) and required 8 wk to reach the same volume (Fig 5B and 5C). Following the resection of the primary tumor, the ability of cancer cells to colonize the lungs was assessed in a follow-up experiment. Two weeks after the removal of the primary tumors, mice were sacrificed and the lungs were collected. Assessment of metastasis formation in the lungs was performed using the Alu-PCR technique. The results evidenced a 3-fold, significant (p = 0.002) increase in the number of cancer cells present in the lung tissue of the control mice compared to the mice with asporin-overexpressing NBFs (Fig 5D). Previously published studies using xenografts based on co-injection of fibroblasts and epithelial cancer cells showed that human fibroblasts are rapidly displaced by murine fibroblasts in vivo [40–42]. We thus sought to verify whether asporin expression remained constant in the tumor during the present experiments. Western blot analysis of tumor tissue extracts showed that asporin expression decreased starting from week 4, reaching low levels at week 6 (Fig 5E). This result suggested the ongoing replacement of human xenografted fibroblasts by murine counterparts. This diminishing asporin expression may underestimate asporin’s effect on tumor growth and metastasis in vivo. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 5. Co-injection of cancer cells and fibroblasts overexpressing asporin reduces primary breast cancer tumor growth and lung metastasis formation in vivo. (A) Western blot analysis of asporin expression in CM of MDA-MB-468 cells and in NBF stable clones used for subcutaneous injection in mice. Ponceau red is shown as loading control. (B) Bioluminescence imaging of control and asporin-expressing xenografts at day 28 after tumor engraftment. The color scale indicates the fluorescent intensity. (C) The volume (in cubic millimeters) of primary tumors measured weekly (from day 7 onwards). The data are presented as mean ± standard error of the mean (SEM) (n = 10 for each group). Statistical significance was calculated using Student’s t-test (**0.01 < p < 0.001; ***0.001 < p < 0.0001). (D) Human-specific Alu-PCR performed on genomic DNA isolated from dissected lungs was used to detect human cancer cells. The data are presented as mean ± SD. (E) Western blot analysis of asporin expression in mice primary tumors monitored for several weeks. HSC70 was used as loading control. https://doi.org/10.1371/journal.pmed.1001871.g005 Therefore, we next sought to engraft asporin-overexpressing cancer cells that would maintain constant asporin expression in the tumor. This was performed with stably transduced asporin-expressing MDA-MB-468 cells (Fig 6A). The control and asporin-expressing MDA-MB-468 cells were implanted subcutaneously in NOD-SCID mice. Primary tumor growth was monitored weekly. The results indicated that asporin-expressing tumors were significantly smaller, reaching up to 2-fold lower volumes at 7 wk post-engraftment (day 49 post-engraftment: 124.1 mm3 smaller than control; 95% CI 75.2–180.4; p = 0.001) (Fig 6B). Histological evaluation demonstrated invasive control tumors developing towards the muscle layers, whereas this was not observed in asporin-expressing counterparts (Fig 6C). Further analysis of asporin-expressing tumors evidenced extensive zones of tumor necrosis in the central areas (Fig 6C), as well as numerous cells with condensed chromatin. In the control conditions necrosis was less pronounced, whereas transparent chromatin staining and the presence of nucleoli further characterized tumor cells. The latter suggested a higher proliferation rate in control tumors. The assessment of tumor proliferation based on Ki67 staining showed stronger and more frequent nuclear positivity in the control tumors in comparison to the asporin-expressing counterparts (Fig 6C). IHC evaluation of asporin expression in the experimental tumors evidenced the expected asporin overexpression. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 6. Asporin reduces primary breast cancer tumor growth and lung metastasis formation in vivo. (A) Western blot analysis of asporin expression in the CM of MDA-MB-468 stable clones expressing asporin, used for subcutaneous injection in mice. Ponceau red is shown as loading control. (B) Bioluminescence imaging of control and asporin-expressing xenografts at day 28 after tumor engraftment (left panel). The color scale indicates the fluorescent intensity. The mean (± SEM) volume (in cubic millimeters) of primary tumors measured weekly (from day 14 onwards) for the time-matched cohort is also shown (n = 20 for each group) (right panel). Statistical significance was calculated using Student’s t-test (**0.01 < p < 0.001; ***0.001 < p < 0.0001). (C) Representative hematoxylin and eosin (H&E), asporin, and Ki67 IHC staining in MDA-MB-468 xenografts collected 7 wk post-engraftment. Control xenografts consistently displayed an invasion in the muscle layer (M). An extended necrotic (N) area was present in the peri-tumoral zone of MDA-MB-468-aspn mice tumors. (D) Quantification of the stem cell population in xenografted tumors expressing asporin using ALDH+ and CD44high/CD24low stemness markers (7 wk post-engraftment). (E) Post-operative follow-up of mice that had primary tumors removed at the same time (time-matched). (F) Mean (± SEM) volume (in cubic millimeters) of primary tumors measured weekly for the size-matched cohort (n = 20 for each group). (G) Post-operative follow-up of mice that had primary tumors removed at the same volume (size-matched). (E and G): IHC evaluation of vimentin in lung necropsies and quantification of metastatic deposits. All images of representative fields were taken at 40×, 100×, and 400× magnification. (D, E, and G): The data are presented as mean ± SD. Statistical significance was calculated using Student’s t-test. https://doi.org/10.1371/journal.pmed.1001871.g006 Considering that asporin blocks TGF-β1 activity and EMT, processes known to enrich stem cells, we hypothesized that the abundance of stem cells would be different in these two experimental conditions. The evaluation of tumor stemness, using two different and independent signatures, showed that asporin-expressing tumors had a significantly lower percentage of stem cells (Fig 6D). As CSCs are essential for tumor survival and metastasis, we sought to evaluate tumor dissemination in mice following tumor resection. For this purpose the animal experiments were divided into two separate cohorts: (i) time-matched and (ii) size-matched. For the time-matched cohort, the tumorectomy was performed at week 7. For the size-matched group, the tumorectomy was conducted at week 9 for control and at week 12 for asporin-expressing tumors (Fig 6F). In both instances the mice were allowed to recover and were observed for axial lymph node and lung metastases during an additional period of 3 wk. As indicated by the time-matched data, control mice developed overt lung metastases, whereas this was not observed in the asporin condition (Fig 6E). Control animals consistently developed frequent and large deposits in the lungs. Animals carrying asporin-expressing tumors also showed lung metastases; however, they were less frequent and of smaller size. The notion that asporin is interfering with the process of metastasis was further confirmed in the size-matched experiments. In this cohort the tumor growth was followed for a longer period of time (control mice 9 wk, asporin 12 wk), highlighting an overall 3-wk delay of tumor growth in asporin-expressing mice. The results quantifying metastases 3 wk post-tumorectomy were similar to those of the time-matched condition (Fig 6G). Collectively, the data obtained with both xenograft models suggested that asporin expression inhibits tumor growth as well as metastatic progression. High Asporin Levels Delineate Breast Cancer Patients with Good Clinical Outcome Considering the in vitro and in vivo data, we expanded our observations using IHC to 180 breast cancer patients, subdivided in four categories with 45 cases each: (i) ER−/PR−/HER2− (triple-negative), (ii) ER+/PR+/HER2+ (triple-positive), (iii) ER+/PR+/HER2− (HR+), and (iv) ER−/PR−/HER2+ (HER2+) (Fig 7A; S1 Table). In our cohort, patients of all subgroups had similar age and showed similar tumor size. Triple-negative and HER2+ cases had higher tumor grade (Bloom 3 versus 2) and a stronger percentage of proliferating cells (Ki67+ cells: ~43% versus ~17%) than the other two subgroups. The frequency of metastasis was highest in TNBC patients (22%), followed by HER2+ and triple-positive breast cancer patients, who displayed similar frequencies (~9%). The IHC results showed that HR+ tumors had a high asporin expression, which was significantly elevated (up to 4-fold) in comparison to TNBC and HER2+ tumors. The latter subgroups had low, and in some cases non-detectable, asporin expression. Intrigued by these findings we aimed to examine the validity of our observations in more individuals. To do so, we used GOBO and publicly deposited mRNA expression data from breast cancer patients [31]. Analysis of asporin mRNA expression in tumors from different molecular subtypes (Fig 7B) confirmed the results obtained with IHC analysis, demonstrating that asporin expression is high in luminal-A and low in basal-like subtypes (n = 1,280). Evaluation of asporin mRNA expression in tumors of different pathological grades showed that its expression is higher in grade 1 and lower in grade 3 tumors (n = 1,411). This gradual decrease of asporin expression with the grade of the tumor suggested a relationship between asporin expression and breast cancer progression. Thus, we sought to verify how asporin expression correlates with breast cancer patient outcome. We assessed asporin protein expression retrospectively using IHC and tissues from 60 breast cancer patients with over 10-y follow-up (Fig 7C; S2 Table). The patients were divided into two groups: (i) good outcome (signified by no metastatic disease in the follow-up period and following the resection of the primary tumor) and (ii) poor outcome (patients who developed metastases). The two groups showed no major difference in age, tumor size, tumor grade, and ER, PR, HER2, and Ki67 status. The IHC results evidenced significantly higher levels of asporin in patients with good outcome than in patients with poor outcome (2-fold; p = 0.001). The suitability of asporin as a biomarker candidate for predicting metastasis in breast cancer patients was evaluated using a ROC curve. As indicated in Fig 7C, the area under the curve was 0.87 (95% CI 0.78–0.96; p = 0.0001). These data warranted further examination in a larger cohort of patients, with longer survival follow-up. We therefore examined mRNA expression in breast cancers using Kaplan-Meier Plotter and publicly deposited data [32], where the corresponding patients had a post-operative follow-up of 25 y and had no adjuvant treatment (n = 375). The Kaplan-Meier survival curve obtained from these data confirmed the IHC results and demonstrated that low asporin mRNA expression is significantly associated with decreased overall survival (hazard ratio = 0.58; 95% CI 0.37–0.91; logrank p = 0.017) (Fig 7D). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 7. High asporin expression in human breast cancer matches with luminal-like tumor type and good patient outcome. (A) Representative IHC staining of asporin expression in human breast cancer tissues (upper panel). Box plots of asporin expression in 180 breast cancer patients with different status of HER2, ER, and PR are also shown (lower panel). The black line denotes the median expression, and the red line the mean expression. Significant differences in asporin expression were detected among all different subtypes of breast cancer. (B) Analysis of asporin mRNA expression in breast cancer tumors from different molecular subtypes (n = 1,280) and evaluation of asporin mRNA expression in breast cancer of different pathological grades (n = 1,411). (C) Representative IHC staining of asporin expression (upper panel), box plot showing the IHC score (middle panel) in breast cancer tissues from 60 patients with different outcomes, and ROC curve analysis of data obtained from 60 patients with different outcome (lower panel). Scoring and statistics were performed as outlined in the Methods section. (D) Kaplan-Meier survival curve based on asporin mRNA expression in untreated breast cancer with post-operative follow-up of 25 y (n = 375). Images in panels were taken at 100× magnification. All analyses outlined in (B) and (D) were performed using publicly deposited gene expression datasets [31,32] and according to procedures outlined in the Methods. AUC, area under the curve; HR, hazard ratio; SE, standard error. https://doi.org/10.1371/journal.pmed.1001871.g007 Discussion A tumor’s ability to successfully grow is increasingly regarded as proportional to the cancer cells’ fitness to survive in a given environment. Their survival is facilitated by adaptation to the environment as well as by actively adapting the environment to the needs of the cancer cell [1,2]. This is in agreement with our key findings that tumor cells with known genetic differences as well as distinct tumorigenic and metastatic potentials have a heterogeneous ability to induce or inhibit asporin expression in stromal fibroblasts. Molecular analysis of all breast cancer cell lines used in the current work [35] suggests that only HR+ cells can induce strong asporin expression in fibroblasts. Contrary to this, CM from TNBC cells strongly inhibits asporin expression in fibroblasts, even when it is induced exogenously by TGF-β1. This, in particular, underlines the evolutionary adaptation of aggressive breast cancer cells. They efficiently exploit a potent cytokine like TGF-β1, yet suppress any unwanted reactions that may result from it (e.g., expression of a natural inhibitor asporin by the stroma). The observations made in vitro were further confirmed in patients, where triple-negative (mainly belonging to basal-like molecular subtype [43]) and HER2+ tumors had the lowest asporin expression. Both are known to be aggressive tumor subtypes with poor clinical outcome [44–46]. The highest asporin levels were observed in the ER+/PR+/HER2− group, which consisted of patients whose tumors were molecularly classified as mainly (~50%) luminal-A subtype [43], which is known to have the best prognosis among all breast tumors [44–46]. Considering this, we were intrigued to identify the mechanism by which TNBC cells manage to suppress asporin expression. Owing to previously published microarray data documenting the differences between triple-negative (e.g., MDA-MB-231) and HR+ (e.g., MCF-7) cells, we identified a strong cluster of several interleukins that were uniquely expressed in TNBC cells. The findings were not surprising, knowing that pro-inflammatory cytokines are essential for “smoldering” inflammation, a key ingredient in cancer and metastasis. The notion that interleukin secretion is responsible for the pro-cancer environment in the tumor stroma context is supported by data from the literature. For example, CXCL1 and CXCL5, secreted by pancreas cancer cells, can activate CXCR2 in fibroblasts to stimulate the production of connective tissue growth factor that, in turn, fuels tumor progression [47]. In the current study we found that IL-1β secreted by MDA-MB-231 cells is indeed responsible for asporin suppression in fibroblasts. The IL-1β protein expression in patient material was inversely correlated with asporin expression. IL-1 consists of two family members, IL-1α and IL-1β, of which only IL-1β is secreted, whereas IL-1α is cytosolic. Earlier studies showed that a single dose of IL-1β is sufficient to significantly increase the number of lung metastases in a melanoma murine tumor model [48]. In line with this, IL-1β expression is elevated in several human tumors (including breast cancer); thus, patients with high IL-1β levels have generally bad clinical outcome [49]. IL-1β induces the expression of several pro-metastatic and pro-angiogenic proteins, among them VEGFA, MMPs, TNFα, and notably TGF-β1 [50]. The present study contributes to further clarifying the intricate mechanism by which IL-1β subverts the tumor stroma into a pro-tumor environment. It does so partly by promoting the expression of tumorigenic cytokines and suppressing their natural inhibitors, in this case asporin, an inhibitor of TGF-β1 [50]. The current findings call for the employment of different strategies to inhibit IL-1β, at least in TNBC. This could be rapidly achieved by employing an anti-IL-1β antibody (canakinumab/Ilaris) or a naturally occurring IL-1β inhibitor IL-1RA (used in vitro in the present work). Recombinant IL-1RA (anakinra/Kineret) is already approved for rheumatoid arthritis treatment, and mounting evidence from gastric [51] and breast cancer [52] research supports its application in treating tumors. A recent phase 1 open-label study with human anti-IL-1 antibody (MABp1) in advanced cancer patients showed encouraging results in terms of disease control, tolerance, and low side effects [53]. Next to considering inhibitors of IL-1β, another axis of treatment could be supported by the asporin peptide (159–205) responsible for TGF-β1 binding. In this work, we showed that the asporin peptide is capable of suppressing different TGF-β1-promoted processes including the acquisition of stem-like phenotype and migration. However, employing the peptide in vivo necessitates further engineering to prevent degradation and increase tissue diffusion. Care would also need to be taken to prevent, or at least diminish, the possibility of an immune reaction against the peptide construct. Future work should certainly address these issues and further explore the possibility of utilizing asporin-derived peptides for the treatment of TNBC, with the aim of slowing progression, suppressing the growth of metastatic lesions, or preventing metastatic dissemination. Decorin and biglycan are other members of the SLRP family that have been shown to be able to bind TGF-β1 [54,55]. Decorin has been labeled as a “guardian from the matrix” because of its ability to sequester a number of cancer-relevant growth factors [56]. However, what makes asporin unique in this context is that decorin and biglycan are expressed during development and broadly expressed in various normal organs [57,58]. The current study underscores the limited expression of asporin in normal adult tissue, qualifying it also as a target for antibody drug conjugates, and highlights its ability to inhibit TGF-β1 downstream signaling, cancer cell migration, and EMT. The in vivo data outlined here support the idea that asporin acts as a tumor suppressor in breast cancer. Asporin-expressing TNBC cells grow significantly slower and are less invasive when xenografted in mice. Following tumor resection, animals with tumors expressing asporin develop fewer and much smaller metastatic deposits in the lungs. One of the possible explanations for this observation may be the interference of asporin with the process of EMT, which is well known to promote tumor dissemination as well as to induce stem cell phenotype [24]. Indeed, the evaluation of two independent and well-established signatures of CSCs, namely ALDH positivity (general) and the abundance of the CD44high/CD24low (breast cancer-specific) population, confirms that asporin-expressing tumors have on average (two signatures together) 50% fewer CSCs. However, recent data on asporin overexpression in gastric cancer show an opposite, pro-invasive function for this stromal protein [17]. For as long as we do not understand all the facets of TGF-β1 biology, the literature may remain a collection of seemingly contradictory findings [59–65]. For example, TGF-β1 inhibition has been previously reported to induce collective cancer cell invasion [66]; hence, it is not surprising that asporin, as a natural inhibitor of TGF-β1, may under certain circumstances contribute to the growth of some tumor types. Therefore, future studies should necessarily take into consideration not only the levels of TGF-β1 but also the expression of its natural inhibitor asporin. The present retrospective study in breast cancer patients with 10-y follow-up underlines the importance of high asporin expression for good clinical outcome. All results obtained by IHC analysis of protein expression levels are further confirmed by the results generated from publicly deposited gene expression data. Collectively, these findings strongly suggest that asporin should be considered as a future diagnostic and prognostic marker, having the potential to stratify breast cancer patients and identify those who are in need for more clinical attention. Therefore, future prospective studies in more patients are required to evaluate the clinical potential of using asporin as a predictive biomarker in breast cancer. Supporting Information S1 Table. Histological characteristics of 180 breast cancer tumors of different subtypes. Values are mean ± SD. Tumor size refers to the diameter (longest axis) of the tumor. Percentage values indicate the proportion of tumor cells that stained positively for the given marker. Frequency of metastasis refers to the respective status at the time point the patient material was collected in the study. https://doi.org/10.1371/journal.pmed.1001871.s001 (DOCX) S2 Table. Histological characteristics of 60 breast cancer tumors of patients with 10-y follow-up. Values are mean ± SD. Tumor size refers to the diameter (longest axis) of the tumor. Percentage values indicate the proportion of tumor cells that stained positively for the given marker. Frequency of metastasis and survival refer to the respective status at the end of the follow-up period. https://doi.org/10.1371/journal.pmed.1001871.s002 (DOCX) S1 Text. ARRIVE checklist. https://doi.org/10.1371/journal.pmed.1001871.s003 (PDF) Acknowledgments The authors acknowledge the experimental support of Dr. Chantal Humblet, Ms. Estelle Dortu, and Ms. Alice Marquet (GIGA—Histology Platform, University of Liège) for the IHC analysis. The authors are thankful to Dr. Sandra Ormenese and Mr. Raafat Stephan (GIGA-Imaging Platform, University of Liège) and Mr. Vincent Hennequière (Metastasis Research Laboratory) for FACS analysis of CSCs. The authors acknowledge the help of Dr. Paul Peixoto and Dr. Barbara Chiavarina (Metastasis Research Laboratory, University of Liège) for isolation of normal tissue fibroblasts. The authors are thankful to Mrs. Naima Maloujahmoum (Metastasis Research Laboratory) for help with the cell lines and to Mr. Luc Duwez (Animal Facility, University of Liège) for help with animal experiments. The authors are also particularly grateful to Dr. Mark E. Sobel and Ms. Ana Turtoi for proofreading the manuscript. The results shown in this work are in part based upon data generated by the The Cancer Genome Atlas Research Network (http://cancergenome.nih.gov/).

doi: 10.1371/journal.pmed.1001876pmid: 26372470

Background Mobile text messaging and medication monitors (medication monitor boxes) have the potential to improve adherence to tuberculosis (TB) treatment and reduce the need for directly observed treatment (DOT), but to our knowledge they have not been properly evaluated in TB patients. We assessed the effectiveness of text messaging and medication monitors to improve medication adherence in TB patients. Methods and Findings In a pragmatic cluster-randomised trial, 36 districts/counties (each with at least 300 active pulmonary TB patients registered in 2009) within the provinces of Heilongjiang, Jiangsu, Hunan, and Chongqing, China, were randomised using stratification and restriction to one of four case-management approaches in which patients received reminders via text messages, a medication monitor, combined, or neither (control). Patients in the intervention arms received reminders to take their drugs and reminders for monthly follow-up visits, and the managing doctor was recommended to switch patients with adherence problems to more intensive management or DOT. In all arms, patients took medications out of a medication monitor box, which recorded when the box was opened, but the box gave reminders only in the medication monitor and combined arms. Patients were followed up for 6 mo. The primary endpoint was the percentage of patient-months on TB treatment where at least 20% of doses were missed as measured by pill count and failure to open the medication monitor box. Secondary endpoints included additional adherence and standard treatment outcome measures. Interventions were not masked to study staff and patients. From 1 June 2011 to 7 March 2012, 4,292 new pulmonary TB patients were enrolled across the 36 clusters. A total of 119 patients (by arm: 33 control, 33 text messaging, 23 medication monitor, 30 combined) withdrew from the study in the first month because they were reassessed as not having TB by their managing doctor (61 patients) or were switched to a different treatment model because of hospitalisation or travel (58 patients), leaving 4,173 TB patients (by arm: 1,104 control, 1,008 text messaging, 997 medication monitor, 1,064 combined). The cluster geometric mean of the percentage of patient-months on TB treatment where at least 20% of doses were missed was 29.9% in the control arm; in comparison, this percentage was 27.3% in the text messaging arm (adjusted mean ratio [aMR] 0.94, 95% CI 0.71, 1.24), 17.0% in the medication monitor arm (aMR 0.58, 95% CI 0.42, 0.79), and 13.9% in the combined arm (aMR 0.49, 95% CI 0.27, 0.88). Patient loss to follow-up was lower in the text messaging arm than the control arm (aMR 0.42, 95% CI 0.18–0.98). Equipment malfunction or operation error was reported in all study arms. Analyses separating patients with and without medication monitor problems did not change the results. Initiation of intensive management was underutilised. Conclusions This study is the first to our knowledge to utilise a randomised trial design to demonstrate the effectiveness of a medication monitor to improve medication adherence in TB patients. Reminders from medication monitors improved medication adherence in TB patients, but text messaging reminders did not. In a setting such as China where universal use of DOT is not feasible, innovative approaches to support patients in adhering to TB treatment, such as this, are needed. Trial Registration Current Controlled Trials, ISRCTN46846388 Background Tuberculosis—a contagious bacterial disease that usually infects the lungs—is a major global public health problem. Every year, about 9 million people develop tuberculosis and at least 1.3 million people die as a result. Mycobacterium tuberculosis, the organism that causes tuberculosis, is spread in airborne droplets when people with tuberculosis cough or sneeze. The symptoms of tuberculosis include cough, weight loss, and fever. Diagnostic tests for tuberculosis include sputum smear microscopy (microscopic analysis of mucus coughed up from the lungs), the growth of M. tuberculosis from sputum samples, and chest X-rays. Tuberculosis can be cured by taking antibiotics daily for several months (usually isoniazid, rifampicin, ethambutol, and pyrazinamide for two months followed by isoniazid and rifampicin for a further four months), but the emergence of multidrug-resistant M. tuberculosis is making tuberculosis increasingly hard to treat. Why Was This Study Done? Because tuberculosis treatment is long and unpleasant, patients often fail to take all their drugs. To improve medication adherence, the World Health Organization recommends that health care workers supervise patients while they take their medication (directly observed treatment, DOT). However, DOT can be hard to implement. In China, for example, where 11% of tuberculosis cases occur, DOT cannot be implemented in many parts of the country, and the national tuberculosis control policy permits self-administered treatment and treatment monitored by family members. It is estimated that over half of individuals with tuberculosis in China self-administer their treatment, but, in 2010, 20% of patients treated using nationally recommended case-management approaches were lost to follow-up or failed to take their medications regularly. In this pragmatic cluster-randomized trial, the researchers investigate whether reminders delivered by mobile phone or by medication monitor boxes (which provide audio reminders to patients and record when the box is opened) might improve tuberculosis medication adherence in China. A pragmatic trial asks whether an intervention works under real-life conditions; a cluster-randomized trial randomly assigns groups of people (here, people living in different counties/districts) to receive alternative interventions and compares outcomes in the differently treated “clusters.” What Did the Researchers Do and Find? The researchers assigned people newly diagnosed with tuberculosis in counties/districts to receive reminders about taking their antibiotics and about monthly follow-up visits via text messaging, a medication monitor, or both text messaging and a medication monitor (the intervention arms), or to receive standard nationally recommended care without electronic reminders (the control arm). All the trial participants (about 1,000 per arm) took their drugs out of a medication monitor box, but the box’s audio reminder function was switched off in the text messaging only and control arms. In the intervention arms, doctors were advised to switch participants with poor medication adherence (evaluated at follow-up visits) to either more intensive management or DOT, depending on the level of missed treatment doses. Compared to the control arm, the percentage of patient-months with at least 20% of the drug doses missed (called “poor adherence” and measured by pill counts and data from the medication monitor) was not significantly reduced in the text messaging arm, whereas poor adherence was significantly reduced by 42% and 51% in the medication monitor and the combined arms, respectively (a significant reduction is unlikely to have happened by chance). Notably, fewer patients were switched to intensive management or DOT than expected based on medication adherence evaluations. What Do These Findings Mean? These findings show that, in China, the use of an electronic medication monitor box to remind patients to take their anti-tuberculosis drugs improved medication adherence. Interestingly, text messaging alone, which has been shown to improve adherence to antiretroviral therapy among HIV-positive individuals, did not improve medication adherence among patients with tuberculosis, possibly because the messages were too frequent or too impersonal, although this intervention (but none of the others) did reduce patient loss to follow-up. Battery problems with the medication monitor may have resulted in over-estimation of poor adherence to treatment. Moreover, the researchers’ assumption that opening the medication monitor box is synonymous with taking the medication may have introduced some inaccuracies into these findings. Despite these limitations and the underuse of more intensive case management in patients with poor adherence, these findings suggest that using medication monitors to deliver electronic drug reminders to patients might improve medication adherence among patients with tuberculosis in China and in other settings. Additional Information This list of resources contains links that can be accessed when viewing the PDF on a device or via the online version of the article at http://dx.doi.org/10.1371/journal.pmed.1001876. This study is further discussed in a PLOS Medicine Perspective by John Metcalfe, Max R. O’Donnell, and David R. Bangsberg The World Health Organization provides information (in several languages) on tuberculosis and on its Directly Observed Treatment Short Course (DOTS) strategy; the Global Tuberculosis Report 2014 provides information about tuberculosis around the world The Stop TB Partnership is working towards tuberculosis elimination and provides personal stories about tuberculosis (in English and Spanish); the Tuberculosis Vaccine Initiative (a not-for-profit organization) also provides personal stories about tuberculosis The US Centers for Disease Control and Prevention provides information about tuberculosis and about treatment for tuberculosis (in English and Spanish) The US National Institute of Allergy and Infectious Diseases also has detailed information on all aspects of tuberculosis MedlinePlus has links to further information about tuberculosis (in English and Spanish) More information about this trial is available Introduction In 2013, China ranked second in the world in number of tuberculosis (TB) cases, accounting for 11% of the estimated 9 million global cases [1]. Implementation of the Directly Observed Treatment, Short Course (DOTS) strategy started in 1992 and covered the entire country by 2005 [2]. Initially, the use of directly observed treatment (DOT) by health care workers was the primary approach to ensure TB treatment adherence. Over time, because of difficulties in carrying out DOT in many parts of the country, national TB control policies also permitted self-administered treatment and treatment monitored by family members. Over half of TB patients now receive self-administered treatment [3]. In the 2010 National Tuberculosis Prevalence Survey, 20% of TB patients treated by the public health system—using national TB case-management approaches—were lost to follow-up or were not taking their medications regularly [4]. Thus, more effective case-management approaches are needed in China. Electronic reminders and monitoring have been used in several disease conditions to improve medication adherence. The potential of mobile phone technology to improve the quality and delivery of health care, including diagnosis, treatment adherence, and data collection, has been recognised [5,6]. Mobile phone text messaging has been shown to improve adherence to antiretroviral treatment and outcomes in HIV-positive patients [7]. However, aside from in a small-scale pilot study [8], the use of text messaging has not been rigorously evaluated in TB patients. Electronic medication packaging (EMP) devices can remind patients to take medications on time, monitor time of drug intake, and alert health care workers to patients who have missed doses [9,10]. The current evidence supporting the use of these devices is limited [10]. In fact, no study to our knowledge has properly evaluated the use of EMP devices in TB patients. Using adherence data from medication monitor boxes (medication monitors) to select less adherent patients for counselling or more intensive forms of case management has been suggested but not studied [11]. To evaluate the use of electronic reminders to improve medication adherence in TB patients, we conducted a cluster-randomised controlled trial to assess the effectiveness of three case-management approaches—using reminders via text messaging, a medication monitor, or both—compared to the standard of care in China. Methods Ethical Approval The study was approved by the ethics committees of the Chinese Center for Disease Control and Prevention (201008) and the London School of Hygiene & Tropical Medicine (5704). All patients provided written consent prior to inclusion in the study. Study Design This study was a pragmatic cluster-randomised trial with one control and three intervention arms. New pulmonary TB patients, starting on standard 6-mo short-course chemotherapy and managed as outpatients, were recruited into the study. Those in the control arm were managed according to the standard of care of the National Tuberculosis Control Program. Those in the three intervention arms also received reminders to take their medications from text messages via short message service (SMS), a medication monitor, or both. If adherence problems were subsequently detected, more intensive management was recommended. For logistical simplicity, randomisation was conducted at the cluster level. Cluster Selection Clusters were defined as rural counties or urban districts within the provinces of Heilongjiang, Jiangsu, Hunan, and Chongqing—located in northern, eastern, central, and western China, respectively. Each cluster had at least 300 active pulmonary TB patients registered in 2009 (S3 Text). Nine clusters, with a rural to urban ratio of 2:1, were selected from two cities in each province. Patient Recruitment In each cluster, consecutive pulmonary TB patients newly registered at the public health TB clinic were screened for study eligibility. Inclusion criteria included the following: no communication impairment (mental, visual, auditory, or speech), patient at least 18 y old, and patient or family member able to use mobile phone to read SMS text messages and use the medication monitor after training. Because of the nature of the study, interventions were not masked to study staff and patients. Randomisation The 36 clusters were randomised to the four arms by rural/urban stratum and restricted such that each province had at least two clusters in each arm. From 5,000 randomly generated acceptable allocations, one was chosen at random as the final allocation using Stata version 12.0. See S3 Text for further details. All Arms All patients were treated according to National Tuberculosis Control Program guidelines including the use of isoniazid, rifampin, ethambutol, and pyrazinamide for 2 mo, followed by isoniazid and rifampin for 4 mo; the programme used every other day dosing for the entire treatment course. Patients received their blister-pack medications in a medication monitor box that electronically collected the date and time of each opening. In the control and text messaging arms, the medication monitor box was in silent mode and was not used as a reminder tool for patients. At each monthly visit, patients were dispensed enough medications for a 1-mo period. Control Arm At the start of treatment, the doctor and the patient selected one of three treatment monitoring approaches as per National Tuberculosis Control Program protocol: self-administered treatment, treatment supervised by family members, or treatment supervised by health care workers. The local doctor monitoring treatment at the township or village/community level was given 60 renminbi (RMB; equivalent to US$10) for each patient. Intervention Arms As in the control arm, patients and their doctors in the intervention arms selected one of the three treatment monitoring approaches as per National Tuberculosis Control Program protocol. The interventions had three common components: reminders for timely drug intake, reminders for monthly follow-up visits, and a recommendation for doctors to switch patients from self-administered treatment to a more intensive treatment monitoring approach when patients were found to have adherence problems based on data available to the managing doctor (Table 1). Download: PPT PowerPoint slide PNG larger image TIFF original image Table 1. Description of the three intervention arms. https://doi.org/10.1371/journal.pmed.1001876.t001 In the text messaging and combined arms, a text message reminded patients to take their medication at the time previously agreed on with the patient. If patient did not reply to the text message, another two text messages would be sent later in the day. Once the patient replied to the SMS reminder, with or without text, the reminders were stopped for that day. Similarly, in the medication monitor and combined arms, an audio reminder from the medication monitor box reminded patients to take their medication. If the patient did not open the medication monitor by a pre-specified time, up to eight additional reminders sounded. Once the box was opened, the reminders were stopped for that day. In all three intervention arms, patients received four reminders to attend the monthly dispensing visit (Table 1). At each monthly follow-up visit, the managing doctor evaluated adherence patterns. Missed doses were defined as the larger of (1) missed doses based on pill count or (2) missed doses from missing SMS reply (in the text messaging only arm) or from failure to open the medication monitor box (in the other two intervention arms). If the patient reported any equipment malfunction or operation error during the previous month, the number of missed doses was based on pill count only. If 1–2 doses were missed in the previous month, we recommended the doctor counsel the patient on the importance of adherence to medication but allowed self-administered treatment to continue. If 3–6 doses were missed, we recommended the doctor switch the patient to “intensive management”, in which township or village/community doctors visited the patient twice a month or once a week, respectively, for the rest of treatment. If seven or more doses were missed the previous month or if 3–6 doses were missed in two prior months, we recommended the doctor switch the patient to DOT, with each dose of treatment supervised by the township or village/community doctor. The local doctors monitoring treatment at the township or village/community level were given 5 RMB (US$0.8) every time they made a visit to a patient as part of the intensive management or DOT, in addition to the 60 RMB (US$10) they received for every patient. Study Endpoints All study endpoints were measured at the individual level. The primary study endpoint for treatment adherence was the percentage of patient-months where at least 20% of doses (equivalent to missing three of 15 doses) were missed (“poor adherence”). The secondary treatment adherence endpoints were (1) percentage of patient-months where at least 47% of doses (equivalent to seven of 15 doses) were missed, (2) percentage of total doses missed over the 6 mo of treatment, (3) binary categorisation of secondary endpoint 2 as <10% versus ≥10% of total doses missed (National Tuberculosis Control Program definition of non-adherent), and (4) percentage of patient-months on TB treatment where at least 20% of doses were missed based on pill count only. Measurement of the adherence endpoints utilised the same data for all four study arms and included data from the medication monitor box, downloaded into a database when patients returned for their monthly medication refill. All adherence endpoints, except secondary endpoint 4, measured the number of missed doses per month as the larger of the number of missed doses from pill count or the number of failures to open the medication monitor box. A month was defined as the number of days between successive appointments, typically 30 d, during which 15 doses should have been taken, but this was adjusted for early/late or missed visits and reduced by the number of days a patient was hospitalised or temporarily discontinued treatment on doctor’s recommendation. Data were censored when a patient died, moved, or permanently discontinued treatment based on a doctor’s decision. For those who were lost to follow-up during treatment, we assumed no drug intake (100% non-adherence) for the period from the date of being lost to follow-up to the date when they should have completed treatment. We also conducted a post hoc sensitivity analysis censoring adherence measurement at the time of loss to follow-up. The secondary TB treatment outcome endpoints, following standard WHO definitions, were (1) poor treatment outcome, defined as death, treatment failure, or patient loss to follow-up and (2) patient loss to follow-up (S3 Text). Routinely recorded data reported to the National Tuberculosis Control Program were used to define TB treatment outcomes. We also included as a secondary endpoint 2-mo smear conversion among those smear-positive at enrolment. Sample Size Sample size calculations were based on a binary endpoint of non-adherence and took into account the study design [12]. Assuming nine clusters per arm, a two-sided type I error of 5%, and a percentage with non-adherence in the control arm of 30%, 110 TB patients per cluster would be required to detect a 40% reduction in the endpoint in the intervention arm with power of 90% and coefficient of variation of 0.25. The sample size was adjusted to 116 per cluster assuming 5% missing endpoint data. An additional power calculation is summarised in S3 Text. Analysis Analysis of all endpoints used standard methods for a small number of clusters [12], accounting for the stratified design and giving each cluster equal weight (S3 Text). Patients who were reassessed as not having TB by the managing doctor or who were switched to a different treatment model within the first month (due to hospitalisation or travel) were excluded from all analyses. Pre-specified sub-group analyses for the primary endpoint were by age group, literacy, gender, and rural/urban setting. There were problems with loose batteries in some of the medication monitors, resulting in a power outage during which data on box openings were not captured. The problem could be easily fixed by the patient or the doctor when they became aware of the problem We conducted a post hoc stratified analysis separating patient-months into those that had a record of a medication monitor problem and those that did not (S3 Text). Analysis was conducted using Stata version 13. Ethical Approval The study was approved by the ethics committees of the Chinese Center for Disease Control and Prevention (201008) and the London School of Hygiene & Tropical Medicine (5704). All patients provided written consent prior to inclusion in the study. Study Design This study was a pragmatic cluster-randomised trial with one control and three intervention arms. New pulmonary TB patients, starting on standard 6-mo short-course chemotherapy and managed as outpatients, were recruited into the study. Those in the control arm were managed according to the standard of care of the National Tuberculosis Control Program. Those in the three intervention arms also received reminders to take their medications from text messages via short message service (SMS), a medication monitor, or both. If adherence problems were subsequently detected, more intensive management was recommended. For logistical simplicity, randomisation was conducted at the cluster level. Cluster Selection Clusters were defined as rural counties or urban districts within the provinces of Heilongjiang, Jiangsu, Hunan, and Chongqing—located in northern, eastern, central, and western China, respectively. Each cluster had at least 300 active pulmonary TB patients registered in 2009 (S3 Text). Nine clusters, with a rural to urban ratio of 2:1, were selected from two cities in each province. Patient Recruitment In each cluster, consecutive pulmonary TB patients newly registered at the public health TB clinic were screened for study eligibility. Inclusion criteria included the following: no communication impairment (mental, visual, auditory, or speech), patient at least 18 y old, and patient or family member able to use mobile phone to read SMS text messages and use the medication monitor after training. Because of the nature of the study, interventions were not masked to study staff and patients. Randomisation The 36 clusters were randomised to the four arms by rural/urban stratum and restricted such that each province had at least two clusters in each arm. From 5,000 randomly generated acceptable allocations, one was chosen at random as the final allocation using Stata version 12.0. See S3 Text for further details. All Arms All patients were treated according to National Tuberculosis Control Program guidelines including the use of isoniazid, rifampin, ethambutol, and pyrazinamide for 2 mo, followed by isoniazid and rifampin for 4 mo; the programme used every other day dosing for the entire treatment course. Patients received their blister-pack medications in a medication monitor box that electronically collected the date and time of each opening. In the control and text messaging arms, the medication monitor box was in silent mode and was not used as a reminder tool for patients. At each monthly visit, patients were dispensed enough medications for a 1-mo period. Control Arm At the start of treatment, the doctor and the patient selected one of three treatment monitoring approaches as per National Tuberculosis Control Program protocol: self-administered treatment, treatment supervised by family members, or treatment supervised by health care workers. The local doctor monitoring treatment at the township or village/community level was given 60 renminbi (RMB; equivalent to US$10) for each patient. Intervention Arms As in the control arm, patients and their doctors in the intervention arms selected one of the three treatment monitoring approaches as per National Tuberculosis Control Program protocol. The interventions had three common components: reminders for timely drug intake, reminders for monthly follow-up visits, and a recommendation for doctors to switch patients from self-administered treatment to a more intensive treatment monitoring approach when patients were found to have adherence problems based on data available to the managing doctor (Table 1). Download: PPT PowerPoint slide PNG larger image TIFF original image Table 1. Description of the three intervention arms. https://doi.org/10.1371/journal.pmed.1001876.t001 In the text messaging and combined arms, a text message reminded patients to take their medication at the time previously agreed on with the patient. If patient did not reply to the text message, another two text messages would be sent later in the day. Once the patient replied to the SMS reminder, with or without text, the reminders were stopped for that day. Similarly, in the medication monitor and combined arms, an audio reminder from the medication monitor box reminded patients to take their medication. If the patient did not open the medication monitor by a pre-specified time, up to eight additional reminders sounded. Once the box was opened, the reminders were stopped for that day. In all three intervention arms, patients received four reminders to attend the monthly dispensing visit (Table 1). At each monthly follow-up visit, the managing doctor evaluated adherence patterns. Missed doses were defined as the larger of (1) missed doses based on pill count or (2) missed doses from missing SMS reply (in the text messaging only arm) or from failure to open the medication monitor box (in the other two intervention arms). If the patient reported any equipment malfunction or operation error during the previous month, the number of missed doses was based on pill count only. If 1–2 doses were missed in the previous month, we recommended the doctor counsel the patient on the importance of adherence to medication but allowed self-administered treatment to continue. If 3–6 doses were missed, we recommended the doctor switch the patient to “intensive management”, in which township or village/community doctors visited the patient twice a month or once a week, respectively, for the rest of treatment. If seven or more doses were missed the previous month or if 3–6 doses were missed in two prior months, we recommended the doctor switch the patient to DOT, with each dose of treatment supervised by the township or village/community doctor. The local doctors monitoring treatment at the township or village/community level were given 5 RMB (US$0.8) every time they made a visit to a patient as part of the intensive management or DOT, in addition to the 60 RMB (US$10) they received for every patient. Study Endpoints All study endpoints were measured at the individual level. The primary study endpoint for treatment adherence was the percentage of patient-months where at least 20% of doses (equivalent to missing three of 15 doses) were missed (“poor adherence”). The secondary treatment adherence endpoints were (1) percentage of patient-months where at least 47% of doses (equivalent to seven of 15 doses) were missed, (2) percentage of total doses missed over the 6 mo of treatment, (3) binary categorisation of secondary endpoint 2 as <10% versus ≥10% of total doses missed (National Tuberculosis Control Program definition of non-adherent), and (4) percentage of patient-months on TB treatment where at least 20% of doses were missed based on pill count only. Measurement of the adherence endpoints utilised the same data for all four study arms and included data from the medication monitor box, downloaded into a database when patients returned for their monthly medication refill. All adherence endpoints, except secondary endpoint 4, measured the number of missed doses per month as the larger of the number of missed doses from pill count or the number of failures to open the medication monitor box. A month was defined as the number of days between successive appointments, typically 30 d, during which 15 doses should have been taken, but this was adjusted for early/late or missed visits and reduced by the number of days a patient was hospitalised or temporarily discontinued treatment on doctor’s recommendation. Data were censored when a patient died, moved, or permanently discontinued treatment based on a doctor’s decision. For those who were lost to follow-up during treatment, we assumed no drug intake (100% non-adherence) for the period from the date of being lost to follow-up to the date when they should have completed treatment. We also conducted a post hoc sensitivity analysis censoring adherence measurement at the time of loss to follow-up. The secondary TB treatment outcome endpoints, following standard WHO definitions, were (1) poor treatment outcome, defined as death, treatment failure, or patient loss to follow-up and (2) patient loss to follow-up (S3 Text). Routinely recorded data reported to the National Tuberculosis Control Program were used to define TB treatment outcomes. We also included as a secondary endpoint 2-mo smear conversion among those smear-positive at enrolment. Sample Size Sample size calculations were based on a binary endpoint of non-adherence and took into account the study design [12]. Assuming nine clusters per arm, a two-sided type I error of 5%, and a percentage with non-adherence in the control arm of 30%, 110 TB patients per cluster would be required to detect a 40% reduction in the endpoint in the intervention arm with power of 90% and coefficient of variation of 0.25. The sample size was adjusted to 116 per cluster assuming 5% missing endpoint data. An additional power calculation is summarised in S3 Text. Analysis Analysis of all endpoints used standard methods for a small number of clusters [12], accounting for the stratified design and giving each cluster equal weight (S3 Text). Patients who were reassessed as not having TB by the managing doctor or who were switched to a different treatment model within the first month (due to hospitalisation or travel) were excluded from all analyses. Pre-specified sub-group analyses for the primary endpoint were by age group, literacy, gender, and rural/urban setting. There were problems with loose batteries in some of the medication monitors, resulting in a power outage during which data on box openings were not captured. The problem could be easily fixed by the patient or the doctor when they became aware of the problem We conducted a post hoc stratified analysis separating patient-months into those that had a record of a medication monitor problem and those that did not (S3 Text). Analysis was conducted using Stata version 13. Results Study Population From 1 June 2011 to 7 March 2012, 6,203 pulmonary TB patients were screened in the 36 clusters, and 5,057 (81.5%) met enrolment criteria, of whom 4,292 (84.9%) gave informed consent. Of these, 61 (1.4%) were reassessed as not having TB by their managing doctor, and 58 (1.4%) were withdrawn from the study as they had switched to a different treatment model within the first month (due to hospitalisation or travel) and were therefore excluded from all analyses (Fig 1; S1 Table). Therefore, 4,173 patients were included in the analysis (Fig 1). There was some variation between arms in the percentages analysed among those screened: 72.5%, 68.2%, 59.4%, and 69.8% in the control, text messaging, medication monitor, and combined arms, respectively (Fig 1). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. Cluster-level CONSORT diagram. Reasons for non-eligibility: SMS req = unable to use mobile phone after training; <18y = less than 18 y of age; comm dis = communication disability. *Withdrew from the study but continued treatment in the local Center for Disease Control and Prevention. https://doi.org/10.1371/journal.pmed.1001876.g001 Overall, 71.0% of participants were male, median age was 43 y (inter-quartile range [IQR] 29 to 56 y), 56.0% were farmers, 7.9% were illiterate, median household income was 20,000 RMB (IQR 10,000 to 30,000 RMB), and 36.3% were smear positive (Table 2). There was some baseline imbalance by study arm for occupation, education level, income, and local residency. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 2. Characteristics at start of tuberculosis treatment for patients enrolled in the four study arms of the study (n = 4,173). https://doi.org/10.1371/journal.pmed.1001876.t002 Endpoints Primary study endpoint. The cluster geometric mean of the percentage of patient-months on TB treatment where at least 20% of doses were missed was 29.9% in the control arm (range 16.0%–48.1%; Fig 2; S2 Table); in comparison, this percentage was 27.3% in the text messaging arm (adjusted mean ratio [aMR] 0.94, 95% CI 0.71, 1.24; Table 3), 17.0% in the medication monitor arm (aMR 0.58, 95% CI 0.42, 0.79), and 13.9% in the combined arm (aMR 0.49, 95% CI 0.27, 0.88). There were no differences in the mean ratios (MRs) for the primary endpoint when stratifying by age, literacy, or gender (S3 Table). There was an indication that the reduction in poor adherence seen in the medication monitor arm compared to the control arm was only in the rural stratum (MR 0.43 for rural and 1.06 for urban, p-value for effect modification 0.011). The coefficient of variation for the primary endpoint among control clusters was 0.24. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. Primary endpoint of poor tuberculosis treatment adherence by study arm. Solid bars represent geometric means of cluster-level proportions. https://doi.org/10.1371/journal.pmed.1001876.g002 Download: PPT PowerPoint slide PNG larger image TIFF original image Table 3. Effectiveness of interventions on tuberculosis treatment adherence and treatment outcomes endpoints. https://doi.org/10.1371/journal.pmed.1001876.t003 Secondary study endpoints. There were similar reductions in the intervention arms versus the control arm in the percentage of months with at least 47% of doses missed (equivalent to 7/15 doses), the percentage of doses missed over the whole treatment period, and the percentage of patients who missed at least 10% of their doses, in both unadjusted and adjusted analyses (Table 3). The percentage of person-months with at least 20% of doses missed as judged by pill count only was much lower than that judged by both pill count and medication monitor data. This secondary endpoint was reduced by 33%–61% in the three intervention arms compared to the control arm, but this reduction was mostly driven by the imputation of months with 100% non-adherence following loss to follow-up (Table 3). The text messaging arm had a lower patient loss to follow-up and occurrence of poor treatment outcome than the control arm. Modest reductions in patient loss to follow-up were also seen for the medication monitor and combined arms, though confidence intervals for the effect estimates included one. A post hoc sensitivity analysis that censored adherence measurement at the time of loss to follow-up showed a strengthening of the evidence for a reduction in poor adherence as measured by pill count in the three intervention arms, but otherwise very similar results (S4 Table). There were too few patients with data on sputum conversion at 2 mo for a formal analysis (summary data in S3 Text). Problems with the medication monitor box, recorded either by the doctor at the monthly visit or by the medication monitor as power interruption, were more common in the medication monitor (49.4% of patients; Table 4) and combined arms (48.0%) than in the control (17.8%) or text messaging arms (16.7%). Stratified analysis of the primary endpoint by noted medication monitor problems showed that the reduction in poor adherence persisted in the medication monitor and combined arms regardless of whether there were power problems (S5 Table). Download: PPT PowerPoint slide PNG larger image TIFF original image Table 4. Intervention process data and medication monitoring data by study arm. https://doi.org/10.1371/journal.pmed.1001876.t004 Process Measures Similar percentages of patients in the three intervention arms were switched to intensive management (3.2%–4.1%) and DOT (0.8%–0.9%) (Table 4). Based on combining the data from pill counts and the medication monitor, the percentages of patients who should have been switched to intensive management or DOT, were 26.3% and 35.4%, respectively, in the text messaging arm, 16.2% and 24.1% in the medication monitor arm and 16.0% and 20.1% in the combined arm. Minor problems with the mobile phones used to receive text messages were also common and were reported by 56.5% of those in the text messaging arm and 27.3% of those in the combined arm (Table 4). These problems included incorrect usage of the phone by the patient (42.0%), network failure (21.1%), and no money on the phone account (14.9%). Problems with the medication monitor or phone were resolved in 88.7% of occurrences (S6 Table). Study Population From 1 June 2011 to 7 March 2012, 6,203 pulmonary TB patients were screened in the 36 clusters, and 5,057 (81.5%) met enrolment criteria, of whom 4,292 (84.9%) gave informed consent. Of these, 61 (1.4%) were reassessed as not having TB by their managing doctor, and 58 (1.4%) were withdrawn from the study as they had switched to a different treatment model within the first month (due to hospitalisation or travel) and were therefore excluded from all analyses (Fig 1; S1 Table). Therefore, 4,173 patients were included in the analysis (Fig 1). There was some variation between arms in the percentages analysed among those screened: 72.5%, 68.2%, 59.4%, and 69.8% in the control, text messaging, medication monitor, and combined arms, respectively (Fig 1). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. Cluster-level CONSORT diagram. Reasons for non-eligibility: SMS req = unable to use mobile phone after training; <18y = less than 18 y of age; comm dis = communication disability. *Withdrew from the study but continued treatment in the local Center for Disease Control and Prevention. https://doi.org/10.1371/journal.pmed.1001876.g001 Overall, 71.0% of participants were male, median age was 43 y (inter-quartile range [IQR] 29 to 56 y), 56.0% were farmers, 7.9% were illiterate, median household income was 20,000 RMB (IQR 10,000 to 30,000 RMB), and 36.3% were smear positive (Table 2). There was some baseline imbalance by study arm for occupation, education level, income, and local residency. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 2. Characteristics at start of tuberculosis treatment for patients enrolled in the four study arms of the study (n = 4,173). https://doi.org/10.1371/journal.pmed.1001876.t002 Endpoints Primary study endpoint. The cluster geometric mean of the percentage of patient-months on TB treatment where at least 20% of doses were missed was 29.9% in the control arm (range 16.0%–48.1%; Fig 2; S2 Table); in comparison, this percentage was 27.3% in the text messaging arm (adjusted mean ratio [aMR] 0.94, 95% CI 0.71, 1.24; Table 3), 17.0% in the medication monitor arm (aMR 0.58, 95% CI 0.42, 0.79), and 13.9% in the combined arm (aMR 0.49, 95% CI 0.27, 0.88). There were no differences in the mean ratios (MRs) for the primary endpoint when stratifying by age, literacy, or gender (S3 Table). There was an indication that the reduction in poor adherence seen in the medication monitor arm compared to the control arm was only in the rural stratum (MR 0.43 for rural and 1.06 for urban, p-value for effect modification 0.011). The coefficient of variation for the primary endpoint among control clusters was 0.24. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. Primary endpoint of poor tuberculosis treatment adherence by study arm. Solid bars represent geometric means of cluster-level proportions. https://doi.org/10.1371/journal.pmed.1001876.g002 Download: PPT PowerPoint slide PNG larger image TIFF original image Table 3. Effectiveness of interventions on tuberculosis treatment adherence and treatment outcomes endpoints. https://doi.org/10.1371/journal.pmed.1001876.t003 Secondary study endpoints. There were similar reductions in the intervention arms versus the control arm in the percentage of months with at least 47% of doses missed (equivalent to 7/15 doses), the percentage of doses missed over the whole treatment period, and the percentage of patients who missed at least 10% of their doses, in both unadjusted and adjusted analyses (Table 3). The percentage of person-months with at least 20% of doses missed as judged by pill count only was much lower than that judged by both pill count and medication monitor data. This secondary endpoint was reduced by 33%–61% in the three intervention arms compared to the control arm, but this reduction was mostly driven by the imputation of months with 100% non-adherence following loss to follow-up (Table 3). The text messaging arm had a lower patient loss to follow-up and occurrence of poor treatment outcome than the control arm. Modest reductions in patient loss to follow-up were also seen for the medication monitor and combined arms, though confidence intervals for the effect estimates included one. A post hoc sensitivity analysis that censored adherence measurement at the time of loss to follow-up showed a strengthening of the evidence for a reduction in poor adherence as measured by pill count in the three intervention arms, but otherwise very similar results (S4 Table). There were too few patients with data on sputum conversion at 2 mo for a formal analysis (summary data in S3 Text). Problems with the medication monitor box, recorded either by the doctor at the monthly visit or by the medication monitor as power interruption, were more common in the medication monitor (49.4% of patients; Table 4) and combined arms (48.0%) than in the control (17.8%) or text messaging arms (16.7%). Stratified analysis of the primary endpoint by noted medication monitor problems showed that the reduction in poor adherence persisted in the medication monitor and combined arms regardless of whether there were power problems (S5 Table). Download: PPT PowerPoint slide PNG larger image TIFF original image Table 4. Intervention process data and medication monitoring data by study arm. https://doi.org/10.1371/journal.pmed.1001876.t004 Primary study endpoint. The cluster geometric mean of the percentage of patient-months on TB treatment where at least 20% of doses were missed was 29.9% in the control arm (range 16.0%–48.1%; Fig 2; S2 Table); in comparison, this percentage was 27.3% in the text messaging arm (adjusted mean ratio [aMR] 0.94, 95% CI 0.71, 1.24; Table 3), 17.0% in the medication monitor arm (aMR 0.58, 95% CI 0.42, 0.79), and 13.9% in the combined arm (aMR 0.49, 95% CI 0.27, 0.88). There were no differences in the mean ratios (MRs) for the primary endpoint when stratifying by age, literacy, or gender (S3 Table). There was an indication that the reduction in poor adherence seen in the medication monitor arm compared to the control arm was only in the rural stratum (MR 0.43 for rural and 1.06 for urban, p-value for effect modification 0.011). The coefficient of variation for the primary endpoint among control clusters was 0.24. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. Primary endpoint of poor tuberculosis treatment adherence by study arm. Solid bars represent geometric means of cluster-level proportions. https://doi.org/10.1371/journal.pmed.1001876.g002 Download: PPT PowerPoint slide PNG larger image TIFF original image Table 3. Effectiveness of interventions on tuberculosis treatment adherence and treatment outcomes endpoints. https://doi.org/10.1371/journal.pmed.1001876.t003 Secondary study endpoints. There were similar reductions in the intervention arms versus the control arm in the percentage of months with at least 47% of doses missed (equivalent to 7/15 doses), the percentage of doses missed over the whole treatment period, and the percentage of patients who missed at least 10% of their doses, in both unadjusted and adjusted analyses (Table 3). The percentage of person-months with at least 20% of doses missed as judged by pill count only was much lower than that judged by both pill count and medication monitor data. This secondary endpoint was reduced by 33%–61% in the three intervention arms compared to the control arm, but this reduction was mostly driven by the imputation of months with 100% non-adherence following loss to follow-up (Table 3). The text messaging arm had a lower patient loss to follow-up and occurrence of poor treatment outcome than the control arm. Modest reductions in patient loss to follow-up were also seen for the medication monitor and combined arms, though confidence intervals for the effect estimates included one. A post hoc sensitivity analysis that censored adherence measurement at the time of loss to follow-up showed a strengthening of the evidence for a reduction in poor adherence as measured by pill count in the three intervention arms, but otherwise very similar results (S4 Table). There were too few patients with data on sputum conversion at 2 mo for a formal analysis (summary data in S3 Text). Problems with the medication monitor box, recorded either by the doctor at the monthly visit or by the medication monitor as power interruption, were more common in the medication monitor (49.4% of patients; Table 4) and combined arms (48.0%) than in the control (17.8%) or text messaging arms (16.7%). Stratified analysis of the primary endpoint by noted medication monitor problems showed that the reduction in poor adherence persisted in the medication monitor and combined arms regardless of whether there were power problems (S5 Table). Download: PPT PowerPoint slide PNG larger image TIFF original image Table 4. Intervention process data and medication monitoring data by study arm. https://doi.org/10.1371/journal.pmed.1001876.t004 Process Measures Similar percentages of patients in the three intervention arms were switched to intensive management (3.2%–4.1%) and DOT (0.8%–0.9%) (Table 4). Based on combining the data from pill counts and the medication monitor, the percentages of patients who should have been switched to intensive management or DOT, were 26.3% and 35.4%, respectively, in the text messaging arm, 16.2% and 24.1% in the medication monitor arm and 16.0% and 20.1% in the combined arm. Minor problems with the mobile phones used to receive text messages were also common and were reported by 56.5% of those in the text messaging arm and 27.3% of those in the combined arm (Table 4). These problems included incorrect usage of the phone by the patient (42.0%), network failure (21.1%), and no money on the phone account (14.9%). Problems with the medication monitor or phone were resolved in 88.7% of occurrences (S6 Table). Discussion Our study found that the use of a medication monitor to remind TB patients to take their drugs reduced poor medication adherence by 40%–50% compared to the standard of care in China’s National Tuberculosis Control Program. This reduction was seen for all TB treatment adherence measures in this study. The use of text messaging did not reduce poor medication adherence but did reduce patient loss to follow-up by 58%. The use of a medication monitor alone resulted in a smaller, and not statistically significant, reduction in patient loss to follow-up compared to control; however, the study was not powered for this treatment outcome. Even though many types of EMP devices exist and have been used for different disease conditions, a recent systematic review concluded that there were limited data supporting their use in improving adherence [9]. This study is the first to our knowledge to utilise a randomised trial design to demonstrate the effectiveness of an EMP device in improving medication adherence in TB patients. The use of our medication monitor box was integrated into the public health management of TB treatment. Such integration seems to be more frequently associated with improved medication adherence [9,10]. As the largest study to date, to our knowledge, to evaluate an EMP device for any disease condition, this study provides important evidence supporting the use of EMP to improve medication adherence. Our results demonstrate that text messaging did not reduce poor medication adherence among TB patients. This contrasts with available evidence supporting the use of text messaging among HIV patients on antiretroviral therapy [7], which led to a strong recommendation from WHO for its use [13]. However, not all text messages are effective. There was a trend for greater effects of an intervention with texts that were less frequent than daily and with more personalised messages [7]. Frequent medication reminders using text messages can result in user fatigue. And some experts suggest that the most important factor is whether patients feel cared for, not the length or frequency of the text messages. Perhaps the lack of a more personalised engagement, the didactic nature of our messages, multiple messages per day, and the SMS message being received when the patient was not in close proximity to his/her medication all contributed to the failure to reduce poor adherence. There has been recent interest in using mobile phone technology to improve adherence to TB medication [5,14] and TB treatment outcomes [15], though, as yet, few studies have reported their findings [16]. In our intervention arms, we recommended that doctors switch patients to intensive patient management or DOT when adherence problems were documented. However, this rarely happened, despite data suggesting that a substantial percentage of patients should have switched. The trial was designed to be pragmatic, and so we did not enforce the initiation of more intensive management or DOT. Because problems with medication monitors, mobile phones, or their use were frequently reported, it is possible that doctors largely chose to ignore the electronic adherence data when deciding whether to switch patients to more intensive case-management approaches. In addition, doctors may not have had sufficient financial incentives to carry out more intensive case management. Even though more intensive case-management approaches were underutilised in the presence of recorded treatment non-adherence, we still observed better adherence in the medication monitor and combined arms. This suggests that the use of a medication monitor to remind patients to take their medications can improve treatment adherence by itself. If information on poor treatment adherence had been used by providers to switch patients to more intensive case-management approaches, as intended, it is likely we would have seen an even greater reduction in poor treatment adherence with the use of medication monitors. Interestingly, text messaging reduced the risk of patient loss to follow-up. Perhaps text messaging is an effective approach to remind patients of follow-up visits and resulted in better attendance at monthly visits. However, a recent meta-analysis suggests that the effectiveness of SMS reminders for appointments is modest at best and not more effective than other types of reminders [6,17]. The differences in the effects of the interventions in terms of adherence and treatment outcome endpoints suggest these do not correlate well. However, adherence is complex, and a recent taxonomy divides it into three constructs—initiation (patient takes the first dose), implementation (measure of how patient’s actual dosing history corresponds to the prescribed dosing regimen from initiation until the last dose is taken), and discontinuation (patient stops taking the prescribed medication) [18]. Given that our primary adherence endpoint is predominantly “implementation”, the seemingly discrepant results are not surprising; the evidence did not change when the primary endpoint was restricted to “implementation” only, in a post hoc sensitivity analysis. We defined poor adherence based on a threshold of 20% missed doses, a threshold commonly used in other disease areas [19]. Our study had several limitations. First, the battery problems with the medication monitors in our study led to loss of data in some patients, potentially resulting in an over-estimation of poor adherence. However, when we performed a stratified analysis using patient-months with or without this problem, we found similar reductions in poor treatment adherence. Second, more intensive case-management approaches were underutilised, possibly because doctors disregarded information from the medication monitor or SMS feedback. In addition, the financial incentives given to the doctors to perform more intensive management may have been inadequate. Third, for the adherence endpoints, we assumed that opening the medication monitor box was synonymous with drug intake, which may not have been the case, though our measure of poor adherence using a combination of this and pill count is arguably more robust than pill count alone. Pill counts have often been shown to under-report poor adherence or non-adherence [18], as is also shown in our study, where the geometric mean for the adherence endpoint measured using pill count only was lower than that of the primary endpoint. Further, data from a separate study indicated high correlation between adherence measured by medication monitor and rifampicin detected in urine [20]. Other limitations included differences in percentages enrolled by study arm, baseline imbalance of some factors, the unmasked nature of the trial, and the study not being powered for the treatment outcome endpoints. In spite of these limitations, this is the largest study to date, to our knowledge, to evaluate the use of text messaging or medication monitors to improve medication adherence in TB patients. As a pragmatic trial, implemented by the National Tuberculosis Control Program and mimicking real-world conditions, this study has produced useful lessons for future study designs. The use of a medication monitor as a reminder for drug intake in combination with the identification of patients requiring more intensive management has been suggested as an approach for improving TB treatment adherence [11]. This is the first study to our knowledge to rigorously evaluate such an approach. Based on our results, the use of a medication monitor shows great promise. In a setting such as China, where universal use of DOT is not feasible, innovative approaches that help patients adhere to TB treatment are needed. The development of a low-cost and reliable medication monitor, as well as evidence that its use can improve clinical outcomes, could enable widespread use of medication monitors in national TB control programmes. Supporting Information S1 Table. Characteristics at start of tuberculosis treatment for patients in the four study arms who were withdrawn from the study due to hospitalisation or travel (n = 58). https://doi.org/10.1371/journal.pmed.1001876.s001 (DOCX) S2 Table. Primary outcome of poor adherence (defined as percentage of patient-months in which a patient missed at least 20% of doses) by cluster, study arm, and rural/urban stratum. https://doi.org/10.1371/journal.pmed.1001876.s002 (DOCX) S3 Table. Pre-specified sub-group analyses of the primary endpoint (percentage of patient-months with at least 20% doses missed). https://doi.org/10.1371/journal.pmed.1001876.s003 (DOCX) S4 Table. Effectiveness of interventions for endpoints of tuberculosis treatment adherence based on adherence measures before imputation of non-adherence for those who were lost to follow-up (post hoc sensitivity analysis). https://doi.org/10.1371/journal.pmed.1001876.s004 (DOCX) S5 Table. Sensitivity analysis of the primary endpoint (percentage of patient-months with at least 20% doses missed): post hoc sub-group analysis. https://doi.org/10.1371/journal.pmed.1001876.s005 (DOCX) S6 Table. Problems with medication monitors and mobile phones by study arm. https://doi.org/10.1371/journal.pmed.1001876.s006 (DOCX) S1 Text. Trial protocol. https://doi.org/10.1371/journal.pmed.1001876.s007 (DOCX) S2 Text. CONSORT checklist. https://doi.org/10.1371/journal.pmed.1001876.s008 (DOCX) S3 Text. Trial methods and results. https://doi.org/10.1371/journal.pmed.1001876.s009 (DOCX) Acknowledgments We thank the thousands of participants who consented to take part in this study.

doi: 10.1371/journal.pmed.1001880pmid: 26418737

Background The continued advance of antibiotic resistance threatens the treatment and control of many infectious diseases. This is exemplified by the largest global outbreak of extensively drug-resistant (XDR) tuberculosis (TB) identified in Tugela Ferry, KwaZulu-Natal, South Africa, in 2005 that continues today. It is unclear whether the emergence of XDR-TB in KwaZulu-Natal was due to recent inadequacies in TB control in conjunction with HIV or other factors. Understanding the origins of drug resistance in this fatal outbreak of XDR will inform the control and prevention of drug-resistant TB in other settings. In this study, we used whole genome sequencing and dating analysis to determine if XDR-TB had emerged recently or had ancient antecedents. Methods and Findings We performed whole genome sequencing and drug susceptibility testing on 337 clinical isolates of Mycobacterium tuberculosis collected in KwaZulu-Natal from 2008 to 2013, in addition to three historical isolates, collected from patients in the same province and including an isolate from the 2005 Tugela Ferry XDR outbreak, a multidrug-resistant (MDR) isolate from 1994, and a pansusceptible isolate from 1995. We utilized an array of whole genome comparative techniques to assess the relatedness among strains, to establish the order of acquisition of drug resistance mutations, including the timing of acquisitions leading to XDR-TB in the LAM4 spoligotype, and to calculate the number of independent evolutionary emergences of MDR and XDR. Our sequencing and analysis revealed a 50-member clone of XDR M. tuberculosis that was highly related to the Tugela Ferry XDR outbreak strain. We estimated that mutations conferring isoniazid and streptomycin resistance in this clone were acquired 50 y prior to the Tugela Ferry outbreak (katG S315T [isoniazid]; gidB 130 bp deletion [streptomycin]; 1957 [95% highest posterior density (HPD): 1937–1971]), with the subsequent emergence of MDR and XDR occurring 20 y (rpoB L452P [rifampicin]; pncA 1 bp insertion [pyrazinamide]; 1984 [95% HPD: 1974–1992]) and 10 y (rpoB D435G [rifampicin]; rrs 1400 [kanamycin]; gyrA A90V [ofloxacin]; 1995 [95% HPD: 1988–1999]) prior to the outbreak, respectively. We observed frequent de novo evolution of MDR and XDR, with 56 and nine independent evolutionary events, respectively. Isoniazid resistance evolved before rifampicin resistance 46 times, whereas rifampicin resistance evolved prior to isoniazid only twice. We identified additional putative compensatory mutations to rifampicin in this dataset. One major limitation of this study is that the conclusions with respect to ordering and timing of acquisition of mutations may not represent universal patterns of drug resistance emergence in other areas of the globe. Conclusions In the first whole genome-based analysis of the emergence of drug resistance among clinical isolates of M. tuberculosis, we show that the ancestral precursor of the LAM4 XDR outbreak strain in Tugela Ferry gained mutations to first-line drugs at the beginning of the antibiotic era. Subsequent accumulation of stepwise resistance mutations, occurring over decades and prior to the explosion of HIV in this region, yielded MDR and XDR, permitting the emergence of compensatory mutations. Our results suggest that drug-resistant strains circulating today reflect not only vulnerabilities of current TB control efforts but also those that date back 50 y. In drug-resistant TB, isoniazid resistance was overwhelmingly the initial resistance mutation to be acquired, which would not be detected by current rapid molecular diagnostics employed in South Africa that assess only rifampicin resistance. Background Tuberculosis (TB)—a contagious bacterial disease that usually infects the lungs—is a global public health problem. Every year, about 9 million people develop active TB disease, and 1.5 million people die from the disease. Mycobacterium tuberculosis, the organism that causes TB, is spread in airborne droplets when people with TB cough. The symptoms of TB include cough, weight loss, and fever. Diagnostic tests for the disease include sputum smear microscopy (microscopic analysis of mucus coughed up from the lungs) and chest X-rays. TB can be cured by taking a regimen of multiple antibiotics daily for 6 mo. However, the emergence of multidrug-resistant tuberculosis (MDR-TB, TB with resistance to both isoniazid and rifampicin) and extensively drug-resistant tuberculosis (XDR-TB, MDR-TB with additional resistance to both quinolones and second-line injectable agents), together with the spread of HIV (which increases susceptibility to TB), is now threatening TB control efforts. MDR-TB is caused by M. tuberculosis strains that have acquired mutations (genetic changes) that make them resistant to isoniazid, rifampicin, and sometimes other anti-TB drugs; XDR-TB is caused by bacteria that are resistant to isoniazid, rifampicin, one or more fluoroquinolones (for example, ofloxacin), and at least one injectable second-line drug (for example, kanamycin). Why Was This Study Done? A better understanding of the origins of drug-resistant TB is essential for effective control of TB. Public health experts need to know whether the emergence of drug-resistant TB is caused by inadequacies in TB control or related to other factors such as the spread of HIV and whether new resistant strains of M. tuberculosis repeatedly emerge during XDR-TB outbreaks or whether the transmission of a single drug-resistant strain drives these outbreaks. Here, the researchers use whole genome sequencing and dating analysis to investigate the origin and evolution of an XDR-TB outbreak identified in 2005 in Tugela Ferry, KwaZulu-Natal, South Africa. The predominant strain of XDR M. tuberculosis isolated during this large XDR-TB outbreak belongs to a subfamily called LAM4. Since the outbreak began, XDR-TB has also been reported in hospitals across KwaZulu-Natal, and some of these outbreaks have been caused by bacterial strains not falling within the LAM4 spoligotype (“spoligotyping” characterizes M. tuberculosis strains based on the presence of unique DNA sequences in a specific region of the bacterial genome). What Did the Researchers Do and Find? The researchers tested the antibiotic susceptibility of 337 clinical isolates of M. tuberculosis collected in KwaZulu-Natal between 2008 and 2013 and of three historical isolates—two collected in the province in the mid-1990s and a third from the Tugela Ferry XDR outbreak. They sequenced the whole genome of these isolates and used comparative techniques to assess the isolates’ relatedness and to investigate the acquisition of drug resistance. This analysis revealed a 50-member clone of XDR bacteria among the isolates collected across KwaZulu-Natal that was highly related to the LAM4 strain (a clone is defined here as a set of strains in which each member differs by no more than ten single nucleotide polymorphisms [SNPs] from at least one other member; an SNP is a type of genetic variant). Mutations that conferred isoniazid resistance in this clone were acquired in about 1957; MDR and XDR strains emerged in about 1984 and 1995, respectively. The analysis also indicates that MDR and XDR evolved de novo 56 times and nine times, respectively, and that isoniazid resistance nearly always evolved before rifampicin resistance. What Do These Findings Mean? These findings provide new information about the ordering and timing of the acquisition of drug-resistance mutations by M. tuberculosis in KwaZulu-Natal but do not necessarily represent the evolution of XDR-TB in other settings. Most notably, these findings indicate that the ancestral precursor of the Tugela Ferry XDR outbreak strain gained resistance to first-line antibiotics shortly after these antibiotics became available for clinical use. Subsequent stepwise accumulation of additional resistance mutations that occurred over decades led to the emergence of MDR and XDR strains. Importantly, the emergence of these strains occurred before the explosion of HIV in KwaZulu-Natal. Thus, these findings highlight the dire repercussions of the failure of historic attempts to control resistance to first-line anti-TB drugs and draw attention to the need for new anti-TB drugs to be used prudently to prevent early fixation of resistance and to protect the useful lifespan of these agents. Finally, the finding that isoniazid resistance is a key initiation event for progression to MDR and XDR suggests that TB control programs should test routinely for both isoniazid and rifampicin resistance to ensure early detection of drug-resistant TB. Additional Information This list of resources contains links that can be accessed when viewing the PDF on a device or via the online version of the article at http://dx.doi.org/10.1371/journal.pmed.1001880. The World Health Organization (WHO) provides information (in several languages) on TB and on MDR-TB; the Global Tuberculosis Report 2014 provides information about TB around the world; a supplement to the report entitled “Drug-Resistant TB—Surveillance and Response” is available The Stop TB Partnership is working towards TB elimination and provides personal stories about TB (in English and Spanish) The United States Centers for Disease Control and Prevention provides information about TB and about drug-resistant TB (in English and Spanish) The US National Institute of Allergy and Infectious Diseases also has detailed information on TB, including a drug-resistant TB visual tour TB & Me, a collaborative blogging project run by patients being treated for MDR-TB and Mèdecins sans Frontiéres, provides more patient stories The not-for-profit organization Global Health Education provides information about TB in South Africa MedlinePlus has links to further information about TB (in English and Spanish) Introduction The global burden of tuberculosis (TB) remains high, with an estimated 9 million active disease cases and 1.5 million deaths in 2013 [1]. Multidrug-resistant (MDR) TB, defined as Mycobacterium tuberculosis with in vitro resistance to both isoniazid and rifampicin, accounted for at least 480,000 incident cases and 210,000 attributed deaths in 2013 [1]. Extensively drug-resistant (XDR) TB, which is MDR with additional resistance to both quinolones and second-line injectable agents [2], has been reported in 100 countries [1]. With high morbidity, XDR poses a dire threat to public health, particularly in populations with high HIV prevalence [1,3]. The incidence of TB in South Africa is estimated by the WHO to be 860 (776–980) per 100,000 population, which is among the highest in the world [1]. With a population of approximately 10 million, KwaZulu-Natal is the easternmost of South Africa’s nine provinces. While its provincial TB incidence is similar to that of the rest of South Africa (889 per 100,000 in 2012, based on treatment initiation data) [4], KwaZulu-Natal has been notable for disproportionately high rates of drug-resistant TB [4,5]. Compounding this epidemic, South Africa has seen a dramatic increase in HIV prevalence in the last 25 y. The Joint United Nations Programme on HIV and AIDS (UNAIDS) estimates that national adult HIV prevalence was only 0.3% in 1990 but rose to 19.1% in 2013 [6]. In KwaZulu-Natal, HIV rates are particularly high, with 37.4% HIV prevalence documented among pregnant women in 2011 [4]. In 2005, the identification of an outbreak of XDR-TB at the Church of Scotland Hospital in KwaZulu-Natal, Tugela Ferry, raised global alarm and called attention to the prospect of dissemination of potentially untreatable TB [7]. Not only was resistance to four or more classes of antibiotics noted in these strains, but also the disease, in the context of HIV coinfection, was rapidly fatal, with 98% mortality [7]. Traditional genotyping by IS6110 fingerprinting and spoligotyping identified a predominant strain of global M. tuberculosis lineage 4 and spoligotype, ST60, later termed LAM4/F15/KZN (henceforth referred to as LAM4), suggestive of a clonal outbreak of a single drug-resistant strain [7–11]. Targeted sequencing of resistance mutations in a subset of these XDR strains revealed identical mutations [9], further supporting the theory of acquisition of XDR-level resistance and subsequent transmission. Nosocomial spread was deemed likely by a social network analysis [9]. Since the events at Church of Scotland Hospital, which still stands as the largest outbreak of XDR-TB ever reported, XDR-TB has been reported in the majority of hospitals across KwaZulu-Natal [12], and more than 516 XDR-TB cases have been reported within Tugela Ferry alone [13]. In addition, XDR-TB caused by strains not falling within the LAM4 spoligotype have been seen, indicating repeated evolutionary emergences of XDR among strains circulating within the region [10,14]; however, the relative contribution of de novo versus vertically inherited resistance of XDR-TB is unknown. It is also unknown how and when XDR-level drug resistance developed, information that could be exploited to detect and prevent higher-level resistances from emerging in South Africa and elsewhere in the world. While accumulation of drug-resistance mutations can confer a fitness cost to bacteria, subsequent development of compensatory mutations can ameliorate these costs by restoring certain affected physiological functions while maintaining drug resistance [15–17]. Identification of compensatory mutations among clinical strains of M. tuberculosis [18–20] has improved our understanding of drug resistance and fitness, but this area remains incompletely studied. Whole genome sequencing efforts that target large collections of M. tuberculosis have provided critical insights into M. tuberculosis population dynamics, including M. tuberculosis transmission and the molecular causes of drug resistance [20–23]. Although some strains from KwaZulu-Natal have been sequenced [24,25], there has been no large-scale sequencing project from this province or studies that have systematically addressed the molecular evolution of XDR. In the largest compilation of whole genome sequences from clinical isolates of M. tuberculosis from South Africa, we used a combination of comparative genomic techniques to elucidate when and how epidemic XDR drug resistance emerged. With knowledge of a strain’s date of collection, determination of the number of single nucleotide polymorphism (SNP) differences between sequenced strains, and the estimated mutation rate of M. tuberculosis, we were able to utilize Bayesian [26] inference to estimate the dates of acquisition of resistance mutations within the Tugela Ferry ancestor. We discuss the implications of these findings with respect to current and future TB control. Methods Specimen Collection and Characterization We selected 337 clinical isolates of M. tuberculosis with diverse drug susceptibility patterns. Strains were collected both retrospectively and prospectively from 2008 to 2013 from all 11 districts of KwaZulu-Natal (Table 1). Strains were chosen for study inclusion on the basis of a predetermined drug resistance pattern so that the dataset was heavily weighted toward drug-resistant strains rather than accurately reflecting the epidemiology of the region. Written informed consent was obtained from study participants prior to cohort enrollment. Biomedical Research Ethics Council (BREC) approval from the University of KwaZulu-Natal was granted for whole genome sequencing of clinical strains. On all study isolates, drug susceptibility testing (DST) was performed by the critical concentration method, using the WHO recommended concentrations [27]. The following drugs were assayed in all strains, with their respective critical concentration in parentheses (in μg/mL): rifampicin (1.0), isoniazid (0.2 and/or 1.0), streptomycin (2.0), kanamycin (6.0), and ofloxacin (2.0). Extended DST was performed for key isolates (Table 1) with the following drugs: capreomycin (10.0), ethambutol (7.5), and ethionamide (10.0). Pyrazinamide resistance testing was performed using PZA MGIT (100.0) or Nicotinamide (500.0). Subject data included age, gender, AFB smear, and HIV status, when available. Study participants were assigned GPS coordinates corresponding to their home provincial district or site of sputum collection. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 1. Description of study isolates. The 337 clinical study isolates derived from five patient cohorts and were both prospectively and retrospectively collected from all 11 districts of KwaZulu-Natal from 2008 to 2013. Culture conditions describe the initial M. tuberculosis isolation method from sputum. If subsequent single colony isolation (SCI) was performed prior to DNA extraction on the entire study cohort or a subset of the cohort, then this is denoted. Drugs for which DST was performed are abbreviated as follows: rifampicin (R), isoniazid (H), nicotinamide (N), pyrazinamide (P), ethambutol (E), streptomycin (S), kanamycin (K), ofloxacin (O), ethionamide (Et), and capreomycin (C). https://doi.org/10.1371/journal.pmed.1001880.t001 We also selected for sequencing three historical strains previously collected in KwaZulu-Natal for resequencing [25,24]: KZN4207 (drug susceptible, collected in Durban in 1995), KZN1435 (MDR, collected in Durban in 1994), and KZN605 (XDR, collected in Tugela Ferry in 2005). Whole Genome Sequencing Genomic DNA was extracted using published methods [31]. The majority of strains were single colony selected prior to DNA isolation (S1 Methods and S1 Table). Library preparation and whole genome sequencing (WGS) were performed as previously described on the Illumina HiSeq 2000 at the Broad Institute [32]. The median depth of sequencing was 143x, and coverage of the H37Rv genome was 99.9%. Sequencing data were submitted to the Sequence Read Archive NCBI under the following umbrella BioProject identifiers: PRJNA183624 and PRJNA235615. Bioinformatic Analysis Primary analysis. Reads were mapped onto a reference strain of H37Rv (GenBank accession number CP003248.2) using BWA version 0.5.9 [33]. In cases in which read coverage of the reference was greater than 200x, reads were down-sampled using Picard [34] prior to mapping. Positions that varied relative to the reference were identified using Pilon version 1.5 as described [32]. Strain diversity and biogeography. We conducted phylogenetic analyses for both the entire set of 340 strains, as well as for a subset of 111 strains belonging to the LAM4 spoligotype. For each set, all sites with unambiguous SNPs in at least one strain were combined into a concatenated alignment. Ambiguous positions were treated as missing data. The concatenated alignment was then used to generate a midpoint rooted phylogenetic tree in RAxML (version 7.3.3) [35] under a GTRCAT substitution model with 1,000 bootstrap replicates. Global M. tuberculosis lineage designations were assigned based on phylogeny and regions of difference [36]. Each strain’s “digital” spoligotype was predicted by statistically testing for the presence of each of 43 unique spacer sequences used in classical spoligotyping from sequence reads. Results were matched to spacer pattern profiles at SITVITWEB to generate a named spoligotype (S1 Methods) [37]. Clonal strains were identified using a density-based clustering algorithm [38] that grouped strains that differ by no more than ten SNPs to at least one other member within a clone (S1 Methods) [39–41]. Mantel tests were performed to evaluate the relationship between genetic and geographic distances among strains using the ZT software v1.1 [42]. Pairwise genetic distances were calculated as the number of SNP differences between strains, and geographic distances were calculated using the haversine formula [43] and points of origin for strain pairs. Ordering and dating evolution of drug resistance. A curated list of genomic polymorphisms associated with drug resistance was defined for each tested drug based on a literature review (S1 Methods). Polymorphisms associated with compensatory mechanisms to isoniazid, rifampicin, and ethambutol were also defined (S1 Methods). Strains with predicted resistance were identified based on the carriage of mutations from the curated list. We used PAUP [44] to reconstruct the patterns of drug resistance mutation gains and losses throughout the phylogenetic tree representing all 340 strains. PAUP was run using a cost matrix that assigned a 10x greater cost for a loss event relative to a gain event. We used BEAST [26] to estimate a mutation rate and to determine dates for the acquisition of mutations within the LAM4 spoligotype. BEAST was run for 50 million iterations, sampling every 1,000 iterations, using the relaxed lognormal clock (uncorrelated) model. The relaxed molecular clock model assumes independent rates on different branches, which was consistent with previously published reports [45], as well as initial BEAST analyses that we conducted involving lineages 2 and 4, indicating that there may be substantial variation in evolutionary rates within M. tuberculosis. In addition, since the BEAST statistic “ucld.stdev” was greater than zero (0.189) for our dataset, this indicated that our data did exhibit rate heterogeneity within the LAM4 spoligotype. The first 5 million iterations were excluded as “burn-in.” We used the GTR + Gamma substitution model, estimated base frequencies, and the “Gamma + invariant sites” site heterogeneity model. We enforced the topology of the SNP-based tree determined using RAxML [35]. We used a starting value for the mean mutation rate of 0.35 SNPs/genome/year [39,41,46–48]. We assayed a range of values for the starting mean mutation rate, covering the range of values previously reported in the literature, with little difference in the output. BEAti was used to construct the BEAST input file, and default values were used for all other priors. The program Tracer was used to examine mixing and effective sample size (ESS) in order to assess chain length and model convergence. ESS indicates the number of effectively independent draws from the posterior distribution to which the Markov chain is equivalent. A low ESS for a particular parameter (ESS < 100) would indicate that the trace contained a lot of correlated samples and thus may not well represent the posterior distribution. In our analysis, all statistics had an ESS greater than 150. The results were consistent across several runs of the same model. Estimated dates are given with 95% highest posterior density (HPD) intervals. Specimen Collection and Characterization We selected 337 clinical isolates of M. tuberculosis with diverse drug susceptibility patterns. Strains were collected both retrospectively and prospectively from 2008 to 2013 from all 11 districts of KwaZulu-Natal (Table 1). Strains were chosen for study inclusion on the basis of a predetermined drug resistance pattern so that the dataset was heavily weighted toward drug-resistant strains rather than accurately reflecting the epidemiology of the region. Written informed consent was obtained from study participants prior to cohort enrollment. Biomedical Research Ethics Council (BREC) approval from the University of KwaZulu-Natal was granted for whole genome sequencing of clinical strains. On all study isolates, drug susceptibility testing (DST) was performed by the critical concentration method, using the WHO recommended concentrations [27]. The following drugs were assayed in all strains, with their respective critical concentration in parentheses (in μg/mL): rifampicin (1.0), isoniazid (0.2 and/or 1.0), streptomycin (2.0), kanamycin (6.0), and ofloxacin (2.0). Extended DST was performed for key isolates (Table 1) with the following drugs: capreomycin (10.0), ethambutol (7.5), and ethionamide (10.0). Pyrazinamide resistance testing was performed using PZA MGIT (100.0) or Nicotinamide (500.0). Subject data included age, gender, AFB smear, and HIV status, when available. Study participants were assigned GPS coordinates corresponding to their home provincial district or site of sputum collection. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 1. Description of study isolates. The 337 clinical study isolates derived from five patient cohorts and were both prospectively and retrospectively collected from all 11 districts of KwaZulu-Natal from 2008 to 2013. Culture conditions describe the initial M. tuberculosis isolation method from sputum. If subsequent single colony isolation (SCI) was performed prior to DNA extraction on the entire study cohort or a subset of the cohort, then this is denoted. Drugs for which DST was performed are abbreviated as follows: rifampicin (R), isoniazid (H), nicotinamide (N), pyrazinamide (P), ethambutol (E), streptomycin (S), kanamycin (K), ofloxacin (O), ethionamide (Et), and capreomycin (C). https://doi.org/10.1371/journal.pmed.1001880.t001 We also selected for sequencing three historical strains previously collected in KwaZulu-Natal for resequencing [25,24]: KZN4207 (drug susceptible, collected in Durban in 1995), KZN1435 (MDR, collected in Durban in 1994), and KZN605 (XDR, collected in Tugela Ferry in 2005). Whole Genome Sequencing Genomic DNA was extracted using published methods [31]. The majority of strains were single colony selected prior to DNA isolation (S1 Methods and S1 Table). Library preparation and whole genome sequencing (WGS) were performed as previously described on the Illumina HiSeq 2000 at the Broad Institute [32]. The median depth of sequencing was 143x, and coverage of the H37Rv genome was 99.9%. Sequencing data were submitted to the Sequence Read Archive NCBI under the following umbrella BioProject identifiers: PRJNA183624 and PRJNA235615. Bioinformatic Analysis Primary analysis. Reads were mapped onto a reference strain of H37Rv (GenBank accession number CP003248.2) using BWA version 0.5.9 [33]. In cases in which read coverage of the reference was greater than 200x, reads were down-sampled using Picard [34] prior to mapping. Positions that varied relative to the reference were identified using Pilon version 1.5 as described [32]. Strain diversity and biogeography. We conducted phylogenetic analyses for both the entire set of 340 strains, as well as for a subset of 111 strains belonging to the LAM4 spoligotype. For each set, all sites with unambiguous SNPs in at least one strain were combined into a concatenated alignment. Ambiguous positions were treated as missing data. The concatenated alignment was then used to generate a midpoint rooted phylogenetic tree in RAxML (version 7.3.3) [35] under a GTRCAT substitution model with 1,000 bootstrap replicates. Global M. tuberculosis lineage designations were assigned based on phylogeny and regions of difference [36]. Each strain’s “digital” spoligotype was predicted by statistically testing for the presence of each of 43 unique spacer sequences used in classical spoligotyping from sequence reads. Results were matched to spacer pattern profiles at SITVITWEB to generate a named spoligotype (S1 Methods) [37]. Clonal strains were identified using a density-based clustering algorithm [38] that grouped strains that differ by no more than ten SNPs to at least one other member within a clone (S1 Methods) [39–41]. Mantel tests were performed to evaluate the relationship between genetic and geographic distances among strains using the ZT software v1.1 [42]. Pairwise genetic distances were calculated as the number of SNP differences between strains, and geographic distances were calculated using the haversine formula [43] and points of origin for strain pairs. Ordering and dating evolution of drug resistance. A curated list of genomic polymorphisms associated with drug resistance was defined for each tested drug based on a literature review (S1 Methods). Polymorphisms associated with compensatory mechanisms to isoniazid, rifampicin, and ethambutol were also defined (S1 Methods). Strains with predicted resistance were identified based on the carriage of mutations from the curated list. We used PAUP [44] to reconstruct the patterns of drug resistance mutation gains and losses throughout the phylogenetic tree representing all 340 strains. PAUP was run using a cost matrix that assigned a 10x greater cost for a loss event relative to a gain event. We used BEAST [26] to estimate a mutation rate and to determine dates for the acquisition of mutations within the LAM4 spoligotype. BEAST was run for 50 million iterations, sampling every 1,000 iterations, using the relaxed lognormal clock (uncorrelated) model. The relaxed molecular clock model assumes independent rates on different branches, which was consistent with previously published reports [45], as well as initial BEAST analyses that we conducted involving lineages 2 and 4, indicating that there may be substantial variation in evolutionary rates within M. tuberculosis. In addition, since the BEAST statistic “ucld.stdev” was greater than zero (0.189) for our dataset, this indicated that our data did exhibit rate heterogeneity within the LAM4 spoligotype. The first 5 million iterations were excluded as “burn-in.” We used the GTR + Gamma substitution model, estimated base frequencies, and the “Gamma + invariant sites” site heterogeneity model. We enforced the topology of the SNP-based tree determined using RAxML [35]. We used a starting value for the mean mutation rate of 0.35 SNPs/genome/year [39,41,46–48]. We assayed a range of values for the starting mean mutation rate, covering the range of values previously reported in the literature, with little difference in the output. BEAti was used to construct the BEAST input file, and default values were used for all other priors. The program Tracer was used to examine mixing and effective sample size (ESS) in order to assess chain length and model convergence. ESS indicates the number of effectively independent draws from the posterior distribution to which the Markov chain is equivalent. A low ESS for a particular parameter (ESS < 100) would indicate that the trace contained a lot of correlated samples and thus may not well represent the posterior distribution. In our analysis, all statistics had an ESS greater than 150. The results were consistent across several runs of the same model. Estimated dates are given with 95% highest posterior density (HPD) intervals. Primary analysis. Reads were mapped onto a reference strain of H37Rv (GenBank accession number CP003248.2) using BWA version 0.5.9 [33]. In cases in which read coverage of the reference was greater than 200x, reads were down-sampled using Picard [34] prior to mapping. Positions that varied relative to the reference were identified using Pilon version 1.5 as described [32]. Strain diversity and biogeography. We conducted phylogenetic analyses for both the entire set of 340 strains, as well as for a subset of 111 strains belonging to the LAM4 spoligotype. For each set, all sites with unambiguous SNPs in at least one strain were combined into a concatenated alignment. Ambiguous positions were treated as missing data. The concatenated alignment was then used to generate a midpoint rooted phylogenetic tree in RAxML (version 7.3.3) [35] under a GTRCAT substitution model with 1,000 bootstrap replicates. Global M. tuberculosis lineage designations were assigned based on phylogeny and regions of difference [36]. Each strain’s “digital” spoligotype was predicted by statistically testing for the presence of each of 43 unique spacer sequences used in classical spoligotyping from sequence reads. Results were matched to spacer pattern profiles at SITVITWEB to generate a named spoligotype (S1 Methods) [37]. Clonal strains were identified using a density-based clustering algorithm [38] that grouped strains that differ by no more than ten SNPs to at least one other member within a clone (S1 Methods) [39–41]. Mantel tests were performed to evaluate the relationship between genetic and geographic distances among strains using the ZT software v1.1 [42]. Pairwise genetic distances were calculated as the number of SNP differences between strains, and geographic distances were calculated using the haversine formula [43] and points of origin for strain pairs. Ordering and dating evolution of drug resistance. A curated list of genomic polymorphisms associated with drug resistance was defined for each tested drug based on a literature review (S1 Methods). Polymorphisms associated with compensatory mechanisms to isoniazid, rifampicin, and ethambutol were also defined (S1 Methods). Strains with predicted resistance were identified based on the carriage of mutations from the curated list. We used PAUP [44] to reconstruct the patterns of drug resistance mutation gains and losses throughout the phylogenetic tree representing all 340 strains. PAUP was run using a cost matrix that assigned a 10x greater cost for a loss event relative to a gain event. We used BEAST [26] to estimate a mutation rate and to determine dates for the acquisition of mutations within the LAM4 spoligotype. BEAST was run for 50 million iterations, sampling every 1,000 iterations, using the relaxed lognormal clock (uncorrelated) model. The relaxed molecular clock model assumes independent rates on different branches, which was consistent with previously published reports [45], as well as initial BEAST analyses that we conducted involving lineages 2 and 4, indicating that there may be substantial variation in evolutionary rates within M. tuberculosis. In addition, since the BEAST statistic “ucld.stdev” was greater than zero (0.189) for our dataset, this indicated that our data did exhibit rate heterogeneity within the LAM4 spoligotype. The first 5 million iterations were excluded as “burn-in.” We used the GTR + Gamma substitution model, estimated base frequencies, and the “Gamma + invariant sites” site heterogeneity model. We enforced the topology of the SNP-based tree determined using RAxML [35]. We used a starting value for the mean mutation rate of 0.35 SNPs/genome/year [39,41,46–48]. We assayed a range of values for the starting mean mutation rate, covering the range of values previously reported in the literature, with little difference in the output. BEAti was used to construct the BEAST input file, and default values were used for all other priors. The program Tracer was used to examine mixing and effective sample size (ESS) in order to assess chain length and model convergence. ESS indicates the number of effectively independent draws from the posterior distribution to which the Markov chain is equivalent. A low ESS for a particular parameter (ESS < 100) would indicate that the trace contained a lot of correlated samples and thus may not well represent the posterior distribution. In our analysis, all statistics had an ESS greater than 150. The results were consistent across several runs of the same model. Estimated dates are given with 95% highest posterior density (HPD) intervals. Results Our study included 337 participants with an average age of 33.8 y and a standard deviation of 10.7 y, of whom 165 (49%) were male (Table 2). Overall, 140 patients were HIV positive, 51 were HIV negative, and 146 had unknown HIV status. Baseline characteristics were similar among HIV-positive and HIV-negative individuals, with the exception that HIV-negative individuals were younger (p = 0.0030), more likely to be smear positive (p = 0.0139), and more likely to live outside of eThekwini, the provincial capital (p = 0.0132). Download: PPT PowerPoint slide PNG larger image TIFF original image Table 2. Demographic characteristics of participants and phenotypic drug susceptibility of strains. Data are n (%) or mean ± standard deviation (SD). https://doi.org/10.1371/journal.pmed.1001880.t002 A clinical sample was obtained from each patient; M. tuberculosis was isolated using standard approaches and phenotypic DST was performed on each isolate using standard methodology (S1 Methods). Phenotypic DST revealed 88 susceptible, 23 monodrug-resistant (defined as phenotypic resistance to only one drug), 19 polydrug-resistant (defined as phenotypic resistance to two drugs that does not meet criteria for MDR), 140 MDR sensu stricto, and 67 XDR M. tuberculosis strains (Table 2). Phenotypic MDR and XDR-TB cases were identified in all 11 districts of KwaZulu-Natal. While we observed a trend toward HIV-negative individuals harboring more drug-susceptible TB, this observation did not meet statistical significance (p = 0.0542). We performed WGS on all 337 clinical strains, as well as on three historical strains isolated prior to the study collection period. We assessed the diversity and phylogenetic relatedness among strains using information from 17,232 variable sites with SNPs relative to the H37Rv reference genome (Fig 1; S1 Fig). The resulting phylogenetic tree revealed four of the seven main global lineages of M. tuberculosis [36,49–51] to be circulating in KwaZulu-Natal during the sampling time frame. The vast majority of isolates (95%) belonged to lineages 2 and 4, with lesser representation from lineages 1 and 3. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. Diverse strains contribute to drug resistance in KwaZulu-Natal. (A) Midpoint rooted maximum-likelihood phylogeny of 340 M. tuberculosis isolates. Four of the seven known M. tuberculosis lineages were identified: CAS (Lin1), Beijing (Lin2), EAI (Lin 3), and Euro-American (Lin4). Digital spoligotyping identified 17 unique spoligotypes in the dataset; spoligotypes are shown on this figure if they are represented by three or more strains. Corresponding spoligotypes and phenotypes are reported for all strains in S4 Table. Phenotypic XDR, MDR, poly- and monodrug resistance (labeled “Drug-resistant other”), and pansusceptible strains are indicated by colored tick marks at the tip of each leaf node. (B) Histogram of pairwise SNP distances between strains. The number of pairs within each SNP distance range is plotted. The peaks correspond to the distance between major lineages. The peak at the far left of the figure corresponds to the distance between pairs of strains within a clone. https://doi.org/10.1371/journal.pmed.1001880.g001 A computational or digital spoligotype prediction was performed, and 17 unique spoligotypes were identified (S1 Table) [37]. Spoligotype diversity was well represented in all districts of KwaZulu-Natal (Fig 2, panel A). Using a Mantel test, we determined that there was very low correlation between geographic and genetic distances among strains (r = -0.067906, p = 0.001760), indicating that strains did not cluster geographically. Older transmission events and/or high patient mobility between districts likely account for this pattern. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. Wide geographic spread of diverse strains across KwaZulu-Natal and wide distribution of XDR and the Tugela Ferry XDR Clone members. A map of the 11 districts of KwaZulu-Natal [52] is shown in which pie charts indicate (A) the fraction of sequenced M. tuberculosis belonging to computationally predicted spoligotypes (see key). In (B), the fraction of strains with a phenotypic classification of XDR and membership in the Tugela Ferry XDR Clone are represented (see key). The size of the pie chart indicates the relative number of strains sequenced from each of the 11 districts within KwaZulu Natal. Tugela Ferry, in the uMzinyathi district, is indicated in red. https://doi.org/10.1371/journal.pmed.1001880.g002 We defined a “clone” as a set of strains in which each member differs by no more than ten SNPs to at least one other member, which is similar to definitions used in previous genomic studies of M. tuberculosis transmission (S2 Fig, S1 Methods) [39–41]. Nearly one-third of the strains (107 of 340, 31%) belonged to 11 such clones (S2 Table), which were distributed across six spoligotypes and three lineages (S3 Fig). All clones were phenotypically drug resistant, indicating recent person-to-person spread of a diverse set of drug-resistant strains that included both HIV-positive and HIV-negative individuals. The “historical” Tugela Ferry XDR strain, KZN605, was nested phylogenetically within a large clone of 50 LAM4 strains with predominantly phenotypic XDR (Fig 3). All of the strains within this clone (henceforth referred to as the Tugela Ferry XDR Clone) possessed the characteristic drug resistance mutations that were previously identified in XDR-TB strains circulating in Tugela Ferry during the outbreak [9,24], further indicating this clone’s continued prevalence within KwaZulu-Natal. Patients in whom the Tugela Ferry XDR Clone was isolated were from ten of the 11 districts within the province (Fig 2, panel B). In addition, the Tugela Ferry XDR Clone was not overrepresented among HIV-positive patients (p = 0.6750) (Table 2). This suggests that strains within this clone were neither geographically constrained nor restricted to immunodeficient hosts. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. Molecular evolution and dating of drug resistance emergence within the Tugela Ferry XDR Clone. Midpoint rooted maximum-likelihood phylogeny of 107 M. tuberculosis isolates of the LAM4 spoligotype. The gray shaded box identifies the Tugela Ferry XDR Clone. KZN605, the historical XDR strain collected in Tugela Ferry during the outbreak, is a member of this clone. Two additional historical isolates, KZN1435 and KZN4207, are not members of the Tugela Ferry XDR Clone. Each evolutionary gain of a drug resistance mutation was assigned to its position on the phylogenetic tree by parsimony (colored circles). A–E traces the stepwise order of drug resistance acquisition in the Tugela Ferry XDR Clone and estimates the year when each mutation was gained. Gray bars indicate the 95% highest posterior density (HPD) intervals. (A) katG S315T (isoniazid); gidB 130 bp deletion (streptomycin); 1957 (95% HPD: 1937–1971); (B) inhA promoter -8 (isoniazid and ethionamide); 1964 (95% HPD: 1948–1976); (C) embB M306V (ethambutol); 1967 (95% HPD: 1950–1978); (D) rpoB L452P (rifampicin); pncA 1bp insertion (pyrazinamide); 1984 (95% HPD: 1974–1992); and (E) rpoB D435G (rifampicin); rrs 1400 (kanamycin); gyrA A90V (ofloxacin); 1995 (95% HPD: 1988–1999). The accumulation of individual drug-resistant mutations within a strain is denoted to the right of the phylogenetic tree. The dates of drug discovery are displayed at the bottom of the figure [53]. Four additional LAM4 strains on a distant branch were not included in this figure because of size constraints. Bootstrap values are provided for lettered nodes, and bootstrap values for all nodes are shown in S5 Fig. https://doi.org/10.1371/journal.pmed.1001880.g003 Many of the sequenced LAM4 strains were closely related to the Tugela Ferry XDR Clone but had different DST profiles (Fig 3 and S2 Fig), giving us an opportunity to finely dissect the order of acquisition of mutations giving rise to the Tugela Ferry XDR Clone [9,24]. LAM4 strain phylogeny was recalculated using data from only LAM4 strains, and parsimony was used to place the origin of known resistance-conferring mutations on the tree. The recalculated LAM4 tree was consistent with our previous tree containing data from all strains with all key internal nodes involved in the evolution of drug resistance having bootstrap values greater than 89%. This enabled us to confidently assign evolutionary ordering of drug resistance mutation acquisition (Fig 3 and S5 Fig). As shown in Fig 3, the first step towards XDR-level resistance in this epidemic clone was the acquisition of isoniazid and streptomycin resistance-conferring mutations in katG and gidB, respectively, which were gained at node A of the phylogenetic tree (100% bootstrap support). With accumulation of successive mutations, the ancestral strain (and its descendants) gained (i) additional polydrug resistance to ethionamide and ethambutol via mutations in the inhA promoter and embB (nodes B and C, respectively, 100% and 89% bootstrap support), (ii) MDR via mutations in rpoB and pncA that conferred resistance to rifampicin and pyrazinamide (node D, 100% bootstrap support), and (iii) XDR via mutations in rrs and gyrA, which conferred resistance to kanamycin and ofloxacin, respectively, and an additional rpoB mutation (node E, 97% bootstrap support). This ordering was highly supported by bootstrapping (all key nodes had bootstrap values ≥89%) in the phylogenetic reconstruction. Thus, the first step towards XDR-level drug resistance in this epidemic clone was the acquisition of isoniazid and streptomycin resistance followed by ethambutol and ethionamide resistance, then rifampicin and pyrazinamide resistance, and, ultimately, kanamycin and ofloxacin resistance. Because we had dates of isolation for all sequenced strains—including strains that were isolated more than 20 y ago—we applied a Bayesian statistical approach to estimate when mutations leading to the Tugela Ferry XDR Clone emerged. Using this approach, which takes into account the phylogeny of LAM4 strains, the dates of their isolation, and published mutation rates for M. tuberculosis [39,41,46–48], we calculated that LAM4 in KZN mutated at a rate of 0.61 SNPs/genome/year. This mutation rate was higher than other previously published mutation rates, regardless of which rate from the literature was used as the starting mean value. Applying this rate, we estimated that drug resistance mutations at node A were acquired in 1957 (95% HPD: 1937–1971), soon after streptomycin and isoniazid were developed. MDR-level resistance was acquired in 1984 (95% HPD: 1974–1992; node D), and XDR-level resistance was acquired in 1995 (95% HPD: 1988–1999; node E), 10 y prior to its acute recognition in 2005 in Tugela Ferry (Fig 3). The dating analysis within the LAM4 spoligotype consistently assigned drug resistance gains after the drug discovery date, indicating that drug resistance emergence in the region mirrored the dates of drug discovery. We also observed multiple drug resistance mutations within LAM4 that emerged outside the Tugela Ferry XDR Clone (Fig 3). Many of these mutations were acquired at leaf nodes, which implied very recent gains of resistance. Including the Tugela Ferry ancestor, we calculated that genotypic MDR sensu stricto—defined as both isoniazid and rifampicin resistance-conferring mutations—independently arose a minimum of 13 times. Within LAM4, the Tugela Ferry XDR Clone represented the single and only evolutionary gain of genotypic XDR—as defined by acquisition of resistance-conferring mutations to the four XDR-defining drugs: isoniazid, rifampicin, ofloxacin, and kanamycin. However, within LAM4 we also observed ten independent gains of either a kanamycin or an ofloxacin resistance-conferring mutation in a background of genotypic MDR sensu stricto. As such, 13 LAM4 strains identified in this study would be considered genotypic “pre-XDR” and only one SNP away from XDR-level resistance. Beyond LAM4, we observed many other independent evolutionary emergences of MDR and XDR across this dataset. Twelve and seven spoligotypes contained strains with phenotypic MDR and XDR, respectively (S3 Table), suggesting that these resistance patterns emerged no fewer than 12 and 7 times. However, when we quantified the total number of independent evolutionary emergences of genotypic MDR and XDR across our entire dataset, we estimated that MDR sensu stricto and XDR evolved no less than 56 and nine independent times, respectively (S3 Table). Remarkably, the first drug resistance acquisition in the Tugela Ferry XDR Clone was consistent with other emergences of MDR and XDR across the entire dataset. For the 214 strains with genotypic resistance to two or more of the MDR and XDR defining drugs, we quantified the number of evolutions in which a specific drug resistance mutation was gained before a second resistance mutation. We observed that isoniazid resistance via nonsynonymous mutation at the katG S315 codon was gained before rifampicin resistance in 46 unique evolutionary events, whereas rifampicin resistance was never acquired before the katG S315 mutation (Fig 4). When we repeated this for all pairwise comparisons, we found that isoniazid resistance, conferred by mutation of the katG S315 codon, preceded or co-occurred with resistance mutations to all other drugs in our dataset. Mutations other than the katG S315 mutations that confer isoniazid resistance (i.e., inhA promoter mutations or katG deletions) occurred before rifampicin resistance mutations in nine unique events, whereas we only observed the reverse ordering twice. These data indicate that, beyond the Tugela Ferry XDR Clone, isoniazid resistance, and in particular the S315 codon mutation in katG, has been the initial resistance-conferring mutation leading to polydrug resistance, including MDR and XDR, among strains from KwaZulu-Natal. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 4. Isoniazid resistance is the first step towards drug resistance. Acquisition of katG S315 mutations precedes all other resistance mutations, including rifampicin, in all instances in which the order of acquisition can be disambiguated. For the 214 strains with genotypic resistance to two or more MDR or XDR defining drugs, and in which the order of acquisition of these mutations could be disambiguated, we quantified the number of evolutions in which resistance to one drug was gained before resistance to a second drug. Isoniazid resistance was divided into mutations conferred by the katGS315 codon versus “Other INH” mutations (defined as loss-of-function mutations in katG that do not involve codon 315 or mutations in the inhA promoter). Reported numbers represent the number of independent evolutionary events (not the number of strains) in which the drug resistance indicated by the row labeled “first resistance” was acquired before the drug resistance indicated by the column labeled “second resistance.” The background color is shaded to indicate the fraction of unambiguous evolutionary events in which the “first resistance” was acquired before the “second resistance” for that given drug pair. https://doi.org/10.1371/journal.pmed.1001880.g004 As in other organisms, in vitro studies have suggested that drug resistance in M. tuberculosis may be associated with a variable fitness cost that can be offset by compensatory evolution [18,15]. Nonsynonymous mutations in the α and β’ subunits of RNA polymerase have been postulated to compensate for fitness costs associated with rifampicin resistance [15,19]. Among our 226 strains with phenotypic rifampicin resistance, 76 strains had mutations known to compensate for rifampicin resistance (S4 Table) [20,19,54,55]. Using the phylogenetic framework and parsimony, we determined that 23 of the 27 previously described compensatory mutations had an evolutionary pattern consistent with rifampicin compensation, i.e., mutations that evolved only after or concurrent with mutations that conferred rifampicin resistance. We also attempted to identify novel rifampicin compensatory mutations with this approach. In addition to the 27 previously described mutations, we detected an additional 38 nonsynonymous polymorphisms in rpoA, rpoC, and the non-rifampicin resistance-determining regions (RRDR) regions of rpoB (S4 Table). By parsimony analysis, we established the acquisition order of these rpoA, rpoC, and non-RRDR rpoB mutations in relation to genotypic rifampicin resistance. An additional 26 of these previously uncharacterized mutations also evolved in a pattern consistent with a role in compensation, which suggests that they may also function in this capacity. While there were ten unique RRDR mutations with subsequent or concurrent evolutionary gain of a putative rifampicin compensatory polymorphism, the vast majority of putative compensatory mutations occurred in association with rpoB S450L (p < 0.001) (S5 Table). This pattern was observed regardless of whether the compensatory mutation was previously known or uncharacterized. Beyond rifampicin compensation, we also applied our combined phylogenetic and parsimony approach to known isoniazid and ethambutol compensatory mutations. With respect to isoniazid compensation, only a single evolution of the ahpC promoter mutation was observed in our dataset (S6 Table). It was gained after genotypic isoniazid resistance, which supports its compensatory role for certain isoniazid resistance mechanisms [41]. Nonsynonymous mutations in ubiA (Rv3086c) have previously been implicated in ethambutol resistance [56]. In our dataset, there were at least two occasions in which these mutations unambiguously arose prior to the acquisition of genotypic ethambutol resistance, suggesting that these are more likely to be stepping-stone mutations rather than compensatory (S6 Table). Discussion We report on the WGS and comparative analysis of the largest collection of drug-resistant M. tuberculosis sequenced to date from South Africa. From analysis of these genomes, we determined the molecular antecedents of the Tugela Ferry XDR Clone and dated the emergence of genotypic resistance to eight drugs. We showed that the development of XDR in KwaZulu-Natal had its roots in first-line drug resistance that arose in the late 1950s and MDR that emerged in the 1980s. Our dating analysis indicated that the Tugela Ferry XDR Clone took nearly four decades to evolve from its initial isoniazid and streptomycin resistances to full-blown XDR. Although our data confirmed that the XDR outbreak in KwaZulu-Natal was indeed a clonal event, we showed that drug resistance in this region is driven by both the development of de novo drug resistance and clonal spread. We elucidated common evolutionary patterns of drug resistance acquisition and determined that isoniazid was overwhelmingly the first drug resistance to be acquired. Lastly, we validated that certain previously described rifampicin compensatory mutations do indeed evolve in a pattern consistent with compensation and have identified 26 novel polymorphisms that may also function in this capacity. Collectively, these data have important implications for the public health control of TB in sub-Saharan Africa and elsewhere. Using a combination of likelihood, parsimony, and Bayesian computational approaches, we observed a decades-long evolutionary trajectory toward XDR-level drug resistance in LAM4 that mirrored the order and timing of introduction of antitubercular drugs into clinical practice [57]. Though parsimony-based approaches can interpret rapid independent evolution of an identical polymorphism in multiple strains as a single evolutionary event (occurring at a single node), our predictions indicated that resistance-conferring mutations evolved only after each drug’s clinical introduction and not before, as might be expected if homoplasy were a major contributor to pattern predictions (Fig 3). In addition, one of the oldest acquired mutations toward XDR-level resistance was a specific 130 bp deletion in gidB, which is extremely unlikely to arise many times independently and supports accurate reconstruction in our evolutionary analysis. Furthermore, though our calculated mutation rate for LAM4 was slightly higher than previous reports [39,41,46–48], our estimate was within the reported 95% HPD interval and was based on a larger fraction of the H37Rv genome than previous studies (99.9% versus <90% H37Rv mapping coverage) [39]. This was due to the inclusion of sequence data generated from both PCR-free short fragment and jumping libraries and analysis with improved bioinformatics tools that enabled us to examine SNPs within more variable and high guanine-cytosine (GC) content regions of the genome, including proline-glutamic acid (PE) and proline-proline-glutamic acid (PPE) genes that have been reported to have a higher mutation rate [58]. Thus, because we are including data from more of the genome, our estimation of the M. tuberculosis mutation rate may more closely approximate the actual mutation rate of the organism as compared to previously published studies. Importantly, from the pattern of drug resistance evolution within LAM4, it is clear that the precursors to XDR evolved well before the explosive South African HIV epidemic of the 1990s, indicating that the selection of transmissible XDR strains can occur in low-prevalence HIV settings. While recent failures in TB and infection control and the current high HIV prevalence rates, combined, undoubtedly contributed to the spread of XDR, they were not the sole causes of XDR in this setting. Indeed, strains that evolved first-line drug resistance soon after the introduction of chemotherapy were a critical entry point to today’s drug-resistant epidemic. Drug-resistant strains that emerged from the mid-20th century were evidently maintained within the population of M. tuberculosis, presenting the opportunity for the acquisition of successive resistance and compensatory mutations that culminated in transmissible XDR and the Tugela Ferry outbreak. Drug-resistant strains may have been maintained within a population over time either by ongoing cycles of infection and transmission or through reactivation of latent disease. It is unclear which of these may have been the most important in this setting, but it suggests that fitness costs due to first-line drug resistance may not be severe. Reactivation was recently shown to be important in the transcontinental spread of MDR-TB from Thailand to California over a 22-y period [59], but it is unclear whether this factor was also critical in KwaZulu-Natal. Beyond LAM4, and as has been shown in other studies [10,14], drug resistance emerged de novo repeatedly in KwaZulu-Natal, as evidenced by our identification of numerous independent evolutionary events of MDR and XDR across multiple lineages and spoligotypes. Of particular note was the detection of multiple independent evolutions of MDR to pre-XDR within LAM4, which may herald a new wave of XDR in the near future. Thus, the repeated emergence of de novo high-level drug resistance underscores the reality that, even in middle-income sub-Saharan African countries, the current approach to TB control is failing to stem the ongoing emergence of drug resistance. In fact, results from our analyses suggest this was not due to infrequent poor adherence to TB drugs but instead to decades of inadequate TB control that has driven resistance development in a stepwise fashion, multiple times over. Given that our estimates of resistance evolution were based on identification of known resistance-conferring mutations and that the majority of sequenced strains were single colony purified, our calculations are likely an underestimation due to incomplete understanding of all mutations that confer drug resistance and the possibility of mixed infections, respectively. Thus, the state of drug resistance emergence is likely more dire than we have described. Recent studies from KwaZulu-Natal have emphasized transmission of a limited number of strains as a driving force behind the emergence of drug resistance [10,14]. Our data also confirm that once drug resistance develops, clonal spread of resistant strains can and does occur in this context. We found that recent person-to-person spread of resistant strains is apparent in KwaZulu-Natal, as evidenced by identification of multiple drug-resistant clones. Importantly, in contrast to initial reports from Tugela Ferry in which nearly all XDR cases were TB/HIV coinfected [7,9], eight patients in our study in whom the Tugela Ferry XDR Clone was identified were HIV negative. This reemphasizes that even XDR drug-resistant strains are sufficiently fit to transmit person to person and cause morbidity in both immunocompetent and immunosuppressed persons. Improved infection control and rapid case finding will be necessary to prevent further spread of drug-resistant strains and to detect such cases in the community as well as in hospital settings [28]. Our genomic analysis uncovered a common initial pattern of drug resistance that is not optimally detected by current diagnostic algorithms. Isoniazid resistance was overwhelmingly the first drug resistance to occur along the pathway to multiple drug resistances. However, current TB control strategies in South Africa focus on early detection of rifampicin resistance as a surrogate marker of MDR and do not include the detection of isoniazid resistance. Clinical diagnostic policies that rely on Xpert MTB/RIF (a WHO-endorsed and widely deployed molecular diagnostic) [60] without more extensive drug resistance testing allow isoniazid resistance to go undetected and unchecked. Moreover, under current short course treatment guidelines that utilize 4 mo of isoniazid and rifampicin in the continuation phase [61], failure to recognize isoniazid monoresistance is tantamount to provision of unopposed rifampicin therapy and may rapidly select for rifampicin resistance. This phenomenon may be underappreciated and incompletely accounted for in mathematical models that recommend continued use of screening tools that identify only rifampicin resistance [62]. Furthermore, if rifampicin resistance is indeed detected by Xpert, failure to implement confirmatory secondary molecular testing for dual rifampin and isoniazid resistance, as is mandated by South African policy, occurs at unacceptably high rates [63]. Our ordering of drug resistance acquisition provides strong evidence that isoniazid monoresistance is a common pathway toward development of MDR and highlights the importance of prompt identification and treatment of isoniazid monoresistance. Failure to do so would be recapitulating the scenario that led to the current XDR problem. Beyond detection, identification of the initial drivers of isoniazid monoresistance is also critical to the prevention of successive resistances. Isoniazid preventive therapy (IPT) has previously been implicated as a potential source of isoniazid monoresistance [64,65]. Our work highlights the need to understand the true risks of mass IPT implementation [66] in high-burden settings. We were able to verify that the evolutionary patterns of select previously described rifampicin and isoniazid compensatory mutations do indeed appear to be consistent with compensation to their respective drug. Similarly, ubiA was observed to evolve in a stepping-stone pattern rather than a compensatory pattern with respect to ethambutol resistance [56]. Furthermore, we have identified novel putative rifampicin compensatory mutations that may have acted to restore bacterial fitness and facilitate transmission of drug-resistant strains. While the majority of the previously described rifampicin compensatory mutations had an evolutionary pattern consistent with this role, four polymorphisms previously associated with rifampicin compensation (rpoB I491F, rpoC G594E and N826K, and rpoA E319K) were not observed to evolve concurrently or subsequent to genotypic rifampicin resistance (S4 Table). These mutations may (i) not be compensatory mutations in the classic sense (i.e., mutations that evolve following gain of genotypic drug resistance to mitigate a fitness cost) but instead serve as stepping-stone mutations, (ii) evolve in concert with non-RRDR genotypic rifampicin resistance, or (iii) have no association with rifampicin resistance. We have proposed 26 novel mutations whose evolutionary patterns are consistent with rifampicin compensation, and these should be investigated in future studies. The most commonly observed genotypic rifampicin resistance mutation among our sequenced strains was rpoB S450L (often referred to as S531L using the Escherichia coli codon numbering scheme), which is known to be the most prevalent RRDR mutation. Laboratory-derived strains carrying the S450L were previously shown to have relatively high fitness in in vitro growth assays [15], supporting the hypothesis that high prevalence of the S450L mutation among clinical strains was due to it imparting few fitness consequences. However, as shown in our study and in several others [20,54], rpoB S450L was the most likely RRDR polymorphism to evolve putative compensatory mutations, which calls into question the low fitness cost of S450L in vivo. Song et al. assessed rifampicin fitness by transcriptional efficiency (rather than growth) and showed that the S450L mutation has half the transcriptional efficiency of WT rpoB [54], which is likely to impart fitness consequences if not compensated. This study has two main limitations. First, as our study isolates derived from only one geographic region, our conclusions regarding the timing and dating of the emergence of resistance may not be universal. However, two recent studies [67,68] have reported results compatible with ours from different settings. Using a similar approach, Eldholm et al. were able to date the first emergence of resistance in an MDR-TB outbreak from Argentina to the early 1970s and found that isoniazid and streptomycin resistance-conferring mutations were the first to be acquired. In a study of the global spread of the Beijing lineage [23], isoniazid and streptomycin resistances were also found to be common to all drug-resistant strains in two clonal complexes that resulted in the epidemic spread of two MDR clones in Russia and Central Asia 20 to 30 y ago. Another limitation, as discussed above, is that parsimony-based dating approaches may fail to distinguish rapid independent evolutions of a commonly occurring resistance mutation as two unique evolutionary events. This could lead to erroneous assignment of a mutation to a more basal part of the phylogenetic tree. While a theoretical risk, we believe the effect was minimal in our dataset since, as described above, our predictions were consistent with the timing of drug introduction, and we included a specific large deletion that is extremely unlikely to arise many times independently. Here, we present the largest WGS study conducted to date of drug-resistant clinical isolates of M. tuberculosis from South Africa. Our dating analysis highlights the dire repercussions of failure to control first-line drug resistance. As acquisition of isoniazid resistance is the key initiation event for progression to MDR and beyond, TB control efforts that focus on the identification of isoniazid as well as rifampicin resistance will result in earlier detection of drug-resistant TB cases. Prudent antibiotic stewardship during the introduction of new antitubercular drugs will be critical to prevent the early fixation of resistance and protect the lifespan of novel agents. Supporting Information S1 Fig. Bootstrap values for phylogenetic tree of 337 strains (1,000 bootstrap replicates). https://doi.org/10.1371/journal.pmed.1001880.s001 (PDF) S2 Fig. Defining clones with varying SNP thresholds. The numbered columns to the right of the phylogenetic tree represent varying SNP thresholds used to define a clone. Strains that would be considered clonal by the SNP threshold listed in the column header are indicated by a unique three-letter code. By the ten-SNP threshold, the Tugela Ferry XDR Clone (labeled 10-AAI, shaded in gray) contains 50 members. https://doi.org/10.1371/journal.pmed.1001880.s002 (PDF) S3 Fig. Drug-resistant clones are distributed widely across the phylogenetic tree. Columns to the right of the phylogenetic tree represent phenotypic DST (as indicated by the colored square), clones defined at the ten-SNP threshold as shown in S2 Table and S2 Fig, and the HIV status of sampled patient. https://doi.org/10.1371/journal.pmed.1001880.s003 (PDF) S4 Fig. Bootstrap values for phylogenetic tree of LAM4 isolates shown in Fig 3 (1,000 bootstrap replicates). https://doi.org/10.1371/journal.pmed.1001880.s004 (PDF) S1 Methods. Microbiologic techniques; digital spoligotyping; clonal strain identification; genotypic definitions of drug resistance, accessory and compensatory mutations; and statistical tests. https://doi.org/10.1371/journal.pmed.1001880.s005 (DOCX) S1 Table. Participant data. Each participant was assigned a strain number with the header Tuberculosis KwaZulu-Natal K-RITH (TKK). Data for each participant included year of collection, specimen type, and smear status (if known). DNA isolation technique via single colony isolation (SCI) or non-single colony selection (non-SCI) is denoted. DST results are reported for each tested drug using the following abbreviations: rifampicin (R), isoniazid (H), nicotinamide (N), pyrazinamide (P), ethambutol (E), streptomycin (S), kanamycin (K), ofloxacin (O), ethionamide (Et), and capreomycin (C). DST results are noted as susceptible (S), resistant (R), or untested (U). Genomic spoligotyping and lineage that were derived from the sequencing data are listed. Lastly, genotypic drug susceptibility prediction and membership in the Tugela Ferry Clone are reported. https://doi.org/10.1371/journal.pmed.1001880.s006 (PDF) S2 Table. Identification of drug-resistant clones indicates recent person-to-person transmission of drug-resistant TB. A linkage analysis identified 11 drug-resistant clones in the entire dataset. The largest clone contained 50 members of the LAM4 spoligotype; this spoligotype was subsequently identified as the Tugela Ferry XDR Clone. Within the LAM4 spoligotype, there were three additional clones identified, and clones were also identified in five other spoligotypes. All clone members were noted to be drug resistant, indicating recent person-to-person transmission of drug-resistant TB. See S1 Methods for definition of a clone. https://doi.org/10.1371/journal.pmed.1001880.s007 (PDF) S3 Table. Diverse drug-resistant strains and frequent de novo development of drug resistance. Drug-resistant strains belonged to many distinct spoligotypes, which highlights the diversity of the drug resistance epidemic in this region. With a parsimony-based analysis, we quantified the independent evolutionary gains of genotypic MDR and XDR in our 340-strain dataset. https://doi.org/10.1371/journal.pmed.1001880.s008 (PDF) S4 Table. Putative rifampicin compensatory mutations were identified in rpoA, rpoC, and non-RRDR regions of rpoB. Polymorphisms were deemed consistent with compensatory mutations when they evolved after or concurrent to genotypic rifampicin resistance. Many previously described putative compensatory mutations occurred in this evolutionary pattern, and 26 novel polymorphisms were newly described. https://doi.org/10.1371/journal.pmed.1001880.s009 (XLSX) S5 Table. Distribution of putative rifampicin compensatory mutations across the RRDR. The vast majority of rpoA, rpoC, and non-RRDR rpoB mutations that evolved with an evolutionary pattern consistent with rifampicin compensation evolved in association with rpoB S450L. https://doi.org/10.1371/journal.pmed.1001880.s010 (XLSX) S6 Table. Ordering of acquisition of polymorphisms with respect to genotypic resistance. ahpC and ubiA mutations were ordered with respect to genotypic isoniazid and ethambutol resistance, respectively. https://doi.org/10.1371/journal.pmed.1001880.s011 (PDF)

doi: 10.1371/journal.pmed.1001864pmid: 26393914

Summary Points The evidence base underpinning clinical practice is deeply flawed. There must be better value gained from resources invested in medical research. We make four proposals: (1) introducing publications officers; (2) developing core competencies for editors and peer reviewers, around which (3) training can be tailored; and (4) training authors to write articles fit for purpose. All of these ideas need to be piloted and evaluated, and implemented if proven effective. We suggest dedicated funding for initiatives aimed at understanding and improving the way that research is conducted and published. Academic institutions, funders, publishers, and others should support and implement effective processes to improve the reliability of the medical research literature. Introducing Publications Officers Universities invest resources at the front end of knowledge generation. Many academic centres employ professionals to help their researchers understand the process of successfully competing for a dizzying number and different types of research applications. In Canada, for example, these people are often known as grants officers or technology transfer officers. Their primary objective is to provide direction, guidance, and timely information to the institute’s scientists relating to grant submissions. After the research monies are secured and the research is conducted, the findings are ready for dissemination through presentations at scientific conferences and journal publications. Unfortunately, this dissemination model is not fully effective. For example, too much research is never published [14,15]. And many research reports that are published display important weaknesses [16]. To help rectify this situation, we propose the introduction of publications officers, who would support and educate researchers, staff, and trainees in universities and research organisations. Their roles and responsibilities could include providing guidance on preparing manuscripts for submission to journals (including adherence to relevant reporting guidelines and the submission procedures); developing seminars on how to write to get published—that is, writing articles that are fit for purpose [17]; harnessing existing resources relevant to manuscript preparation and publication, addressing research integrity and publication ethics; and facilitating internal peer review of manuscripts before submission to journals. Other activities might include facilitating in-depth training on using reporting guidelines when preparing manuscript submissions, regular seminars on issues about publication ethics and research integrity and responsibility, explaining open access options, and providing seminars to the local community on “making sense of science,” such as how authors use “spin” to interpret the results of their research [18]. Some organizations, for example, the International Society for Medical Publication Professionals (http://www.ismpp.org/), provide resources for professional medical writers. They also offer certification (Certified Medical Publication Professional). But there seem to be few comparable opportunities for academic researchers. Currently there is an inequity in academic institutional thinking—great interest in maximizing the chances of succeeding in grant applications, yet little attention given to maximizing success when the research project has been completed, namely, dissemination, although universities typically employ press officers (communications officers) to maximise opportunities of the media reporting on their scientists’ research. Publications are the tangible output from all of the research activity, so they surely merit serious investment to ensure complete and transparent reports. We believe that introducing publications officers within academic institutions would help to reduce nonpublication and selective publication of research findings, improve the clarity and transparency of the institution’s research output, and help raise the quality and value of their researchers’ publications. There is, as yet, little experience regarding the ideal publications officer. Backgrounds in education, clinical epidemiology, medical writing, research ethics, or a combination of these would seem appropriate. One of our institutions (Ottawa Hospital Research Institute) very recently employed a publications officer with a background in biology and psychology. As part of this early pilot endeavour, the institution is conducting an evaluation of the effectiveness of the position. Core Competencies Editors Scientific editors (and ultimately, editors-in-chief) are accountable for all published material in their journals. Readers should expect them to have processes in place to ensure the quality of the papers they publish and to strive constantly to improve their journals. While well-resourced medical journals [19] have full-time, paid, professional scientific editors, and their publishers may have resources to provide some formal training for their position, the majority of medical journal editors work on a voluntary basis. Such “pro bono” activities are useful if the scientific activities associated with being an editor are of the highest possible standards. Unfortunately, many medical editors who oversee their journals are largely untrained and certainly uncertified. We think this is not the optimal way to instil confidence in readers, provide value for money to funders, or ensure the public can trust the research record. Some organizations, for example, the World Association of Medical Editors (WAME), provide resources for editors. There are some good websites, such as Committee on Publication Ethics (COPE; http://publicationethics.org/), and blogs, such as Journalology (http://journalology.blogspot.ca/), that provide important information for editors. There are also several short courses on being an editor offered by commercial groups (http://www.pspconsulting.org/medical-short.shtml) and a few large, well-resourced journals offer in-house training for editors (e.g., The BMJ). For a substantial editor training program to work optimally, it must be based primarily on what the broad medical editor community considers to be core competencies for all editors. Other stakeholders need to contribute to this effort, such as publishers, peer reviewers, and authors (researchers). We are unaware, however, of any body of literature identifying what these core competencies are [20]. Given one recent recommendation to use reporting guidelines [16], a core competency might be for editors to have a more thorough knowledge of them, including how best to endorse and implement them and facilitate their use by peer reviewers [20]. One of us is leading the development of a core competency for editors program, one result of which will be a minimum set of evidence-based core competencies. The program has several elements similar to how some reporting guidelines have been successfully developed [21,22]. Authors Any efforts aimed at training authors might be considered too late in the knowledge generation cycle. The research has been completed and all that can be done, realistically, is to try to ensure that authors transparently and completely tell readers what they did (methods) and found (results). Recent examples of using guidelines earlier, during the design and conduct of research, include SPIRIT [23] for preparing protocols of randomized trials and PRISMA-P for preparing protocols of systematic reviews [22]. Also, some groups are focused on helping researchers to improve the design and conduct of clinical trials, such as the Core Outcome Measures in Effectiveness Trials (COMET) [24] and TrialForge [25]. Even if such initiatives succeed in improving trial methods, training researchers to write high-quality articles will remain an important and relevant goal. As authors, many researchers have difficulty reporting what they did and found, completely and transparently. Authors need to write papers that are fit for purpose [17]. They need to ensure that every report that includes their name is a completely reported and transparent account of what was done and found, to enable interested readers to replicate the methods and use the results [13]. Collectively, authors are currently not doing a good job reporting their research [16]. If the introduction of publications officers, described above, proves successful, this is one way institutions can help ensure that the manuscripts submitted for publication consideration are the highest quality. This might help reduce waste and increase journal efficiencies, including the peer review process. For many researchers, writing is difficult. It is an acquired skill that often starts during graduate studies, at which time it should be formally taught and discussed. Developing good writing skills early on can reap benefits throughout a researcher’s career. Formal training in writing, use of reporting guidelines, and issues related to authorship (e.g., attributing authorship, authorship order, author responsibilities) should be mandatory [26]. It is unfortunate that most universities do not promote this type of formal training, given the evidence of the tarnished published literature [1–11,16]. Editors Scientific editors (and ultimately, editors-in-chief) are accountable for all published material in their journals. Readers should expect them to have processes in place to ensure the quality of the papers they publish and to strive constantly to improve their journals. While well-resourced medical journals [19] have full-time, paid, professional scientific editors, and their publishers may have resources to provide some formal training for their position, the majority of medical journal editors work on a voluntary basis. Such “pro bono” activities are useful if the scientific activities associated with being an editor are of the highest possible standards. Unfortunately, many medical editors who oversee their journals are largely untrained and certainly uncertified. We think this is not the optimal way to instil confidence in readers, provide value for money to funders, or ensure the public can trust the research record. Some organizations, for example, the World Association of Medical Editors (WAME), provide resources for editors. There are some good websites, such as Committee on Publication Ethics (COPE; http://publicationethics.org/), and blogs, such as Journalology (http://journalology.blogspot.ca/), that provide important information for editors. There are also several short courses on being an editor offered by commercial groups (http://www.pspconsulting.org/medical-short.shtml) and a few large, well-resourced journals offer in-house training for editors (e.g., The BMJ). For a substantial editor training program to work optimally, it must be based primarily on what the broad medical editor community considers to be core competencies for all editors. Other stakeholders need to contribute to this effort, such as publishers, peer reviewers, and authors (researchers). We are unaware, however, of any body of literature identifying what these core competencies are [20]. Given one recent recommendation to use reporting guidelines [16], a core competency might be for editors to have a more thorough knowledge of them, including how best to endorse and implement them and facilitate their use by peer reviewers [20]. One of us is leading the development of a core competency for editors program, one result of which will be a minimum set of evidence-based core competencies. The program has several elements similar to how some reporting guidelines have been successfully developed [21,22]. Authors Any efforts aimed at training authors might be considered too late in the knowledge generation cycle. The research has been completed and all that can be done, realistically, is to try to ensure that authors transparently and completely tell readers what they did (methods) and found (results). Recent examples of using guidelines earlier, during the design and conduct of research, include SPIRIT [23] for preparing protocols of randomized trials and PRISMA-P for preparing protocols of systematic reviews [22]. Also, some groups are focused on helping researchers to improve the design and conduct of clinical trials, such as the Core Outcome Measures in Effectiveness Trials (COMET) [24] and TrialForge [25]. Even if such initiatives succeed in improving trial methods, training researchers to write high-quality articles will remain an important and relevant goal. As authors, many researchers have difficulty reporting what they did and found, completely and transparently. Authors need to write papers that are fit for purpose [17]. They need to ensure that every report that includes their name is a completely reported and transparent account of what was done and found, to enable interested readers to replicate the methods and use the results [13]. Collectively, authors are currently not doing a good job reporting their research [16]. If the introduction of publications officers, described above, proves successful, this is one way institutions can help ensure that the manuscripts submitted for publication consideration are the highest quality. This might help reduce waste and increase journal efficiencies, including the peer review process. For many researchers, writing is difficult. It is an acquired skill that often starts during graduate studies, at which time it should be formally taught and discussed. Developing good writing skills early on can reap benefits throughout a researcher’s career. Formal training in writing, use of reporting guidelines, and issues related to authorship (e.g., attributing authorship, authorship order, author responsibilities) should be mandatory [26]. It is unfortunate that most universities do not promote this type of formal training, given the evidence of the tarnished published literature [1–11,16]. Training Peer Reviewers Evidence of the effectiveness of peer review of medical research articles is minimal at best [27]. A recent study found that peer reviewers identified only a fifth of the reporting deficiencies in a cohort of randomised trials [28]. Hundreds of reviews of published articles have shown that peer reviewers fail to detect widespread methodological errors and reporting deficiencies [5–11]. Such dismal results have led some to argue that we should abandon peer review completely [29]. Teaching peer review is probably one of the most important ways to increase trust and confidence in the published research record. There are some commercial short courses on peer review, but to the best of our knowledge there is minimal formal teaching of peer review in academic institutions, the very places where a large amount of research originates. This is also where the next generation of researchers are being trained and cultivated. Furthermore, most journals fail to provide guidance to reviewers on what they expect from them [30]. There are several motivations for completing peer review, including requests from a supervisor, wanting to keep abreast of the latest developments in a specific content area, and altruism. Regardless of the reason for doing peer review, most reviewers do it without training or reward. Most people who perform peer review learn by trial and error and perhaps some mentorship, as we did. There has been a recent focus on the need to provide comprehensive peer reviewer training and view it more professionally [31]. We believe that to make peer review more effective requires the development of a set of core competencies, after which training programs can be developed. An approach similar to that described for developing core competencies for editors, described above, deserves serious consideration. Such training could be integrated into a broad training curriculum within universities. The objective of formal training is to provide students with the skill set needed to detect manuscripts that are not fit for purpose and help authors to improve them. These skills learned as a peer reviewer can also be used when the peer reviewers write complete and transparent articles as authors—how their study was conducted and what they found. Additionally, such courses need to teach about the role and responsibilities of peer reviewers. Academic institutions need to take peer review seriously and develop full- or half-semester courses that can be used by students towards their degree. Sufficient institutional resources need to be set aside to ensure these courses can be appropriately developed. These courses should be mandatory for all new graduate students and young researchers. Funding One estimate is that US$240 billion is spent globally, every year, on health research [32]. The outputs from this research are documented in about 3 million articles, of which about half are published by 6,000 publishers in 25,000 journals (with a much larger number of editors). We have briefly described some of the serious problems associated with the published record. In many cases, the information reported cannot be used; in many more, the reporting is biased; and much research is never published at all. It is estimated that 85% [32] of the global investment is avoidable waste (i.e., it is modifiable). This is a bad return on such a large fiscal investment, particularly when there are substantial public monies involved. The status quo is unacceptable, yet there are few signs that major bodies (funders, universities, professional bodies) recognise the need for change. Just as the problems we have discussed here are large, complex, and not the sole responsibility of any single group, no single stakeholder can or should fund “journalology” (research on research) investigations. That said, funding agencies and others, particularly publishers, are likely central in helping to promote and support the development of sustained programs of investigation in journalology. Perhaps it is possible to start with leadership and support from heads of major funding organizations and large publishers heavily invested in publishing biomedical research. Once such an initial commitment is secured, other relevant groups may join and support the collective effort. It is unclear how best to leverage academic institutions to join this funding effort. One possibility is for them to commit in-kind resources for faculty to have time to develop and pilot these and other similar initiatives. A very small fraction of funders’ and publishers’ expenditures (say 0.1%) could be set aside for initiatives to reduce waste and improve the quality, and thus value, of research publications, including journalology investigations; it is a legitimate and arguably essential research endeavour. Some small percent of these monies could be used to fund the certification and continuing education training for editors, as well as training for authors and peer reviewers. Some of the investment could be targeted to reach attainable increases in research value, annually, over the next decade. The precise increase could be agreed upon by key players. In conclusion, publishing medical research is complex; the biomedical research community is failing at it intellectually, fiscally, morally, and ethically [33]. The present state of research publication is unacceptable. We have made four proposals here to help improve the situation, complementing other proposals [19,34]. For these initiatives to succeed, there must be a fundamental shift in how our academic institutions, funding bodies, journals, and publishers perceive the importance of publication practices. Working together, these organizations can help test and potentially implement our proposals and indicate the importance of these and other similar efforts. Ensuring high-quality research publications must become a core activity within their portfolios. The medical publication business needs to be taken much more seriously. Everybody deserves a guarantee of reliable evidence resulting from our research endeavours.

doi: 10.1371/journal.pmed.1001877pmid: 26372602

Mycobacterium tuberculosis often develops resistance in the setting of monotherapy, either de facto (when an organism is susceptible to only one drug of an intended multidrug regimen) or actual (historically, or in the setting of extensively drug resistant tuberculosis (XDR-TB) salvage regimens) [1–4]. With current regimens, sterilization of drug-susceptible organisms requires at least six months of treatment to prevent disease relapse. Unfortunately, pill burden, drug toxicity, stigma, and poor provider practice often complicate these prolonged treatment courses, and nonadherence is widely blamed for the global epidemic of drug-resistant TB. Directly observed therapy (DOT) has for decades been considered crucial to ensuring anti-TB medication adherence worldwide, but carries important individual and health policy concerns. This week in PLOS Medicine, Fielding and colleagues present a cluster randomized trial of an alternative strategy—managing adherence with reminders delivered by an electronic pillbox or text messaging. DOT is variously defined as 5–7 daily doses per week observed by facility-, workplace-, or community-based healthcare workers or confidants. Although DOT is employed selectively for other communicable and noncommunicable diseases for which incomplete adherence threatens treatment success, DOT in TB treatment has become enshrined as “canon in a field long characterized by fervor in principle and practice” as in no other disease [5]. First implemented in Madras and Hong Kong in the 1950s [6], DOT became globally endorsed (as one of five key components of the World Health Organization directly observed therapy short-course [DOTS] strategy) [7] in 1994 in the aftermath of public health alarm created by an outbreak of multidrug-resistant (MDR)-TB in New York City [8]. Operationalization of DOT varies according to resource availability, cultural factors, and individual provider perception, and is often incomplete, even among populations for which strict adherence is considered essential [9]. Although in its best form DOT can be a platform for patient social support and guidance, it has long been debated whether the requirement for witnessed dosing represents the least restrictive alternative in pursuit of public health goals [10]. Conceived of as a principal standard of care to protect individuals from drug resistance amplification and communities from a potentially devastating airborne disease, DOT in some cases adversely impacts the dignity, autonomy, and livelihoods of patients who are often already poor and disenfranchised [11]. For example, among patients receiving treatment for both HIV and drug-resistant TB, loss of a sense of agency due to provider supervision of TB treatment contributes to preferential adherence to antiretrovirals over drug-resistant TB therapy, complicating combined treatment regimens that may surpass 30 pills daily [12]. Nevertheless, alternative strategies to DOT have received little attention. To what degree does nonadherence lead to treatment failure or acquisition of drug resistance? In contrast to HIV, for which the complex relationship between adherence, pharmacokinetics, and resistance for each antiretroviral class is defined [13], the levels and patterns of adherence that lead to TB treatment failure and drug resistance remain largely unknown [14]. Observational studies from the United States [15] and Botswana [16] in the 1990s documented associations between DOT-supported adherence and prevention of drug resistance. However, a meta-analysis of randomized trials [17] comparing DOT to self-administration of TB treatment failed to demonstrate improved treatment outcome (though the component trials were heterogeneous [18] and were not powered to examine drug resistance amplification). In addition, preclinical studies suggest that TB treatment is robust to relatively high levels of nonadherence [19], and certain drugs and drug combinations are more or less forgiving than others [4,20]. Nevertheless, adherence challenges with prior TB treatment are a common refrain among people living with MDR-TB [21]. The study by Fielding and colleagues was a four-arm cluster randomized trial, involving 4,173 patients in 36 health centers across China, designed to assess the effect of dosing and refill reminders delivered by cell phone text message, electronic pillbox, or both, on patient adherence to six months of intermittent (every other day) TB treatment. The primary outcome was the proportion of patient-months in which ≥20% of doses were missed (“nonadherent months”), measured by either monthly pill counts or electronic pillbox openings (which did not produce audible reminders in the control arm of the study). Despite use of medication pillboxes, various forms of DOT, and data-informed (pill counts) counseling in the control arm, substantial nonadherence was observed (30% nonadherent months). Electronic pillbox reminders, but not text messages alone, significantly decreased the primary outcome (17% nonadherent months). This effect was stronger when text message reminders were added to pillbox reminders (14% nonadherent months). Approximately one-tenth of patients were excluded because of inability to use mobile phones after training, and nearly one-third of randomized patients (and one-half of those in the combined text messaging and electronic pill box arm) experienced technical problems with their device (S5 Table in the supplementary material of the paper by Fielding and colleagues). Despite substantial nonadherence to an intermittent regimen, adverse end-of-treatment outcomes, other than loss to follow up, were rare in this low HIV burden setting. Significant technical complications in the trial related to battery connectivity suggest that additional work is needed to both confirm benefit and facilitate scale-up, and the data provide little indication of whether durable effects on treatment outcomes (beyond loss to follow-up) should be anticipated. Yet, the study by Fielding and colleagues is unusual in providing objective TB treatment adherence data combined with a scalable intervention strategy to improve adherence. If replicated, it will have important implications for global TB treatment in moving away from witnessed dosing, which is not universally feasible, towards a more personalized adherence model of patient–provider communication in which intervention is delivered where, when, and in whom it is needed to efficiently prevent adverse treatment outcome. Will such technology aid transition to a “post-DOT” era in low- as well as high-income settings? In the field of HIV, “just-in-time” adherence interventions linked in real-time to objectively monitored adherence have overcome several limitations in traditional counseling-based interventions [22], which cannot anticipate decline in adherence over time or interruptions in daily adherence routines. Such approaches can potentially reduce healthcare costs by selectively targeting other medical resources (e.g., laboratory monitoring, provider visits, pharmacokinetic testing) to patients at greatest risk, while foregoing these for patients with near-perfect adherence and negligible risk of disengagement, treatment failure, or drug resistance [23]. In its long history with humanity, tuberculosis has in many ways provided the prototype for chronic disease management. The cornerstone of successful TB treatment has long been recognized to lie within the complex inter-relationship between patients and clinical staff and to hinge strongly on structural facilitators to the treatment experience, along with empathy [24]. Shortened, simplified TB treatment regimens in conjunction with technologies facilitating adherence, such as real-time adherence monitoring and novel drug-delivery systems, could support this foundation to speed TB elimination while delivering truly patient-centered care.

doi: 10.1371/journal.pmed.1001873pmid: 26348035

Background Home-based HIV testing and counselling (HTC) achieves high uptake, but is difficult and expensive to implement and sustain. We investigated a novel alternative based on HIV self-testing (HIVST). The aim was to evaluate the uptake of testing, accuracy, linkage into care, and health outcomes when highly convenient and flexible but supported access to HIVST kits was provided to a well-defined and closely monitored population. Methods and Findings Following enumeration of 14 neighbourhoods in urban Blantyre, Malawi, trained resident volunteer-counsellors offered oral HIVST kits (OraQuick ADVANCE Rapid HIV-1/2 Antibody Test) to adult (≥16 y old) residents (n = 16,660) and reported community events, with all deaths investigated by verbal autopsy. Written and demonstrated instructions, pre- and post-test counselling, and facilitated HIV care assessment were provided, with a request to return kits and a self-completed questionnaire. Accuracy, residency, and a study-imposed requirement to limit HIVST to one test per year were monitored by home visits in a systematic quality assurance (QA) sample. Overall, 14,004 (crude uptake 83.8%, revised to 76.5% to account for population turnover) residents self-tested during months 1–12, with adolescents (16–19 y) most likely to test. 10,614/14,004 (75.8%) participants shared results with volunteer-counsellors. Of 1,257 (11.8%) HIV-positive participants, 26.0% were already on antiretroviral therapy, and 524 (linkage 56.3%) newly accessed care with a median CD4 count of 250 cells/μl (interquartile range 159–426). HIVST uptake in months 13–24 was more rapid (70.9% uptake by 6 mo), with fewer (7.3%, 95% CI 6.8%–7.8%) positive participants. Being “forced to test”, usually by a main partner, was reported by 2.9% (95% CI 2.6%–3.2%) of 10,017 questionnaire respondents in months 1–12, but satisfaction with HIVST (94.4%) remained high. No HIVST-related partner violence or suicides were reported. HIVST and repeat HTC results agreed in 1,639/1,649 systematically selected (1 in 20) QA participants (99.4%), giving a sensitivity of 93.6% (95% CI 88.2%–97.0%) and a specificity of 99.9% (95% CI 99.6%–100%). Key limitations included use of aggregate data to report uptake of HIVST and being unable to adjust for population turnover. Conclusions Community-based HIVST achieved high coverage in two successive years and was safe, accurate, and acceptable. Proactive HIVST strategies, supported and monitored by communities, could substantially complement existing approaches to providing early HIV diagnosis and periodic repeat testing to adolescents and adults in high-HIV settings. Background Every year, about 2.1 million people (70% of whom live in sub-Saharan Africa) are newly infected with HIV, the virus that causes AIDS, and 1.5 million people (again, mainly in sub-Saharan Africa) die as a result. HIV, which is usually transmitted through unprotected sex with an infected individual, gradually destroys CD4 lymphocytes and other immune system cells, leaving HIV-positive individuals susceptible to other serious infections and to unusual cancers. HIV is diagnosed by looking for antibodies to HIV in blood or saliva. After diagnosis, the progression of HIV infection is monitored by regularly counting the number of CD4 cells in the blood. Initiation of antiretroviral therapy—a combination of drugs that keeps HIV replication in check but that does not cure the infection—is recommended when an individual’s CD4 count falls below 500 cells/μl or when he or she develops an AIDS-defining condition. Why Was This Study Done? HIV-positive individuals need to know their status so that they can take steps to avoid transmitting the virus to other people (for example, by always using a condom during sexual intercourse) and so that they can begin treatment. Treatment helps to keep HIV-positive individuals healthy but also reduces their chances of transmitting the virus to their sexual partners. Unfortunately, many HIV-positive individuals are unaware of their status. The situation is particularly bad in sub-Saharan Africa, where, despite major investments in facility-based and community-based HIV testing and counseling (HTC) programs, only a quarter of adults have had a recent HIV test, and only half of the people living with HIV know they are HIV positive. Barriers to facility-based HTC include concern about lack of confidentiality and fears of stigmatization. Home-based HTC avoids some of these barriers and can achieve high uptake of testing, but doubts have been expressed about the sustainability of this approach to testing. Here, the researchers evaluate an alternative to home-based HTC—HIV self-testing (HIVST)—by undertaking a community-based prospective study of HIVST in Blantyre, Malawi. HIVST involves individuals performing and interpreting their own HIV test and has the potential to be widely implemented with minimal involvement of trained healthcare workers. What Did the Researchers Do and Find? Trained resident volunteer-counselors offered one oral HIVST kit (a kit that measures HIV in saliva) per year for a two-year period to 16,660 adult residents in 14 neighborhoods in urban Blantyre. All the participants received instructions on how to use the kits, pre- and post-counseling, and, for participants self-testing HIV positive, a referral card to attend an HIV care clinic. The residents also completed a questionnaire about their experience of HIVST. Three-quarters of the residents self-tested in the first and second year of the study. HIVST uptake was more rapid in the second year than in the first year and was high among men and adolescents, two hard-to-reach populations. Three-quarters of the residents who self-tested during the first year of the study shared their results with a volunteer-counselor. Of the 1,257 participants who discovered they were HIV positive during the first year of the study, more than half accessed HIV care. Importantly, 94.4% of the participants reported that they were happy with HIVST even though 2.9% reported being forced to take the test, usually by a main partner; no HIVST-related partner violence or suicides were reported by the study’s community surveillance system. Finally, HIVST and repeat HTC results agreed in 99.4% of participants selected as a quality assurance sample (one in 20 of the participants). What Do These Findings Mean? These findings show that, in urban neighborhoods in Malawi, coverage with community-based HIVST was high (particularly among adolescents and men) in two successive years and that HIVST was safe, accurate, and acceptable. Importantly, HIVST using a delivery model based on trained volunteers led to acceptable linkage into HIV care services, and the approach had a very low incidence of major social harms such as partner violence. Uncertainty about estimates of uptake and linkage to care and other aspects of the study design may limit the accuracy of these results. Nevertheless, these findings suggest that scaling up HIVST could complement existing strategies for providing early HIV diagnosis and periodic repeat testing and could thus have a sustained impact on the coverage of HIV testing and care in Africa and on the control of the HIV/AIDS epidemic. Additional Information This list of resources contains links that can be accessed when viewing the PDF on a device or via the online version of the article at http://dx.doi.org/10.1371/journal.pmed.1001873. Information is available from the US National Institute of Allergy and Infectious Diseases on all aspects of HIV infection and AIDS, including testing and diagnosis NAM/aidsmap provides basic information about HIV/AIDS, summaries of recent research findings on HIV care and treatment, and personal stories about living with HIV/AIDS Information is available from Avert, an international AIDS charity, on many aspects of HIV/AIDS, including HIV testing, HIV/AIDS treatment and care, and HIV prevention, and on HIV/AIDS in Malawi and in sub-Saharan Africa; Avert also provides personal stories about living with HIV/AIDS The World Health Organization provides information on all aspects of HIV/AIDS (in several languages), including its new consolidated guidelines on HIV testing The UNAIDS Fast-Track Strategy to End the AIDS Epidemic by 2030 provides up-to-date information about the AIDS epidemic and efforts to halt it; UNAIDS also provides detailed region-specific information and policy news. Introduction Sub-Saharan Africa is still disproportionately affected by the HIV epidemic, accounting for 71% (24.7 million) of people living with HIV globally; in 2013, 71% of the 2.1 million global new infections, and 73% of the 1.5 million HIV-related deaths, occurred in the region [1]. Despite major investments in HIV testing, treatment, and prevention programmes, only one-quarter of adult Africans have had a recent HIV test, and half of people living with HIV in sub-Saharan Africa do not know they are HIV positive [1–3]. Barriers to HIV testing and counselling (HTC) and initiation of antiretroviral therapy (ART) include overly busy health facilities, concerns about lack of confidentiality and privacy, and high out-of-pocket costs [4–6]. Community-based HTC approaches, including home-based and mobile services, can overcome some of these problems, achieving high population uptake of HTC [7–10]. Compared to facility-based approaches, community-based HTC provides earlier HIV diagnosis and increases uptake of couples testing [4,5]. Nevertheless, evaluation of community-based HTC and HIV services has raised concerns about cost and sustainability [11,12], especially for delivering services to more rural settings [12,13]. For example, despite community-based HTC being national policy in Malawi and Zimbabwe, only 2% of Malawians and 4% of Zimbabweans in 2010 were reached by mobile or door-to-door services [3]. HIV self-testing (HIVST), defined as an individual performing and interpreting his/her own HIV test [14], has the potential to be implemented at a wide scale with a minimal requirement for trained health-workers. As such, HIVST could improve population coverage of regular HTC, recognised as being a critical component of all strategies to further intensify HIV prevention and care in countries with generalised HIV epidemics. We have previously demonstrated very high uptake and accuracy of HIVST in a small feasibility study [7]. However, critical, unanswered questions that need to be addressed before considering large-scale interventions based on HIVST include the following: what levels of HIVST uptake and accuracy can be achieved with population-wide implementation, and do safety concerns, including the potential for coercive testing, suicide, and gender-based violence, preclude implementation [15–17]? We, therefore, investigated the uptake, accuracy, and outcomes of implementation of community-wide HIVST delivered by trained resident volunteer-counsellors in Blantyre, Malawi [18]. A delivery system based on service provision from the houses of volunteer-counsellors was designed. The aim was to evaluate uptake, accuracy, linkage into care, and health outcomes when highly convenient and flexible but supported access to HIVST kits was provided to a well-defined and closely monitored population. HIVST services were flexibly provided, with facilitated access to HIV care for those willing to share positive results. Participants could opt for support ranging from standard provider-conducted HTC to HIVST at home either in complete privacy or assisted by an attendant volunteer-counsellor. Methods Ethical Statement Ethical approval was obtained from the College of Medicine Ethics Review Committee, University of Malawi; London School of Hygiene & Tropical Medicine; and Liverpool School of Tropical Medicine. All participants opting for HIVST provided written (or witnessed thumbprint) informed consent. Study Design This study was a prospective study nested within a cluster-randomised trial (ISRCTN02004005) comparing health outcomes between 14 clusters randomised to HIVST and 14 clusters randomised to routine (facility-based) HTC [18]. The data reported here relate only to the 14 clusters where HIVST was provided. HIVST was provided for a 2-y period in any given cluster, starting between February and May 2012; active surveillance for harms continued for 4–6 mo after the 2-y HIVST period. Study Setting and Study Population The study took place in three high-density informal residential settlements in urban Blantyre, as described elsewhere [10,18]. In brief, neighbourhood clusters were defined on the basis of existing community health worker catchment areas and enumerated between April and June 2011. In clusters randomised to the intervention arm, community-based HIVST was available for all adults (≥16 y). Services were provided by two resident volunteer-counsellors in each cluster of ~1,200 adults; the volunteer-counsellors were identified using participatory methods [19] and were paid a monthly stipend similar to that of Malawi Ministry of Health community health workers. Volunteer-counsellors received Malawi Ministry of Health HTC training and study-specific HIVST and protocol training. Targets within each cluster were to reach >80% of adult residents each year through promoting HIVST door to door and leafleting. Participants could opt to test at home, with or without the volunteer-counsellor present to provide help as needed. HIV Self-Testing Kit Provision Participants (individuals or couples) received pre-test counselling, received instructions on performing HIVST, and were asked to demonstrate understanding using a cotton bud and vial of water in place of the kit itself. An anonymous self-completed questionnaire (SCQ) was provided with an opaque envelope for return of the used kit and SCQ, either to the volunteer-counsellor or into a locked “ballot” box kept at the volunteer-counsellor’s house (S1 Questionnaire). The test kit used was OraQuick ADVANCE Rapid HIV-1/2 Antibody Test (OraSure Technologies). User instructions were modified and included pictures. The ten-item SCQ included questions about the self-read HIVST result, satisfaction indicators, and the results of the individual’s most recent previous HIV test, if applicable. The question “If you were forced to test, who forced you?” was used to define coercion. Residents were asked to limit HIVST to one test in each 12-mo time period. Post-test counselling was recommended, but not required. All participants received a “self-referral card” allowing them to directly access one of two study clinics, but were encouraged to share results with their resident volunteer-counsellor for standard results-based post-test counselling and referral. A modified counselling protocol (including written information on all local HIV care options) was used for participants unwilling to share their results. Within seven of the 14 study clusters, a second cluster-randomised controlled trial was conducted that investigated the effect of optional home-based initiation of HIV care (ART eligibility assessment and 2 wk of treatment including ART if indicated) on uptake of ART [10]. This intervention was extended to all 14 HIVST clusters from January 2013 onwards. At health facilities, a study nurse provided confirmatory HIV testing (Determine HIV-1/2, Alere; and Uni-Gold Recombigen HIV, Trinity Biotech), CD4 count measurement (Cyflow SL-3 platform, Partec), tuberculosis screening (with isoniazid preventive therapy for those eligible [20]), WHO clinical staging, and cotrimoxazole. Participants who met national ART eligibility criteria (CD4 count < 350 cells/μl or WHO stage 3 or 4 or breastfeeding or pregnant) were registered for ART. Ascertainment of Outcomes Volunteer-counsellors recorded each individual/couple with nature of support provided for the test, age, and sex of the individual(s), and whether they had tested before. Estimates of linkage into care were based on the number of participants who disclosed positive results to counsellors during the first 12 mo compared to the number of participants accessing study clinic confirmatory testing and HIV care over the same time period. Confirmation of participation in the study was based on presentation of the self-referral card. Recording Social Harms In each cluster, four community members (key informants) provided weekly reports of all deaths and any known episodes of intimate partner violence. Study nurses conducted verbal autopsies for all reported deaths, including temporal relatedness to HIVST. Quality Assurance A systematic sample of HIVST participants was selected for home visit by study nurses, aiming for minimum 5% coverage. Nurses selected from participants tested in the previous week using counsellors’ HIVST logs that recorded one participant or couple per row, with 20 rows per page. A random number between 1 and 20 was generated on a weekly basis and provided to nurses on the day of use. Nurses selected the corresponding row number (e.g., each row 11 participant if the number 11 had been supplied that week). If the selected number exceeded the number of participants on any given page, then the nurses continued counting out from row 1 of the same page until that week’s number was reached. Checks during the home visit included age, confirmation of residency, whether or not HIVST kits had been used, and self-read result, with offer of confirmatory testing (finger-prick blood parallel testing with Determine HIV-1/2 and Uni-Gold Recombigen HIV). Statistical Analysis and Sample Size Stata version 13.0 (StataCorp) and R version 2.15.3 (R Foundation for Statistical Computing) were used for analyses. The sample size for the parent cluster-randomised trial was determined by the primary outcome (cluster-level tuberculosis case notification rates) and not by HIVST uptake or linkage. Of note, however, primary outcome assumptions were that population uptake of HIVST would be ≥70% per year [7,8], with ≥80% linkage into HIV care [21] and HIVST accuracy of ≥90% [7]. The proportion of residents accepting HIVST was estimated both overall and within sex, age, and neighbourhood strata, using population denominators from the study census (i.e., proportions were calculated using a fixed denominator that was determined before the start of the study, rather than as cumulative incidence, which would have required individual cohort follow-up for all residents) conducted in the year preceding the rollout of the intervention. Since crude uptake in some sex-age-neighbourhood subgroups exceeded the population denominators from the study census, the number of residents accepting HIVST within any single sex-age-neighbourhood subgroup was capped at the census denominator for that subgroup to provide an adjusted uptake. The first estimate of linkage into care was calculated with the number of participants who presented at a study clinic with a volunteer-counsellor-provided self-referral card as the numerator and the number of participants who disclosed a positive HIV result to the volunteer-counsellor as the denominator. The second estimate was calculated after adjusting for a proportion assumed to be already aware of their positive HIV status and in care. Participant characteristics in months 1–12 and months 13–24 were compared using design-based F-tests calculated by applying the second-order Rao and Scott correction [22,23] to the usual Pearson chi-squared test statistic for two-way tables to allow for the clustered sampling design. The accuracy of self-reported HIVST results in quality assurance (QA) participants was assessed using finger-prick rapid diagnostic test results to calculate sensitivity, specificity, and exact binomial 95% confidence intervals. Univariate and multivariate random effects logistic regression models accounting for clustering at the neighbourhood level were fitted in order to obtain odds ratios (ORs) and 95% CIs for associations between prespecified exposures of interest (age, sex, previous testing, testing alone/with partner, self-read HIVST result) and reported coercion. A substantial proportion of SCQ participants had missing data for at least one of the exposures of interest. Comparison of characteristics of participants with and without complete data showed no significant differences, and, therefore, findings from complete case analysis are presented [24]. Sensitivity analysis was undertaken using multiple imputation methods to handle missing data. Ethical Statement Ethical approval was obtained from the College of Medicine Ethics Review Committee, University of Malawi; London School of Hygiene & Tropical Medicine; and Liverpool School of Tropical Medicine. All participants opting for HIVST provided written (or witnessed thumbprint) informed consent. Study Design This study was a prospective study nested within a cluster-randomised trial (ISRCTN02004005) comparing health outcomes between 14 clusters randomised to HIVST and 14 clusters randomised to routine (facility-based) HTC [18]. The data reported here relate only to the 14 clusters where HIVST was provided. HIVST was provided for a 2-y period in any given cluster, starting between February and May 2012; active surveillance for harms continued for 4–6 mo after the 2-y HIVST period. Study Setting and Study Population The study took place in three high-density informal residential settlements in urban Blantyre, as described elsewhere [10,18]. In brief, neighbourhood clusters were defined on the basis of existing community health worker catchment areas and enumerated between April and June 2011. In clusters randomised to the intervention arm, community-based HIVST was available for all adults (≥16 y). Services were provided by two resident volunteer-counsellors in each cluster of ~1,200 adults; the volunteer-counsellors were identified using participatory methods [19] and were paid a monthly stipend similar to that of Malawi Ministry of Health community health workers. Volunteer-counsellors received Malawi Ministry of Health HTC training and study-specific HIVST and protocol training. Targets within each cluster were to reach >80% of adult residents each year through promoting HIVST door to door and leafleting. Participants could opt to test at home, with or without the volunteer-counsellor present to provide help as needed. HIV Self-Testing Kit Provision Participants (individuals or couples) received pre-test counselling, received instructions on performing HIVST, and were asked to demonstrate understanding using a cotton bud and vial of water in place of the kit itself. An anonymous self-completed questionnaire (SCQ) was provided with an opaque envelope for return of the used kit and SCQ, either to the volunteer-counsellor or into a locked “ballot” box kept at the volunteer-counsellor’s house (S1 Questionnaire). The test kit used was OraQuick ADVANCE Rapid HIV-1/2 Antibody Test (OraSure Technologies). User instructions were modified and included pictures. The ten-item SCQ included questions about the self-read HIVST result, satisfaction indicators, and the results of the individual’s most recent previous HIV test, if applicable. The question “If you were forced to test, who forced you?” was used to define coercion. Residents were asked to limit HIVST to one test in each 12-mo time period. Post-test counselling was recommended, but not required. All participants received a “self-referral card” allowing them to directly access one of two study clinics, but were encouraged to share results with their resident volunteer-counsellor for standard results-based post-test counselling and referral. A modified counselling protocol (including written information on all local HIV care options) was used for participants unwilling to share their results. Within seven of the 14 study clusters, a second cluster-randomised controlled trial was conducted that investigated the effect of optional home-based initiation of HIV care (ART eligibility assessment and 2 wk of treatment including ART if indicated) on uptake of ART [10]. This intervention was extended to all 14 HIVST clusters from January 2013 onwards. At health facilities, a study nurse provided confirmatory HIV testing (Determine HIV-1/2, Alere; and Uni-Gold Recombigen HIV, Trinity Biotech), CD4 count measurement (Cyflow SL-3 platform, Partec), tuberculosis screening (with isoniazid preventive therapy for those eligible [20]), WHO clinical staging, and cotrimoxazole. Participants who met national ART eligibility criteria (CD4 count < 350 cells/μl or WHO stage 3 or 4 or breastfeeding or pregnant) were registered for ART. Ascertainment of Outcomes Volunteer-counsellors recorded each individual/couple with nature of support provided for the test, age, and sex of the individual(s), and whether they had tested before. Estimates of linkage into care were based on the number of participants who disclosed positive results to counsellors during the first 12 mo compared to the number of participants accessing study clinic confirmatory testing and HIV care over the same time period. Confirmation of participation in the study was based on presentation of the self-referral card. Recording Social Harms In each cluster, four community members (key informants) provided weekly reports of all deaths and any known episodes of intimate partner violence. Study nurses conducted verbal autopsies for all reported deaths, including temporal relatedness to HIVST. Quality Assurance A systematic sample of HIVST participants was selected for home visit by study nurses, aiming for minimum 5% coverage. Nurses selected from participants tested in the previous week using counsellors’ HIVST logs that recorded one participant or couple per row, with 20 rows per page. A random number between 1 and 20 was generated on a weekly basis and provided to nurses on the day of use. Nurses selected the corresponding row number (e.g., each row 11 participant if the number 11 had been supplied that week). If the selected number exceeded the number of participants on any given page, then the nurses continued counting out from row 1 of the same page until that week’s number was reached. Checks during the home visit included age, confirmation of residency, whether or not HIVST kits had been used, and self-read result, with offer of confirmatory testing (finger-prick blood parallel testing with Determine HIV-1/2 and Uni-Gold Recombigen HIV). Statistical Analysis and Sample Size Stata version 13.0 (StataCorp) and R version 2.15.3 (R Foundation for Statistical Computing) were used for analyses. The sample size for the parent cluster-randomised trial was determined by the primary outcome (cluster-level tuberculosis case notification rates) and not by HIVST uptake or linkage. Of note, however, primary outcome assumptions were that population uptake of HIVST would be ≥70% per year [7,8], with ≥80% linkage into HIV care [21] and HIVST accuracy of ≥90% [7]. The proportion of residents accepting HIVST was estimated both overall and within sex, age, and neighbourhood strata, using population denominators from the study census (i.e., proportions were calculated using a fixed denominator that was determined before the start of the study, rather than as cumulative incidence, which would have required individual cohort follow-up for all residents) conducted in the year preceding the rollout of the intervention. Since crude uptake in some sex-age-neighbourhood subgroups exceeded the population denominators from the study census, the number of residents accepting HIVST within any single sex-age-neighbourhood subgroup was capped at the census denominator for that subgroup to provide an adjusted uptake. The first estimate of linkage into care was calculated with the number of participants who presented at a study clinic with a volunteer-counsellor-provided self-referral card as the numerator and the number of participants who disclosed a positive HIV result to the volunteer-counsellor as the denominator. The second estimate was calculated after adjusting for a proportion assumed to be already aware of their positive HIV status and in care. Participant characteristics in months 1–12 and months 13–24 were compared using design-based F-tests calculated by applying the second-order Rao and Scott correction [22,23] to the usual Pearson chi-squared test statistic for two-way tables to allow for the clustered sampling design. The accuracy of self-reported HIVST results in quality assurance (QA) participants was assessed using finger-prick rapid diagnostic test results to calculate sensitivity, specificity, and exact binomial 95% confidence intervals. Univariate and multivariate random effects logistic regression models accounting for clustering at the neighbourhood level were fitted in order to obtain odds ratios (ORs) and 95% CIs for associations between prespecified exposures of interest (age, sex, previous testing, testing alone/with partner, self-read HIVST result) and reported coercion. A substantial proportion of SCQ participants had missing data for at least one of the exposures of interest. Comparison of characteristics of participants with and without complete data showed no significant differences, and, therefore, findings from complete case analysis are presented [24]. Sensitivity analysis was undertaken using multiple imputation methods to handle missing data. Results Uptake of HIVST In 2011, 16,660 adults (16 y or older) were enumerated in the 14 HIVST clusters. During months 1–12 and months 13–24, a total of 14,004 (84.1%) and 13,785 (82.7%) participants accessed the HIVST service, respectively (Figs 1 and 2). Compared to months 1–12, the second year saw higher proportions of men (46.1% versus 43.8%; p = 0.057), adolescents (24.7% versus 22.2%; p < 0.001), participants with a sexual partner (59.3% versus 37.5%; p < 0.001), and participants who had tested for HIV ever (82.2% versus 64.9%, and for testing within the last 12 mo, 61.2% versus 27.3%; p < 0.001 for both) (Table 1). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. Flow of study participants in months 1–12 of HIV self-testing. https://doi.org/10.1371/journal.pmed.1001873.g001 Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. Flow of study participants in months 13–24 of HIV self-testing. https://doi.org/10.1371/journal.pmed.1001873.g002 Download: PPT PowerPoint slide PNG larger image TIFF original image Table 1. Characteristics of HIV self-testing participants in the first and second years of HIV self-testing availability. https://doi.org/10.1371/journal.pmed.1001873.t001 The estimated uptake of HIVST, based on study census denominators, was 84.1% and 82.7% in months 1–12 and months 13–24, respectively. Crude uptake in some age-sex-neighbourhood subgroups (notably among adolescent women [aged 16–19 y]) exceeded population denominators from the census conducted in the year preceding the study (Table 2). Capping uptake in any single age-sex-neighbourhood subgroup at 100% led to revised uptake estimates of 76.5% and 74.4% in months 1–12 and months 13–24, respectively. With both approaches, there was significantly higher uptake each year amongst women than men, and for progressively younger age groups (p < 0.001 for both). Download: PPT PowerPoint slide PNG larger image TIFF original image Table 2. Age-sex distribution of study population and study participants with and without adjustment by study census maximum denominators in age-sex-neighbourhood subgroups. https://doi.org/10.1371/journal.pmed.1001873.t002 The time course of HIVST uptake within each annual period for which HIVST was restricted to a single test per person (Methods and QA results) is shown by time point, sex, and age group in Fig 3. In comparison to months 1–12, uptake during the second year of availability was more rapid, with a higher proportion accessing services soon after they became available (Fig 3), notably so for adolescents (aged 16–19 y). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. Cumulative uptake of HIV self-testing by sex, age group, and time point. (A) Cumulative uptake of HIVST during the first 12 mo of availability among all HIVST cluster residents by age and time point among men and women. HIVST uptake increased with time, rising to close to 100% by 12 mo in adolescents (age group 16–19 y); uptake for men was lower than for women at every time point. (B) Cumulative uptake of HIVST during months 13–24 of HIVST availability among all cluster residents by age and time point. Uptake defined as an individual having collected an HIVST kit from a community counsellor. Since crude uptake of HIVST exceeded 100% in some age-sex-neighbourhood subgroups, likely explained by migration, revised estimates were calculated where uptake in any single age-sex-neighbourhood subgroup was censored at 100%; study census data were used for denominators. https://doi.org/10.1371/journal.pmed.1001873.g003 HIV Prevalence in HIVST Participants and Linkage into Care In the first year of HIVST, HIV prevalence in participants sharing results with volunteer-counsellors was 11.8% (95% CI 11.2%–12.5%), similar to the estimate from the rereading of returned kits (10.1%, 95% CI 9.6%–10.7%) (Fig 1). These estimates, however, were substantially higher than the respective figures from months 13–24, which were 6.8% (95% CI 6.3%–7.2%) and 7.3% (95% CI 6.8%–7.8%). HIV prevalence among self-testing participants (shown separately for men and women in Fig 4) was highest in the age group 40–49 y, with a pooled prevalence of 22.5% (95% CI 19.4%–25.8%) in months 1–12; the pooled rate in participants aged 16–19 y (2.5%, 95% CI 1.9%–3.2%) was much lower. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 4. HIV prevalence in self-testing participants who returned used test kits by sex and age group and time of HIV self-testing availability. This figure shows HIV prevalence in HIVST participants for men (A) and women (B), stratified by time of HIVST availability. Bars show HIV prevalence (percent); error bars show 95% confidence intervals. Estimates are based on denominators determined through enumeration. Numerators were based on a reread of used and returned HIVST kits by a laboratory technician within 2 wk of use. Individuals were asked to test only once within each 12-mo time period, and retesting in people already aware of their positive HIV status was discouraged. https://doi.org/10.1371/journal.pmed.1001873.g004 In total, 75.8% (95% CI 75.1%–76.5%; 10,614/14,004) of participants who underwent HIVST in months 1–12 reported their result to a volunteer-counsellor, with 1,257 (11.8%, 95% CI 11.2%–12.5%) reporting a positive result. During this same time period, 524 participants presented for HIV care, with all presenting cards identifying them as having been directly referred in by a volunteer-counsellor (Fig 5). Thus, our first estimate of linkage is 41.7% (524 of 1,257 self-testing positive). However, in a subset of 3,016 participants in months 1–12, 2,380 (78.9%; 95% CI 77.4%–80.4%) responded to a question about ART. Of these, 219 (9.2%, 95% CI 8.1%–10.4%) were HIV positive, and 57 (26.0%, 95% CI 20.3%–32.4%) of these individuals stated that they were already on ART, consequently increasing our estimate of linkage to 56.3% (524/930). The median CD4 count from 415 participants (72.9% of those attending care) was 250 cells/μl (interquartile range [IQR] 159–426), with 66.3% (275/415) of CD4 counts being below 350 cells/μl. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 5. Linkage into HIV care after HIV self-testing (months 1–12). https://doi.org/10.1371/journal.pmed.1001873.g005 Accuracy A total of 2,361 (8.5%) of 27,789 HIVST participants were included in QA tracing (shown for separate years in Figs 1 and 2). Only 54 (2.3%) were found not to be cluster residents, while 1,649 (69.8%) agreed to confirmatory HIV testing. Results were positive in 141 (8.6%, 95% CI 7.2%–10.0%). Compared to stated HIVST results, there were 9/1,508 (0.6%) false negatives (including four participants already on ART) and 1/133 false positives, giving agreement of 1,639/1,649 (99.4%, 95% CI 98.9%–99.7%), sensitivity of 93.6% (95% CI 88.2%–97.0%), and specificity of 99.9% (95% CI 99.6%–100%) (Table 3). Download: PPT PowerPoint slide PNG larger image TIFF original image Table 3. Summary of quality assurance process and accuracy results. https://doi.org/10.1371/journal.pmed.1001873.t003 Acceptability of Self-Testing and Social Harms, Including Reported Coercive Testing During months 1–12, 81.1% (95% CI 80.5%–81.8%; 11,359/14,004) participants returned a SCQ to the counsellor, with 7,014 (61.7%) completing all key fields including self-read HIVST result, coercion, and acceptability indicators (S1 Questionnaire). There was acceptable internal consistency (Cronbach’s alpha = 0.64) for the four variables relating to acceptability: overall satisfaction with HIVST, whether or not they would recommend HIVST to friends and family, how hard it was to self-test, and whether or not they trusted the results of an oral test [25]. Acceptability indicators were high in all age group and sex strata, with 94.6% (1,446/1,635) reporting that they were “highly satisfied” with the HIVST process and 97.1% (6,683/6,883) reporting they would “definitely recommend HIVST to their friends and family”. These indicators did not vary significantly by self-reported HIV status, with those testing positive having OR 0.60 (95% CI 0.34–1.05) and OR 0.92 (95% CI 0.56–1.50) relative to HIV-negative participants for being “very satisfied” with the HIVST process and for “definitely” recommending HIVST to friends and family, respectively. In total, 288/10,017 participants (2.9%, 95% CI 2.6%–3.2%) reported having been coerced into participating in HIVST. Notably, however, satisfaction indicators in the group reporting coercion were high, with 94.4% (252/267) stating that they would recommend HIVST to friends and family, and 92.2% (130/141) reporting that they were highly satisfied with HIVST. In the univariate analysis, men and participants who self-tested with their partner were significantly more likely to report having been coerced into HIVST (Table 4). In multivariate analysis, male sex (adjusted OR [aOR] 1.83, 95% CI 1.38–2.43) and having tested with a partner (aOR 3.86, 95% CI 2.82–5.29) remained significantly associated with reported coercion. There was no significant difference in reporting of coercion by reported HIVST result to volunteer-counsellors (aOR 1.00, 95% CI 0.59–1.71). The findings were comparable when multiple imputation methods were used to handle missing data (S1 Table). Download: PPT PowerPoint slide PNG larger image TIFF original image Table 4. Factors associated with reported coercion during months 1–12 of HIV self-testing (n = 7,014). https://doi.org/10.1371/journal.pmed.1001873.t004 A total of 132 adult deaths were reported through the community liaison system during the first 12 mo of follow-up, including one suicide in an individual who had not self-tested and four murders, none of which had any known or close temporal relationship to self-testing. No intimate partner violence episodes were reported through the community liaison system. Uptake of HIVST In 2011, 16,660 adults (16 y or older) were enumerated in the 14 HIVST clusters. During months 1–12 and months 13–24, a total of 14,004 (84.1%) and 13,785 (82.7%) participants accessed the HIVST service, respectively (Figs 1 and 2). Compared to months 1–12, the second year saw higher proportions of men (46.1% versus 43.8%; p = 0.057), adolescents (24.7% versus 22.2%; p < 0.001), participants with a sexual partner (59.3% versus 37.5%; p < 0.001), and participants who had tested for HIV ever (82.2% versus 64.9%, and for testing within the last 12 mo, 61.2% versus 27.3%; p < 0.001 for both) (Table 1). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. Flow of study participants in months 1–12 of HIV self-testing. https://doi.org/10.1371/journal.pmed.1001873.g001 Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. Flow of study participants in months 13–24 of HIV self-testing. https://doi.org/10.1371/journal.pmed.1001873.g002 Download: PPT PowerPoint slide PNG larger image TIFF original image Table 1. Characteristics of HIV self-testing participants in the first and second years of HIV self-testing availability. https://doi.org/10.1371/journal.pmed.1001873.t001 The estimated uptake of HIVST, based on study census denominators, was 84.1% and 82.7% in months 1–12 and months 13–24, respectively. Crude uptake in some age-sex-neighbourhood subgroups (notably among adolescent women [aged 16–19 y]) exceeded population denominators from the census conducted in the year preceding the study (Table 2). Capping uptake in any single age-sex-neighbourhood subgroup at 100% led to revised uptake estimates of 76.5% and 74.4% in months 1–12 and months 13–24, respectively. With both approaches, there was significantly higher uptake each year amongst women than men, and for progressively younger age groups (p < 0.001 for both). Download: PPT PowerPoint slide PNG larger image TIFF original image Table 2. Age-sex distribution of study population and study participants with and without adjustment by study census maximum denominators in age-sex-neighbourhood subgroups. https://doi.org/10.1371/journal.pmed.1001873.t002 The time course of HIVST uptake within each annual period for which HIVST was restricted to a single test per person (Methods and QA results) is shown by time point, sex, and age group in Fig 3. In comparison to months 1–12, uptake during the second year of availability was more rapid, with a higher proportion accessing services soon after they became available (Fig 3), notably so for adolescents (aged 16–19 y). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. Cumulative uptake of HIV self-testing by sex, age group, and time point. (A) Cumulative uptake of HIVST during the first 12 mo of availability among all HIVST cluster residents by age and time point among men and women. HIVST uptake increased with time, rising to close to 100% by 12 mo in adolescents (age group 16–19 y); uptake for men was lower than for women at every time point. (B) Cumulative uptake of HIVST during months 13–24 of HIVST availability among all cluster residents by age and time point. Uptake defined as an individual having collected an HIVST kit from a community counsellor. Since crude uptake of HIVST exceeded 100% in some age-sex-neighbourhood subgroups, likely explained by migration, revised estimates were calculated where uptake in any single age-sex-neighbourhood subgroup was censored at 100%; study census data were used for denominators. https://doi.org/10.1371/journal.pmed.1001873.g003 HIV Prevalence in HIVST Participants and Linkage into Care In the first year of HIVST, HIV prevalence in participants sharing results with volunteer-counsellors was 11.8% (95% CI 11.2%–12.5%), similar to the estimate from the rereading of returned kits (10.1%, 95% CI 9.6%–10.7%) (Fig 1). These estimates, however, were substantially higher than the respective figures from months 13–24, which were 6.8% (95% CI 6.3%–7.2%) and 7.3% (95% CI 6.8%–7.8%). HIV prevalence among self-testing participants (shown separately for men and women in Fig 4) was highest in the age group 40–49 y, with a pooled prevalence of 22.5% (95% CI 19.4%–25.8%) in months 1–12; the pooled rate in participants aged 16–19 y (2.5%, 95% CI 1.9%–3.2%) was much lower. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 4. HIV prevalence in self-testing participants who returned used test kits by sex and age group and time of HIV self-testing availability. This figure shows HIV prevalence in HIVST participants for men (A) and women (B), stratified by time of HIVST availability. Bars show HIV prevalence (percent); error bars show 95% confidence intervals. Estimates are based on denominators determined through enumeration. Numerators were based on a reread of used and returned HIVST kits by a laboratory technician within 2 wk of use. Individuals were asked to test only once within each 12-mo time period, and retesting in people already aware of their positive HIV status was discouraged. https://doi.org/10.1371/journal.pmed.1001873.g004 In total, 75.8% (95% CI 75.1%–76.5%; 10,614/14,004) of participants who underwent HIVST in months 1–12 reported their result to a volunteer-counsellor, with 1,257 (11.8%, 95% CI 11.2%–12.5%) reporting a positive result. During this same time period, 524 participants presented for HIV care, with all presenting cards identifying them as having been directly referred in by a volunteer-counsellor (Fig 5). Thus, our first estimate of linkage is 41.7% (524 of 1,257 self-testing positive). However, in a subset of 3,016 participants in months 1–12, 2,380 (78.9%; 95% CI 77.4%–80.4%) responded to a question about ART. Of these, 219 (9.2%, 95% CI 8.1%–10.4%) were HIV positive, and 57 (26.0%, 95% CI 20.3%–32.4%) of these individuals stated that they were already on ART, consequently increasing our estimate of linkage to 56.3% (524/930). The median CD4 count from 415 participants (72.9% of those attending care) was 250 cells/μl (interquartile range [IQR] 159–426), with 66.3% (275/415) of CD4 counts being below 350 cells/μl. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 5. Linkage into HIV care after HIV self-testing (months 1–12). https://doi.org/10.1371/journal.pmed.1001873.g005 Accuracy A total of 2,361 (8.5%) of 27,789 HIVST participants were included in QA tracing (shown for separate years in Figs 1 and 2). Only 54 (2.3%) were found not to be cluster residents, while 1,649 (69.8%) agreed to confirmatory HIV testing. Results were positive in 141 (8.6%, 95% CI 7.2%–10.0%). Compared to stated HIVST results, there were 9/1,508 (0.6%) false negatives (including four participants already on ART) and 1/133 false positives, giving agreement of 1,639/1,649 (99.4%, 95% CI 98.9%–99.7%), sensitivity of 93.6% (95% CI 88.2%–97.0%), and specificity of 99.9% (95% CI 99.6%–100%) (Table 3). Download: PPT PowerPoint slide PNG larger image TIFF original image Table 3. Summary of quality assurance process and accuracy results. https://doi.org/10.1371/journal.pmed.1001873.t003 Acceptability of Self-Testing and Social Harms, Including Reported Coercive Testing During months 1–12, 81.1% (95% CI 80.5%–81.8%; 11,359/14,004) participants returned a SCQ to the counsellor, with 7,014 (61.7%) completing all key fields including self-read HIVST result, coercion, and acceptability indicators (S1 Questionnaire). There was acceptable internal consistency (Cronbach’s alpha = 0.64) for the four variables relating to acceptability: overall satisfaction with HIVST, whether or not they would recommend HIVST to friends and family, how hard it was to self-test, and whether or not they trusted the results of an oral test [25]. Acceptability indicators were high in all age group and sex strata, with 94.6% (1,446/1,635) reporting that they were “highly satisfied” with the HIVST process and 97.1% (6,683/6,883) reporting they would “definitely recommend HIVST to their friends and family”. These indicators did not vary significantly by self-reported HIV status, with those testing positive having OR 0.60 (95% CI 0.34–1.05) and OR 0.92 (95% CI 0.56–1.50) relative to HIV-negative participants for being “very satisfied” with the HIVST process and for “definitely” recommending HIVST to friends and family, respectively. In total, 288/10,017 participants (2.9%, 95% CI 2.6%–3.2%) reported having been coerced into participating in HIVST. Notably, however, satisfaction indicators in the group reporting coercion were high, with 94.4% (252/267) stating that they would recommend HIVST to friends and family, and 92.2% (130/141) reporting that they were highly satisfied with HIVST. In the univariate analysis, men and participants who self-tested with their partner were significantly more likely to report having been coerced into HIVST (Table 4). In multivariate analysis, male sex (adjusted OR [aOR] 1.83, 95% CI 1.38–2.43) and having tested with a partner (aOR 3.86, 95% CI 2.82–5.29) remained significantly associated with reported coercion. There was no significant difference in reporting of coercion by reported HIVST result to volunteer-counsellors (aOR 1.00, 95% CI 0.59–1.71). The findings were comparable when multiple imputation methods were used to handle missing data (S1 Table). Download: PPT PowerPoint slide PNG larger image TIFF original image Table 4. Factors associated with reported coercion during months 1–12 of HIV self-testing (n = 7,014). https://doi.org/10.1371/journal.pmed.1001873.t004 A total of 132 adult deaths were reported through the community liaison system during the first 12 mo of follow-up, including one suicide in an individual who had not self-tested and four murders, none of which had any known or close temporal relationship to self-testing. No intimate partner violence episodes were reported through the community liaison system. Discussion The main finding of this study was the high population uptake of HIVST and retesting during 2 y of highly decentralised service provision in an urban community in Malawi. HIVST was safe and accurate, with uptake highest among adolescents, and with acceptable linkage into HIV care services using a delivery model based on trained volunteers. No suicides or other serious unintended consequences related to HIVST were detected by an active community surveillance system, including systematic death reporting and verbal autopsies. Feeling coerced into self-testing (usually by a main partner) was common (2.9% respondents), but was nonetheless associated with a high satisfaction rating for HIVST for all but a small minority of respondents. This model of HIVST is potentially scalable to other low-income settings where annual repeat HIV testing is recommended. HIV testing needs in Africa have changed dramatically in the last decade due to the massive scale-up of ART services and an increasing focus on early diagnosis and treatment of HIV for prevention [26,27], as well as other biomedical HIV prevention strategies [28,29]. Population surveys and qualitative studies report high readiness to test, but there exist substantial barriers to accessing free clinic-based HIV testing services [30–33]. The high acceptability and ease of distribution of oral test kits makes HIVST of special interest in high-HIV settings, where the aim is to achieve affordable universal coverage and regular repeat testing [34]. Here we report considerable complementarity of this model of HIVST with existing strategies. Although our urban population was already served by free facility-based services, 35% of participants in the first 12 mo had never previously tested, and uptake was high in two important hard-to-reach groups: men and adolescents. Our estimates of adolescent population uptake (~100% for women aged 16–19 y and ~90% for men aged 16–19 y) are in stark contrast with reported adolescent HTC uptake in African DHS surveys [3]. Ideally, HIVST services would capitalise on high acceptability among key populations, facilitating linkage into HIV prevention programmes, such as pre-exposure prophylaxis and voluntary medical male circumcision, as well as ensuring prompt linkage into HIV care [14]. The per-episode costs of providing HIVST compared to the costs of facility-based testing will be reported fully elsewhere. Our data from the second year of HIVST availability (participants were asked to test only once in each year) show high readiness to retest, as well as reduced numbers of first-time testers and new positive HIV diagnoses, which is consistent with the high coverage reported from the first year. Importantly, population uptake in the second year was faster, suggesting that under programmatic conditions, experienced volunteer-counsellors could cover larger populations as soon as communities have been familiarised with HIVST concepts. Optimum systems for linking clients into HIV care/prevention programmes are not well established in Africa [35–38] but are critical to the public health impact and cost-effectiveness of HTC [39]. Here we estimate a timely linkage into confirmatory testing and HIV care following HIVST of 56%, which compares favourably with many other approaches [40] and is well within the expected range for African HTC services [35,36]. This linkage estimate, however, reflects that, in addition to HIVST, participants were asked to attend post-test counselling and were advised to share their HIVST results. Facilitated HIV care assessment and initiation was provided following a successful trial in the first 6 mo of this study [10]. Despite reluctance to be tested by a volunteer-counsellor who is a neighbour, willingness to take kits and to share results was high. Although at first seemingly paradoxical, other studies have also reported that learning one’s HIV status demands a moment of complete privacy, but that being able to turn to someone familiar can then make the next steps of accessing HIV care less daunting [41]. Some of the benefits of community-based HTC are reaching HIV-positive individuals earlier [42], improving survival [43], and reducing costs [44] and onward transmission. A recent meta-analysis has found that when CD4 measurement was offered in tandem with home-based HIV testing, approximately 60% of those who tested HIV positive had CD4 counts greater than 350 cells/μl [9]. Here we report a CD4 count profile below this ideal (median 250 cells/μl, IQR 159–426) for HIVST participants who subsequently attended care, but still considerably higher than that of HIV care attendees diagnosed from our study clusters following standard non-study HTC (median 154 cells/μl, IQR 116–249) [10]. Concerns about the potential impact of user error on diagnostic accuracy from HIVST [45,46] have been widely discussed [14]. Here we report an HIVST accuracy (93.6% sensitivity, 99.9% specificity) very similar to that of unobserved HIVST using the OraQuick ADVANCE Rapid HIV-1/2 Antibody Test in American participants [47]. We have previously reported 97.9% sensitivity and 100% specificity for a small observed/controlled-setting study in Blantyre [7]. In the HIVST model evaluated, users were given a short simple demonstration by trained lay volunteers, and this may have been a key factor in maintaining high accuracy in this relatively low literacy setting. Both accuracy and uptake of services post-testing will need revaluation if different test kits or less supportive models are considered, for example, over-the-counter or vending machine sales. Also of note, a much higher than anticipated proportion (26%) of our HIV-positive HIVST participants were on ART already, as were two of our four participants found to have false-negative results. ART is known to reduce sensitivity especially for oral fluid-based rapid diagnostic tests [48]. In Malawi, faith healing, whereby HIV is considered curable through prayer, is widely preached and may prompt ART patients to reconsider their status and need for ART if they get a negative test result via HIVST [49]. Based on our experience, we would recommend careful messaging about retesting while on ART in HIVST package inserts and education campaigns. Coercion was reported by 3% of our SCQ respondents and was the major social harm, with no suicides or intimate partner violence attributed to HIVST despite active surveillance. Comparable data suggest that feeling coerced affects other modalities of HTC, with an estimated 7% of HTC episodes in Africa occurring without consent [50]. Both pregnant women and their male partners commonly report feeling coerced into testing by health professionals [51]. Among our participants, men and those who tested with their partners were more likely to report coercion. HIVST programmes need to anticipate and guard against coercive and mandatory testing, and to ensure that information about rights is disseminated and that systems for reporting social harms are in place. Study limitations include uncertainty around our linkage and uptake estimates, and use of aggregate-level data reporting rather than individual cohort follow-up. Population turnover, typically high in urban slums, was not factored into our population denominators, and may in part explain why our crude uptake estimates for adolescent women were >100%. Importantly, our QA programme results ruled out a major contribution to our findings from HIVST offered to non-eligible individuals (non-residents and individuals taking multiple tests). Estimates of linkage into care always have a wide uncertainty (Fig 5), but as disclosure of positive HIVST results was voluntary, even our precise denominators are unknown. Furthermore, we under-appreciated the extent of retesting while already on ART, adding to the uncertainty around numbers of newly identified HIV-positive participants. However, these sources of imprecision are unlikely to have affected our overall messages. In summary, community-level HIVST service provision along with supportive post-test services resulted in high and rapid uptake of accurate HIVST, with very low incidence of major social harms, and acceptable linkage into HIV care. The continued high uptake in the second year suggests that scaling up HIVST could have a sustained impact on the coverage of HIV testing and care in Africa, especially for men and adolescents. Supporting Information S1 Checklist. STROBE checklist. https://doi.org/10.1371/journal.pmed.1001873.s001 (DOCX) S1 Questionnaire. Self-completed questionnaire. https://doi.org/10.1371/journal.pmed.1001873.s002 (PDF) S1 Table. Comparison between complete case analysis (n = 7,014) presented in Table 4 and analysis based on imputed data (n = 11,359). https://doi.org/10.1371/journal.pmed.1001873.s003 (DOCX) Acknowledgments We thank the community members who participated in the study and the patients and staff at Ndirande Health Centre, Chilomoni Health Centre, and Queen Elizabeth Central Hospital in Blantyre. The Blantyre District Health Office and the HIV Department of the Ministry of Health of Malawi provided invaluable technical support.

doi: 10.1371/journal.pmed.1001875pmid: 26372359

Background Although selective serotonin reuptake inhibitors (SSRIs) are widely prescribed, associations with violence are uncertain. Methods and Findings From Swedish national registers we extracted information on 856,493 individuals who were prescribed SSRIs, and subsequent violent crimes during 2006 through 2009. We used stratified Cox regression analyses to compare the rate of violent crime while individuals were prescribed these medications with the rate in the same individuals while not receiving medication. Adjustments were made for other psychotropic medications. Information on all medications was extracted from the Swedish Prescribed Drug Register, with complete national data on all dispensed medications. Information on violent crime convictions was extracted from the Swedish national crime register. Using within-individual models, there was an overall association between SSRIs and violent crime convictions (hazard ratio [HR] = 1.19, 95% CI 1.08–1.32, p < 0.001, absolute risk = 1.0%). With age stratification, there was a significant association between SSRIs and violent crime convictions for individuals aged 15 to 24 y (HR = 1.43, 95% CI 1.19–1.73, p < 0.001, absolute risk = 3.0%). However, there were no significant associations in those aged 25–34 y (HR = 1.20, 95% CI 0.95–1.52, p = 0.125, absolute risk = 1.6%), in those aged 35–44 y (HR = 1.06, 95% CI 0.83–1.35, p = 0.666, absolute risk = 1.2%), or in those aged 45 y or older (HR = 1.07, 95% CI 0.84–1.35, p = 0.594, absolute risk = 0.3%). Associations in those aged 15 to 24 y were also found for violent crime arrests with preliminary investigations (HR = 1.28, 95% CI 1.16–1.41, p < 0.001), non-violent crime convictions (HR = 1.22, 95% CI 1.10–1.34, p < 0.001), non-violent crime arrests (HR = 1.13, 95% CI 1.07–1.20, p < 0.001), non-fatal injuries from accidents (HR = 1.29, 95% CI 1.22–1.36, p < 0.001), and emergency inpatient or outpatient treatment for alcohol intoxication or misuse (HR = 1.98, 95% CI 1.76–2.21, p < 0.001). With age and sex stratification, there was a significant association between SSRIs and violent crime convictions for males aged 15 to 24 y (HR = 1.40, 95% CI 1.13–1.73, p = 0.002) and females aged 15 to 24 y (HR = 1.75, 95% CI 1.08–2.84, p = 0.023). However, there were no significant associations in those aged 25 y or older. One important limitation is that we were unable to fully account for time-varying factors. Conclusions The association between SSRIs and violent crime convictions and violent crime arrests varied by age group. The increased risk we found in young people needs validation in other studies. Background Antidepressants—drugs that treat depression (unbearable feelings of sadness and despair caused by changes in brain chemistry)—are widely prescribed in many countries. In the US, for example, about one in ten people over 12 years old take antidepressants. The first antidepressants—monoamine oxidase inhibitors and tricyclic antidepressants—were developed in the 1950s. Experts think that both these classes of drugs treat depression by increasing serotonin levels in the brain. Serotonin, which is thought to improve mood, emotion, and sleep, is a neurotransmitter, a chemical that carries messages between nerve cells. However, monoamine oxidase inhibitors and tricyclic antidepressants had many adverse side effects unrelated to their effects on serotonin levels. So, in the late 1980s, a new class of antidepressant drugs was launched known as selective serotonin reuptake inhibitors (SSRIs). After serotonin delivers a message between nerve cells, it is usually reabsorbed by the nerve cells. Fluoxetine (Prozac), paroxetine (Seroxat), and other SSRIs block this “reuptake,” thereby increasing serotonin levels in the brain. Why Was This Study Done? SSRIs (which are also used to treat several other mental health conditions) have fewer side effects than the older antidepressants, although they can cause headache, nausea, sleep problems, restlessness, and sexual problems. However, SSRIs are not recommended for use in people under the age of 18 years because there is some evidence that SSRIs increase the risk of self-harm and suicidal thoughts in this age group. Moreover, there is limited and inconclusive evidence linking SSRI use with violent behavior. Because SSRIs are widely prescribed, it is important to clarify this latter issue. In this cohort study—an observational study that follows a group of individuals who are identical with the exception of exposure to a specific factor to determine whether exposure to that factor increases the likelihood of a specific outcome—the researchers investigate the association between violent crime and SSRIs in Sweden. What Did the Researchers Do and Find? The researchers extracted information on SSRIs prescribed in Sweden between 2006 and 2009 from the Swedish Prescribed Drug Register and information on convictions for violent crimes for the same period from the Swedish national crime register. They then compared the rate of violent crime while individuals were prescribed SSRIs with the rate of violent crime in the same individuals while not receiving medication. This “within-individual” design accounts for time-invariant factors such as genetic and early environmental factors that might otherwise lead to confounding. In observational studies, participants exposed to a specific factor can also share another unknown characteristic (confounder) that is actually responsible for the outcome of interest. During the study period, about 850,000 individuals (10.8% of the Swedish population) were prescribed SSRIs, and 1% of these individuals were convicted of a violent crime. Using within-individual statistical models, there was a significant but modest overall association (an association unlikely to have occurred by chance) between SSRIs and convictions for violent crime. After adjustment for age, the association between SSRIs and convictions for violent crimes remained significant for individuals (males and females combined or males and females considered separately) aged 15 to 24 years but became non-significant among older individuals. What Do These Findings Mean? These findings show an association between SSRIs and violent crime that varies by age group. They cannot, however, prove that taking SSRIs actually causes an increase in violent crime among young people because the analytical approach used does not fully account for time-varying risk factors such as symptom severity or alcohol misuse that might affect an individual’s risk of committing a violent crime (residual confounding). In addition, some people who committed a violent crime might have subsequently taken SSRIs to cope with the anxiety and stress of arrest (reverse causation). The lack of a significant association between SSRIs and violent crime among most people taking SSRIs is reassuring; the association between violent crimes and SSRIs among individuals younger than 25 years is worrying. However, this finding needs confirming in studies with other designs undertaken in other settings. If confirmed, warnings about the increased risk of violent behavior among young people when being treated with SSRIs might be needed. But, note the researchers, it might be inappropriate to restrict the use of SSRIs in this age group because increases in adverse outcomes associated with poorly treated depression, such as suicide, might outweigh the public health benefit accruing from decreases in violence. Additional Information. This list of resources contains links that can be accessed when viewing the PDF on a device or via the online version of the article at http://dx.doi.org/10.1371/journal.pmed.1001875. The UK National Health Service Choices website provides information about depression (including personal stories) and about SSRIs; a “behind the headlines” article discusses a research article on recent increases in the use of SSRIs across Europe The UK Royal College of Psychiatrists provides leaflets on depression and on antidepressants Mind, a UK noMind, a UK not-for-profit organization, also provides information about depression (including personal stories) and about antidepressants The US National Institute of Mental Health provides information about depression and about antidepressant medications for children and adolescents MedlinePlus provides links to additional resources about depression and antidepressants Introduction Selective serotonin reuptake inhibitors (SSRIs) are among the most widely prescribed psychiatric medications in many countries [1–6]. At the same time, concerns about their adverse effects, including suicide and violence, have been widely discussed and remain controversial. Observational and trial data have shown that although SSRIs appear not to elevate the risk for suicidal behaviour in adults, they may increase the risk of suicide ideation in children, adolescents, and young adults. This weak age-related association is consistent across studies [7–11] but inconsistently supported by ecological data [12–15]. Despite a number of legal cases linking SSRIs and violent behaviour [16], empirical research on the association is limited and inconclusive. Ecological studies suggest that increased SSRI prescriptions have been associated with decreases in violent crimes in the US [17] and lethal violence in the Netherlands [18]. In contrast, an expert review of clinical trials concluded that there was an excess of violence in both adults and children on SSRIs compared with placebo [16]. Furthermore, drug safety (or pharmacovigilance) data have shown a disproportionate association between SSRIs and violent behaviours [19] and serious violent acts [20], and an observational study found an association of work-related violence with antidepressant purchases [21]. However, these study designs are limited: findings from ecological data fail to relate the use of SSRIs at the individual level and are liable to be influenced by secular changes, including legislation, reporting of violence, and unaccounted changes in the impact of other risk factors such as drug and alcohol use [15,22]. Pharmacovigilance data are subject to reporting bias, changes in patient awareness about adverse outcomes, confounding by indication, and failure to account for exposure to other medications [23]. Pharmacoepidemiological studies provide one approach to deal with these limitations [12,23]. Our objective was thus to investigate the association between SSRIs and violence outcomes by linking data from Swedish national registers on individual SSRI prescriptions, use of other psychotropic drugs, and violent crimes in a large population-based cohort. We have primarily used a “within-individual” design [24–27], where the risk of violent crime is determined when an individual is taking an SSRI as compared to when the same person is not. Using this design, all time-invariant factors (i.e., genetic factors, all factors before the start of follow-up, and factors that remain constant during follow-up) are accounted for; thus, this design more fully adjusts for unmeasured time-invariant confounding and confounding by indication than other observational designs, but does not account for time-varying factors such as symptom severity. Our null hypothesis was that no associations between SSRI medication and violent outcomes would be demonstrated using a within-individual design, including in different age groups. Methods In the total population of Sweden aged 15 y or older in 2006 (n = 7,917,854) and residing in Sweden during follow-up (January 1, 2006, to December 31, 2009), we identified 856,493 individuals who were prescribed SSRI treatment. Information on individuals receiving SSRI treatment was collected from Swedish population-based registers with national coverage, and registers were linked using each individual’s unique identification number. The project was approved by the ethics committee at Karolinska Institutet (2005/4:5). Measures SSRI treatment. Information on medication and the date prescriptions were dispensed was extracted from the Swedish Prescribed Drug Register, with complete national data on all prescribed and dispensed medical drugs from all pharmacies in Sweden since July 2005 [28]. A previous comparison between post-mortem toxicology and SSRI purchases in the Swedish Prescribed Drug Register indicated good medication compliance [29]. In our initial analysis, we included all individuals with dispensed SSRI prescriptions. However, as prescriptions are typically restricted to at most 3 mo and we wanted to restrict the sample to those adherent to SSRIs, individuals with a single SSRI prescription within a 6-mo period were excluded from stratified and sensitivity analyses as no assumptions could be made about their medication adherence. A separate analysis was also carried out including only individuals with a single dispensed prescription. A treatment period was thus defined as a series of SSRI prescriptions with no more than 6 mo between two consecutive prescriptions. The start of a treatment period was defined as the date an SSRI prescription was first dispensed during our follow-up. The end of a treatment period was defined as the date that the last SSRI prescription in that treatment period was dispensed. Periods of more than 6 mo between prescriptions were considered non-treatment periods. A new treatment period was considered to have started at the first date of the next series of consecutive prescriptions (see S1 Methods for details on SSRI medications). For individuals with a single prescription, the start of their treatment period was defined as the date their prescription was dispensed, and the end of that treatment period was defined as 14 d after the prescription was dispensed. Other psychotropic medications. Adjustments were made for concurrent psychotropic medications other than SSRIs, which included antipsychotics, hypnotics, sedatives, anxiolytics, drugs used in addictive disorders, mood stabilisers, antiepileptics, and antidepressant medications other than SSRIs (venlafaxine, duloxetine, tricyclics, heterocyclics, mirtazapine, non-selective monoamine oxidase inhibitors, moclobemide, and bupropion). Treatment periods were defined in the same manner as SSRI treatment periods (see S1 Methods for details). Violent crimes. Information on convictions for violent crimes for individuals aged 15 y and older (the age of criminal responsibility) was extracted from the Swedish national crime register. Violent crimes were defined as crimes against persons as per previous work [30], and included attempted, completed, and aggravated forms of homicide, manslaughter, unlawful threats, harassment, robbery, arson, assault, assault on an official, kidnapping, stalking, coercion, and all sexual offences (see S1 Methods for more details). Alternative outcomes. Examinations of individual SSRIs and alternative outcomes were also carried out, including (1) convictions for substance-related crimes, (2) convictions for non-violent crimes, (3) arrests with preliminary investigations (hereafter “arrests”, as distinct from convictions; described as “suspicions” in the Swedish crime register) for violent crimes, (4) arrests for substance-related crimes, (5) arrests for non-violent crimes, (6) non-fatal injuries (hospitalisations) from accidents; (7) emergency inpatient or outpatient treatment for alcohol intoxication or misuse, (8) and psychiatric hospitalisations (see S1 Methods for details on alternative outcomes). Statistical Analyses Individuals were followed from January 1, 2006, to December 31, 2009, and follow-up was adjusted for migration, periods in prison or institutional youth care, hospitalisation, and death through linkage to the Swedish migration, prison, patient, and cause of death registers. Unobservable time, i.e., time abroad, in prison, or in hospital, was removed (truncated) from the follow-up time. Time after hospital discharge, release from prison, or immigration was added to the observable cohort again. A between-individual Cox proportional hazards regression compared the average rate of violent crime convictions during SSRI medication with the rate during non-medication for all individuals. In this analysis, follow-up period was split into the period before the first outcome, periods between outcomes, and the period after the last outcome. Time at risk was measured from the start of each period, and medication was used as a time-varying covariate. Robust standard errors were calculated to account for correlations between periods within the same individual. This analysis was adjusted for sex and age. The principal analyses were within-individual stratified Cox proportional hazards regressions, with each individual entering as a separate stratum in the analysis and serving as his/her own control. The obtained hazard ratio (HR) is thus adjusted for (i.e., stratified by) all potential time-invariant confounders within each individual. To adjust for age, which is a time-varying potential confounder, age was added to the model as a time-varying covariate, with one factor for each whole year. In the within-individual stratified Cox proportional hazards regression, only individuals who changed medication status contributed directly to the estimate. All other individuals contributed indirectly through the estimates of other covariates. Since the covariates in the within-individual stratified Cox proportional hazards regression were time-varying, we did not test for the proportional hazards assumption. More information on this approach is provided in [31]; this approach has been applied in studies of attention deficit hyperactivity disorder medication, antipsychotics, and mood stabilisers [24–27]. To ensure that outcomes were measured appropriately, all crimes were included from the date of perpetration (rather than conviction), and those with uncertain date of perpetration were excluded from the analyses, resulting in the exclusion of 1.3% (1,241) of violent crime convictions, 1.0% (9,108) of non-violent crime convictions, and 1.8% (5,187) of substance-related convictions during the period from 2006 to 2009. To test for confounding by other psychotropic medications, we first adjusted for concurrent exposure to other psychotropic medications as a time-varying covariate. Then we excluded individuals with other psychotropic medications during follow-up from the within-individual stratified Cox proportional hazards regression. Analyses were also stratified by sex, by age (from age 15 y, the age of criminal responsibility, in 10-y bands [32] up to age 44 y; the age bands for ages 45 y and over were combined as event rates were low), and by type of SSRI medication (fluoxetine, citalopram, paroxetine, sertraline, or escitalopram). To estimate cumulative exposure to SSRIs, the defined daily dose (DDD) of SSRI medication [33] was calculated through summing dispensed medication and then dividing the sum by the number of days in the treatment period. DDDs were categorised into four groups; (1) no exposure, (2) low SSRI exposure (<1 DDD/day), (3) moderate SSRI exposure (1–2 DDD/day), and (4) high SSRI exposure (>2 DDD/day). Sensitivity analyses. In sensitivity analyses, within-individual stratified Cox proportional hazards regressions were carried out with the following alternative outcomes: convictions for non-violent crimes, convictions for substance-related crimes, arrests for violent crimes, arrests for non-violent crimes, arrests for substance-related crimes, non-fatal injuries from accidents, emergency treatment for alcohol intoxication or misuse, and psychiatric hospitalisations. Furthermore, each SSRI medication was analysed separately, and periods of using of two or more SSRI medications were excluded to adjust for switching effects between SSRI medications. Furthermore, all SSRIs were entered in the same model as covariates to adjust for concurrent use of other SSRIs. Analyses were also stratified by type of SSRI medication with violent crime arrests as an alternative outcome. Additionally, other antidepressants (venlafaxine, duloxetine, tricyclics, heterocyclics, mirtazapine, moclobemide, and bupropion) were used as an alternative exposure for violent crime convictions. Further sensitivity analyses were carried out to test for non-specific treatment effects where diuretics were used as an alternative exposure for violent crime convictions to test the model. For individuals who started SSRI treatment after being convicted of a violent crime, the number of days between the date of committing the crime and the start of SSRI treatment was calculated. To exclude the possibility of reverse causation, i.e., if committing a violent crime increased the probability of subsequent SSRI treatment, new within-individual stratified Cox proportional hazards regressions were carried out excluding from the analysis all individuals who received SSRI treatment within 7, 14, 30, or 60 d after committing a violent crime. Finally, the robustness of results was tested by undertaking four alternative analyses. First, a conditional Poisson regression examined how changes in medication exposure were associated with changes in violent crime convictions within the same person, thus adjusting for time-invariant confounders. Second, we repeated the main models with different definitions of a treatment period: (1) a series of SSRI prescriptions with no more than 3 mo between two consecutive prescriptions and (2) a series of SSRI prescriptions with no more than 4 mo between two consecutive prescriptions. Third, we tested for delayed onset of action of SSRIs by setting the first day of the treatment period to 8 wk after the date of the first dispensed prescription. Fourth, we tested for SSRI discontinuation effects by extending the end of the treatment period to 3 wk and 12 wk after the date that the last SSRI prescription in a treatment period was dispensed. SAS version 9.4 (SAS Institute) was used for all analyses, except for the conditional Poisson regression, for which STATA 13.1 (StataCorp) was used. For SAS, software function “proc phreg” was used for both stratified and marginal Cox regressions, and for STATA, software function “xtpoisson” was used for the conditional Poisson regression. STROBE guidelines were followed (S1 STROBE). Measures SSRI treatment. Information on medication and the date prescriptions were dispensed was extracted from the Swedish Prescribed Drug Register, with complete national data on all prescribed and dispensed medical drugs from all pharmacies in Sweden since July 2005 [28]. A previous comparison between post-mortem toxicology and SSRI purchases in the Swedish Prescribed Drug Register indicated good medication compliance [29]. In our initial analysis, we included all individuals with dispensed SSRI prescriptions. However, as prescriptions are typically restricted to at most 3 mo and we wanted to restrict the sample to those adherent to SSRIs, individuals with a single SSRI prescription within a 6-mo period were excluded from stratified and sensitivity analyses as no assumptions could be made about their medication adherence. A separate analysis was also carried out including only individuals with a single dispensed prescription. A treatment period was thus defined as a series of SSRI prescriptions with no more than 6 mo between two consecutive prescriptions. The start of a treatment period was defined as the date an SSRI prescription was first dispensed during our follow-up. The end of a treatment period was defined as the date that the last SSRI prescription in that treatment period was dispensed. Periods of more than 6 mo between prescriptions were considered non-treatment periods. A new treatment period was considered to have started at the first date of the next series of consecutive prescriptions (see S1 Methods for details on SSRI medications). For individuals with a single prescription, the start of their treatment period was defined as the date their prescription was dispensed, and the end of that treatment period was defined as 14 d after the prescription was dispensed. Other psychotropic medications. Adjustments were made for concurrent psychotropic medications other than SSRIs, which included antipsychotics, hypnotics, sedatives, anxiolytics, drugs used in addictive disorders, mood stabilisers, antiepileptics, and antidepressant medications other than SSRIs (venlafaxine, duloxetine, tricyclics, heterocyclics, mirtazapine, non-selective monoamine oxidase inhibitors, moclobemide, and bupropion). Treatment periods were defined in the same manner as SSRI treatment periods (see S1 Methods for details). Violent crimes. Information on convictions for violent crimes for individuals aged 15 y and older (the age of criminal responsibility) was extracted from the Swedish national crime register. Violent crimes were defined as crimes against persons as per previous work [30], and included attempted, completed, and aggravated forms of homicide, manslaughter, unlawful threats, harassment, robbery, arson, assault, assault on an official, kidnapping, stalking, coercion, and all sexual offences (see S1 Methods for more details). Alternative outcomes. Examinations of individual SSRIs and alternative outcomes were also carried out, including (1) convictions for substance-related crimes, (2) convictions for non-violent crimes, (3) arrests with preliminary investigations (hereafter “arrests”, as distinct from convictions; described as “suspicions” in the Swedish crime register) for violent crimes, (4) arrests for substance-related crimes, (5) arrests for non-violent crimes, (6) non-fatal injuries (hospitalisations) from accidents; (7) emergency inpatient or outpatient treatment for alcohol intoxication or misuse, (8) and psychiatric hospitalisations (see S1 Methods for details on alternative outcomes). SSRI treatment. Information on medication and the date prescriptions were dispensed was extracted from the Swedish Prescribed Drug Register, with complete national data on all prescribed and dispensed medical drugs from all pharmacies in Sweden since July 2005 [28]. A previous comparison between post-mortem toxicology and SSRI purchases in the Swedish Prescribed Drug Register indicated good medication compliance [29]. In our initial analysis, we included all individuals with dispensed SSRI prescriptions. However, as prescriptions are typically restricted to at most 3 mo and we wanted to restrict the sample to those adherent to SSRIs, individuals with a single SSRI prescription within a 6-mo period were excluded from stratified and sensitivity analyses as no assumptions could be made about their medication adherence. A separate analysis was also carried out including only individuals with a single dispensed prescription. A treatment period was thus defined as a series of SSRI prescriptions with no more than 6 mo between two consecutive prescriptions. The start of a treatment period was defined as the date an SSRI prescription was first dispensed during our follow-up. The end of a treatment period was defined as the date that the last SSRI prescription in that treatment period was dispensed. Periods of more than 6 mo between prescriptions were considered non-treatment periods. A new treatment period was considered to have started at the first date of the next series of consecutive prescriptions (see S1 Methods for details on SSRI medications). For individuals with a single prescription, the start of their treatment period was defined as the date their prescription was dispensed, and the end of that treatment period was defined as 14 d after the prescription was dispensed. Other psychotropic medications. Adjustments were made for concurrent psychotropic medications other than SSRIs, which included antipsychotics, hypnotics, sedatives, anxiolytics, drugs used in addictive disorders, mood stabilisers, antiepileptics, and antidepressant medications other than SSRIs (venlafaxine, duloxetine, tricyclics, heterocyclics, mirtazapine, non-selective monoamine oxidase inhibitors, moclobemide, and bupropion). Treatment periods were defined in the same manner as SSRI treatment periods (see S1 Methods for details). Violent crimes. Information on convictions for violent crimes for individuals aged 15 y and older (the age of criminal responsibility) was extracted from the Swedish national crime register. Violent crimes were defined as crimes against persons as per previous work [30], and included attempted, completed, and aggravated forms of homicide, manslaughter, unlawful threats, harassment, robbery, arson, assault, assault on an official, kidnapping, stalking, coercion, and all sexual offences (see S1 Methods for more details). Alternative outcomes. Examinations of individual SSRIs and alternative outcomes were also carried out, including (1) convictions for substance-related crimes, (2) convictions for non-violent crimes, (3) arrests with preliminary investigations (hereafter “arrests”, as distinct from convictions; described as “suspicions” in the Swedish crime register) for violent crimes, (4) arrests for substance-related crimes, (5) arrests for non-violent crimes, (6) non-fatal injuries (hospitalisations) from accidents; (7) emergency inpatient or outpatient treatment for alcohol intoxication or misuse, (8) and psychiatric hospitalisations (see S1 Methods for details on alternative outcomes). Statistical Analyses Individuals were followed from January 1, 2006, to December 31, 2009, and follow-up was adjusted for migration, periods in prison or institutional youth care, hospitalisation, and death through linkage to the Swedish migration, prison, patient, and cause of death registers. Unobservable time, i.e., time abroad, in prison, or in hospital, was removed (truncated) from the follow-up time. Time after hospital discharge, release from prison, or immigration was added to the observable cohort again. A between-individual Cox proportional hazards regression compared the average rate of violent crime convictions during SSRI medication with the rate during non-medication for all individuals. In this analysis, follow-up period was split into the period before the first outcome, periods between outcomes, and the period after the last outcome. Time at risk was measured from the start of each period, and medication was used as a time-varying covariate. Robust standard errors were calculated to account for correlations between periods within the same individual. This analysis was adjusted for sex and age. The principal analyses were within-individual stratified Cox proportional hazards regressions, with each individual entering as a separate stratum in the analysis and serving as his/her own control. The obtained hazard ratio (HR) is thus adjusted for (i.e., stratified by) all potential time-invariant confounders within each individual. To adjust for age, which is a time-varying potential confounder, age was added to the model as a time-varying covariate, with one factor for each whole year. In the within-individual stratified Cox proportional hazards regression, only individuals who changed medication status contributed directly to the estimate. All other individuals contributed indirectly through the estimates of other covariates. Since the covariates in the within-individual stratified Cox proportional hazards regression were time-varying, we did not test for the proportional hazards assumption. More information on this approach is provided in [31]; this approach has been applied in studies of attention deficit hyperactivity disorder medication, antipsychotics, and mood stabilisers [24–27]. To ensure that outcomes were measured appropriately, all crimes were included from the date of perpetration (rather than conviction), and those with uncertain date of perpetration were excluded from the analyses, resulting in the exclusion of 1.3% (1,241) of violent crime convictions, 1.0% (9,108) of non-violent crime convictions, and 1.8% (5,187) of substance-related convictions during the period from 2006 to 2009. To test for confounding by other psychotropic medications, we first adjusted for concurrent exposure to other psychotropic medications as a time-varying covariate. Then we excluded individuals with other psychotropic medications during follow-up from the within-individual stratified Cox proportional hazards regression. Analyses were also stratified by sex, by age (from age 15 y, the age of criminal responsibility, in 10-y bands [32] up to age 44 y; the age bands for ages 45 y and over were combined as event rates were low), and by type of SSRI medication (fluoxetine, citalopram, paroxetine, sertraline, or escitalopram). To estimate cumulative exposure to SSRIs, the defined daily dose (DDD) of SSRI medication [33] was calculated through summing dispensed medication and then dividing the sum by the number of days in the treatment period. DDDs were categorised into four groups; (1) no exposure, (2) low SSRI exposure (<1 DDD/day), (3) moderate SSRI exposure (1–2 DDD/day), and (4) high SSRI exposure (>2 DDD/day). Sensitivity analyses. In sensitivity analyses, within-individual stratified Cox proportional hazards regressions were carried out with the following alternative outcomes: convictions for non-violent crimes, convictions for substance-related crimes, arrests for violent crimes, arrests for non-violent crimes, arrests for substance-related crimes, non-fatal injuries from accidents, emergency treatment for alcohol intoxication or misuse, and psychiatric hospitalisations. Furthermore, each SSRI medication was analysed separately, and periods of using of two or more SSRI medications were excluded to adjust for switching effects between SSRI medications. Furthermore, all SSRIs were entered in the same model as covariates to adjust for concurrent use of other SSRIs. Analyses were also stratified by type of SSRI medication with violent crime arrests as an alternative outcome. Additionally, other antidepressants (venlafaxine, duloxetine, tricyclics, heterocyclics, mirtazapine, moclobemide, and bupropion) were used as an alternative exposure for violent crime convictions. Further sensitivity analyses were carried out to test for non-specific treatment effects where diuretics were used as an alternative exposure for violent crime convictions to test the model. For individuals who started SSRI treatment after being convicted of a violent crime, the number of days between the date of committing the crime and the start of SSRI treatment was calculated. To exclude the possibility of reverse causation, i.e., if committing a violent crime increased the probability of subsequent SSRI treatment, new within-individual stratified Cox proportional hazards regressions were carried out excluding from the analysis all individuals who received SSRI treatment within 7, 14, 30, or 60 d after committing a violent crime. Finally, the robustness of results was tested by undertaking four alternative analyses. First, a conditional Poisson regression examined how changes in medication exposure were associated with changes in violent crime convictions within the same person, thus adjusting for time-invariant confounders. Second, we repeated the main models with different definitions of a treatment period: (1) a series of SSRI prescriptions with no more than 3 mo between two consecutive prescriptions and (2) a series of SSRI prescriptions with no more than 4 mo between two consecutive prescriptions. Third, we tested for delayed onset of action of SSRIs by setting the first day of the treatment period to 8 wk after the date of the first dispensed prescription. Fourth, we tested for SSRI discontinuation effects by extending the end of the treatment period to 3 wk and 12 wk after the date that the last SSRI prescription in a treatment period was dispensed. SAS version 9.4 (SAS Institute) was used for all analyses, except for the conditional Poisson regression, for which STATA 13.1 (StataCorp) was used. For SAS, software function “proc phreg” was used for both stratified and marginal Cox regressions, and for STATA, software function “xtpoisson” was used for the conditional Poisson regression. STROBE guidelines were followed (S1 STROBE). Sensitivity analyses. In sensitivity analyses, within-individual stratified Cox proportional hazards regressions were carried out with the following alternative outcomes: convictions for non-violent crimes, convictions for substance-related crimes, arrests for violent crimes, arrests for non-violent crimes, arrests for substance-related crimes, non-fatal injuries from accidents, emergency treatment for alcohol intoxication or misuse, and psychiatric hospitalisations. Furthermore, each SSRI medication was analysed separately, and periods of using of two or more SSRI medications were excluded to adjust for switching effects between SSRI medications. Furthermore, all SSRIs were entered in the same model as covariates to adjust for concurrent use of other SSRIs. Analyses were also stratified by type of SSRI medication with violent crime arrests as an alternative outcome. Additionally, other antidepressants (venlafaxine, duloxetine, tricyclics, heterocyclics, mirtazapine, moclobemide, and bupropion) were used as an alternative exposure for violent crime convictions. Further sensitivity analyses were carried out to test for non-specific treatment effects where diuretics were used as an alternative exposure for violent crime convictions to test the model. For individuals who started SSRI treatment after being convicted of a violent crime, the number of days between the date of committing the crime and the start of SSRI treatment was calculated. To exclude the possibility of reverse causation, i.e., if committing a violent crime increased the probability of subsequent SSRI treatment, new within-individual stratified Cox proportional hazards regressions were carried out excluding from the analysis all individuals who received SSRI treatment within 7, 14, 30, or 60 d after committing a violent crime. Finally, the robustness of results was tested by undertaking four alternative analyses. First, a conditional Poisson regression examined how changes in medication exposure were associated with changes in violent crime convictions within the same person, thus adjusting for time-invariant confounders. Second, we repeated the main models with different definitions of a treatment period: (1) a series of SSRI prescriptions with no more than 3 mo between two consecutive prescriptions and (2) a series of SSRI prescriptions with no more than 4 mo between two consecutive prescriptions. Third, we tested for delayed onset of action of SSRIs by setting the first day of the treatment period to 8 wk after the date of the first dispensed prescription. Fourth, we tested for SSRI discontinuation effects by extending the end of the treatment period to 3 wk and 12 wk after the date that the last SSRI prescription in a treatment period was dispensed. SAS version 9.4 (SAS Institute) was used for all analyses, except for the conditional Poisson regression, for which STATA 13.1 (StataCorp) was used. For SAS, software function “proc phreg” was used for both stratified and marginal Cox regressions, and for STATA, software function “xtpoisson” was used for the conditional Poisson regression. STROBE guidelines were followed (S1 STROBE). Results Sample Description Of 7,917,854 individuals in the general population investigated (individuals in Sweden aged 15 y or older in 2006), 856,493 (10.8%) were prescribed SSRIs during the time period 2006–2009, or 14.1% of all women and 7.5% of all men in the investigated population (see Table 1 for background characteristics). Of those prescribed SSRIs, 9.9% were aged 15–24 y, 12.7% were aged 25–34 y, 16.5% were aged 35–44 y, 15.6% were aged 45–54 y, 15.5% were aged 55–64 y, and 29.7% were aged 65 y or over at baseline in 2006. In the SSRI cohort, 8,377 individuals (1.0%) were convicted of a violent crime during the period 2006–2009. Among the individuals who were prescribed SSRI treatment, 65,862 individuals were prescribed fluoxetine, 389,857 citalopram, 46,615 paroxetine, 215,873 sertraline, 1,198 fluvoxamine, and 84,934 escitalopram. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 1. Characteristics at baseline and during follow-up for SSRI-medicated and non-medicated individuals in a population sample in Sweden 2006–2009. https://doi.org/10.1371/journal.pmed.1001875.t001 Main Analyses Within-individual Cox proportional hazards analyses were carried out to compare violent crime rates within the same individuals during periods when they were on medication compared to periods when they were not, and the results showed an increased risk of violent crime conviction during medicated periods (HR = 1.19, 95% CI 1.08–1.32, p < 0.001; Table 2). The estimated hazard did not materially change when we adjusted for concurrent psychotropic medications (HR = 1.22, 95% CI 1.11–1.32, p < 0.001), nor when we excluded individuals with only one dispensed prescription (HR = 1.22, 95% CI 1.10–1.35, p < 0.001). Additionally, when we excluded all individuals who had received other psychotropic medications during follow-up from the analysis, the estimated hazard was similar (HR = 1.20, 95% CI 1.04–1.38, p = 0.014). The between-individual Cox proportional hazards analysis also demonstrated an association between SSRI prescriptions and being convicted of a violent crime (HR = 2.66, 95% CI 2.54–2.78, p < 0.001) when comparing individuals on SSRIs to individuals who were not taking SSRIs. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 2. Violent crime convictions in 2006–2009 in individuals treated with SSRI medication as compared to non-medicated individuals, and comparing treatment to non-treatment periods within the same person. https://doi.org/10.1371/journal.pmed.1001875.t002 The analyses were then stratified by sex and age band (Table 3). This demonstrated an increased risk of violent crime conviction for those aged 15 to 24 y (HR = 1.43, 95% CI 1.19–1.73, p < 0.001) but not for the other age bands investigated (25–34 y, 35–44 y, and 45 y and older). When stratified by sex and age, associations were significant for both genders in the age group 15 to 24 y (HR = 1.40, 95% CI 1.13–1.73, p = 0.002, and HR = 1.75, 95% CI 1.08–2.84, p = 0.023, for males and females respectively). Next, the role of cumulative SSRI exposure was examined using DDDs. The results showed that low SSRI exposure was associated with an increased risk of being convicted of a violent crime as compared to periods of non-exposure (HR = 1.27, 95% CI 1.10–1.47, p = 0.001). However, no significant association with violent crime conviction was found for periods of moderate or high SSRI exposure (Table 3). Download: PPT PowerPoint slide PNG larger image TIFF original image Table 3. Violent crime convictions in 2006–2009 in individuals treated with SSRI medication compared to non-treatment periods in the same person stratified by sex, age, dose, and medication type using stratified Cox regression models. https://doi.org/10.1371/journal.pmed.1001875.t003 Sensitivity Analyses In sensitivity analyses, our results showed some differences for individual SSRIs; there was a significantly higher hazard for violent crime conviction in individuals prescribed sertraline (Table 3) and for violent crime conviction in individuals prescribed citalopram and sertraline after eliminating periods of concurrent use of two different SSRIs, thus adjusting for switching effects between SSRIs (Table 4). For violent crime arrests, the increased association with citalopram remains (Table 4). Download: PPT PowerPoint slide PNG larger image TIFF original image Table 4. Sensitivity analyses: rates of different adverse outcomes in individuals treated with SSRI medication and other antidepressants compared to non-treatment periods in the same person using stratified Cox regression models. https://doi.org/10.1371/journal.pmed.1001875.t004 In further analyses, the relationship between SSRI treatment and other outcomes was examined (Table 4), and the results showed an increased risk of violent crime arrests, non-violent crime convictions, and non-violent crime arrests with SSRI treatment. Furthermore, an increased risk of non-fatal injuries from accidents was found (HR = 1.20, 95% CI 1.18–1.23, p < 0.001). The possible role of alcohol misuse as a time-varying confounder was tested by using emergency inpatient and or outpatient treatment for alcohol intoxication or misuse as an outcome, showing an increased risk during times of medication (HR = 1.06, 95% CI 1.03–1.09, p < 0.001). The risk of hospitalisation for psychiatric care was also examined, showing a slightly decreased risk with SSRI treatment (HR = 0.96, 95% CI 0.93–0.99, p = 0.012). When we investigated other antidepressant classes, we found a significant association between medication use and violent crime conviction for individuals prescribed venlafaxine. The risk of being convicted of a violent crime was reduced when on mirtazapine. Finally, an inverse association between violent crime conviction and diuretics was found using the within-individual model (HR = 0.80, 95% CI 0.67–0.95, p = 0.012). When all analyses were stratified by age (S1 Table), the increased risk of being convicted of a violent crime remained in individuals aged 15 to 24 y after adjustment for concurrent psychotropic medications (HR = 1.45, 95% CI 1.21–1.74, p < 0.001). Results also showed that low SSRI exposure was associated with an increased risk of being convicted of a violent crime in this age band only (HR = 1.62, 95% CI 1.23–2.13, p < 0.002). Furthermore, significant associations were shown for violent crime arrests and non-violent crime arrests and convictions for individuals aged 15 to 24 y, and also for individuals aged 25 to 34 y, although associations were weaker in the latter age band. The increased risk of non-fatal injuries from accidents remained significant for all ages. Results also showed that individuals aged 15–24, 25–34, and 35–44 y had an increased risk of emergency inpatient or outpatient treatment for alcohol intoxication or misuse (HR = 1.98, 95% CI 1.76–2.21, p < 0.001; HR = 1.33, 95% CI 1.21–1.46, p < 0.001; HR = 1.08, 95% CI 1.01–1.14, p = 0.015, respectively). However, individuals aged 45 y and older showed a slightly decreased risk of emergency inpatient or outpatient treatment for alcohol intoxication or misuse (HR = 0.96, 95% CI 0.93–0.99, p = 0.028). To test whether individuals who had been dispensed only one prescription differed from the rest of the cohort, we also carried out a within-individual analysis including these individuals only. No significant association of SSRI treatment with violent crime conviction was found for this group (HR = 0.73, 95% CI 0.45–1.17, p = 0.193). To account for the possibility of reverse causation, i.e., that individuals are more likely to take SSRIs after committing a crime, we excluded 996 individuals who received SSRIs within 60 d of committing a violent crime, and the risk increase remained (S2 Table). We then excluded 608 individuals who received SSRIs within 30 d of committing of a violent crime, and the risk increase remained similar (S2 Table). When we excluded those who received medication within 14 d (356 individuals) or 7 d (197 individuals) of committing a violent crime, similar risk increases were found (S2 Table). No material differences were found when we repeated this analysis with violent crime arrests as an outcome (S2 Table). When we carried out a conditional Poisson regression, a similar pattern of findings was found (incidence rate ratio for violent crime conviction = 1.18, 95% CI 1.09–1.27, p = 0.001). When treatment periods were defined as no breaks in prescription coverage of more than 3 or 4 mo, instead of 6 mo, no material differences were found in the within-individual models (S2 Table). When we tested for delayed treatment effects, no material differences were found for the association between SSRI treatment and violent crime convictions when the treatment period was considered to start 8 wk after the SSRI prescription was dispensed (HR = 1.21, 95% CI 1.02–1.43, p = 0.031). Similar effects were found when testing for SSRI discontinuation effects up to 3 wk or 12 weeks, respectively, after the last dispensed prescription (S2 Table). Sample Description Of 7,917,854 individuals in the general population investigated (individuals in Sweden aged 15 y or older in 2006), 856,493 (10.8%) were prescribed SSRIs during the time period 2006–2009, or 14.1% of all women and 7.5% of all men in the investigated population (see Table 1 for background characteristics). Of those prescribed SSRIs, 9.9% were aged 15–24 y, 12.7% were aged 25–34 y, 16.5% were aged 35–44 y, 15.6% were aged 45–54 y, 15.5% were aged 55–64 y, and 29.7% were aged 65 y or over at baseline in 2006. In the SSRI cohort, 8,377 individuals (1.0%) were convicted of a violent crime during the period 2006–2009. Among the individuals who were prescribed SSRI treatment, 65,862 individuals were prescribed fluoxetine, 389,857 citalopram, 46,615 paroxetine, 215,873 sertraline, 1,198 fluvoxamine, and 84,934 escitalopram. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 1. Characteristics at baseline and during follow-up for SSRI-medicated and non-medicated individuals in a population sample in Sweden 2006–2009. https://doi.org/10.1371/journal.pmed.1001875.t001 Main Analyses Within-individual Cox proportional hazards analyses were carried out to compare violent crime rates within the same individuals during periods when they were on medication compared to periods when they were not, and the results showed an increased risk of violent crime conviction during medicated periods (HR = 1.19, 95% CI 1.08–1.32, p < 0.001; Table 2). The estimated hazard did not materially change when we adjusted for concurrent psychotropic medications (HR = 1.22, 95% CI 1.11–1.32, p < 0.001), nor when we excluded individuals with only one dispensed prescription (HR = 1.22, 95% CI 1.10–1.35, p < 0.001). Additionally, when we excluded all individuals who had received other psychotropic medications during follow-up from the analysis, the estimated hazard was similar (HR = 1.20, 95% CI 1.04–1.38, p = 0.014). The between-individual Cox proportional hazards analysis also demonstrated an association between SSRI prescriptions and being convicted of a violent crime (HR = 2.66, 95% CI 2.54–2.78, p < 0.001) when comparing individuals on SSRIs to individuals who were not taking SSRIs. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 2. Violent crime convictions in 2006–2009 in individuals treated with SSRI medication as compared to non-medicated individuals, and comparing treatment to non-treatment periods within the same person. https://doi.org/10.1371/journal.pmed.1001875.t002 The analyses were then stratified by sex and age band (Table 3). This demonstrated an increased risk of violent crime conviction for those aged 15 to 24 y (HR = 1.43, 95% CI 1.19–1.73, p < 0.001) but not for the other age bands investigated (25–34 y, 35–44 y, and 45 y and older). When stratified by sex and age, associations were significant for both genders in the age group 15 to 24 y (HR = 1.40, 95% CI 1.13–1.73, p = 0.002, and HR = 1.75, 95% CI 1.08–2.84, p = 0.023, for males and females respectively). Next, the role of cumulative SSRI exposure was examined using DDDs. The results showed that low SSRI exposure was associated with an increased risk of being convicted of a violent crime as compared to periods of non-exposure (HR = 1.27, 95% CI 1.10–1.47, p = 0.001). However, no significant association with violent crime conviction was found for periods of moderate or high SSRI exposure (Table 3). Download: PPT PowerPoint slide PNG larger image TIFF original image Table 3. Violent crime convictions in 2006–2009 in individuals treated with SSRI medication compared to non-treatment periods in the same person stratified by sex, age, dose, and medication type using stratified Cox regression models. https://doi.org/10.1371/journal.pmed.1001875.t003 Sensitivity Analyses In sensitivity analyses, our results showed some differences for individual SSRIs; there was a significantly higher hazard for violent crime conviction in individuals prescribed sertraline (Table 3) and for violent crime conviction in individuals prescribed citalopram and sertraline after eliminating periods of concurrent use of two different SSRIs, thus adjusting for switching effects between SSRIs (Table 4). For violent crime arrests, the increased association with citalopram remains (Table 4). Download: PPT PowerPoint slide PNG larger image TIFF original image Table 4. Sensitivity analyses: rates of different adverse outcomes in individuals treated with SSRI medication and other antidepressants compared to non-treatment periods in the same person using stratified Cox regression models. https://doi.org/10.1371/journal.pmed.1001875.t004 In further analyses, the relationship between SSRI treatment and other outcomes was examined (Table 4), and the results showed an increased risk of violent crime arrests, non-violent crime convictions, and non-violent crime arrests with SSRI treatment. Furthermore, an increased risk of non-fatal injuries from accidents was found (HR = 1.20, 95% CI 1.18–1.23, p < 0.001). The possible role of alcohol misuse as a time-varying confounder was tested by using emergency inpatient and or outpatient treatment for alcohol intoxication or misuse as an outcome, showing an increased risk during times of medication (HR = 1.06, 95% CI 1.03–1.09, p < 0.001). The risk of hospitalisation for psychiatric care was also examined, showing a slightly decreased risk with SSRI treatment (HR = 0.96, 95% CI 0.93–0.99, p = 0.012). When we investigated other antidepressant classes, we found a significant association between medication use and violent crime conviction for individuals prescribed venlafaxine. The risk of being convicted of a violent crime was reduced when on mirtazapine. Finally, an inverse association between violent crime conviction and diuretics was found using the within-individual model (HR = 0.80, 95% CI 0.67–0.95, p = 0.012). When all analyses were stratified by age (S1 Table), the increased risk of being convicted of a violent crime remained in individuals aged 15 to 24 y after adjustment for concurrent psychotropic medications (HR = 1.45, 95% CI 1.21–1.74, p < 0.001). Results also showed that low SSRI exposure was associated with an increased risk of being convicted of a violent crime in this age band only (HR = 1.62, 95% CI 1.23–2.13, p < 0.002). Furthermore, significant associations were shown for violent crime arrests and non-violent crime arrests and convictions for individuals aged 15 to 24 y, and also for individuals aged 25 to 34 y, although associations were weaker in the latter age band. The increased risk of non-fatal injuries from accidents remained significant for all ages. Results also showed that individuals aged 15–24, 25–34, and 35–44 y had an increased risk of emergency inpatient or outpatient treatment for alcohol intoxication or misuse (HR = 1.98, 95% CI 1.76–2.21, p < 0.001; HR = 1.33, 95% CI 1.21–1.46, p < 0.001; HR = 1.08, 95% CI 1.01–1.14, p = 0.015, respectively). However, individuals aged 45 y and older showed a slightly decreased risk of emergency inpatient or outpatient treatment for alcohol intoxication or misuse (HR = 0.96, 95% CI 0.93–0.99, p = 0.028). To test whether individuals who had been dispensed only one prescription differed from the rest of the cohort, we also carried out a within-individual analysis including these individuals only. No significant association of SSRI treatment with violent crime conviction was found for this group (HR = 0.73, 95% CI 0.45–1.17, p = 0.193). To account for the possibility of reverse causation, i.e., that individuals are more likely to take SSRIs after committing a crime, we excluded 996 individuals who received SSRIs within 60 d of committing a violent crime, and the risk increase remained (S2 Table). We then excluded 608 individuals who received SSRIs within 30 d of committing of a violent crime, and the risk increase remained similar (S2 Table). When we excluded those who received medication within 14 d (356 individuals) or 7 d (197 individuals) of committing a violent crime, similar risk increases were found (S2 Table). No material differences were found when we repeated this analysis with violent crime arrests as an outcome (S2 Table). When we carried out a conditional Poisson regression, a similar pattern of findings was found (incidence rate ratio for violent crime conviction = 1.18, 95% CI 1.09–1.27, p = 0.001). When treatment periods were defined as no breaks in prescription coverage of more than 3 or 4 mo, instead of 6 mo, no material differences were found in the within-individual models (S2 Table). When we tested for delayed treatment effects, no material differences were found for the association between SSRI treatment and violent crime convictions when the treatment period was considered to start 8 wk after the SSRI prescription was dispensed (HR = 1.21, 95% CI 1.02–1.43, p = 0.031). Similar effects were found when testing for SSRI discontinuation effects up to 3 wk or 12 weeks, respectively, after the last dispensed prescription (S2 Table). Discussion In this study, we examined the possible association between SSRIs and violent crime using a large population-based cohort that included 856,493 individuals prescribed SSRIs. There were three main findings. First, using a within-individual design, there was an association between SSRI prescriptions and violent crime convictions. With age stratification, there was an increased hazard of violent crime convictions in individuals aged 15 to 24 y, and no significant association in older individuals. A second finding was that the association in individuals aged 15–24 y was consistent when looking at a related antidepressant (venlafaxine, a serotonin–norepinephrine reuptake inhibitor), considering four other outcomes (violent crime arrests, non-violent crime convictions and arrests, and non-fatal accidental injuries), or using another design (conditional Poisson regression). Third, the association of SSRI treatment with violent crime was not found for moderate or high SSRI use, including in those aged 15–24 y. The finding of a modest risk association in younger people is consistent with trial data showing that children and adolescents respond differently than adults to SSRIs [34], and with reported increases in suicide-related outcomes in adolescents prescribed SSRI medication in both observational studies and clinical trials [11,32,35], although this finding is not supported by meta-analyses of trial data [36,37]. These associations may be moderated by impulsivity and risk-taking, which could explain the similar association we report with accidents, and the weaker associations with non-violent crime convictions and arrests. A recent observational investigation also found increases in suicidal behaviour in a large US cohort aged less than 25 y [38]. The US investigation found that younger people receiving the modal antidepressant dose were at increased risk of deliberate self-harm compared to adults, and this risk was further increased in individuals receiving higher doses. The apparent contrast with our findings on medication dose may be because the US study looked at risks in those initiating treatment, while our study examined all treatment periods. Importantly, this US study saw no increased hazard of self-harm in those over 25 y, analogous to our null finding for crime outcomes for those over 25 y. The reasons for the age-dependent differences are still poorly understood, but the adolescent brain may be particularly sensitive to pharmacological interference, as has been demonstrated in animal studies [38–43]. Yet, the possible adverse effects associated with SSRI use appear to be separate from its therapeutic ones; treatment effects have been demonstrated [10,37], particularly for fluoxetine and escitalopram, which are approved by the US Food and Drug Administration for treating adolescent depression [34]. The reported association between SSRIs and violent crime in young people cannot be interpreted causally because of confounding by indication. This confounding was confirmed in our study by the difference between the hazards reported in between-individual and within-individual analyses. Hence we focused on the within-individual analyses: crime outcomes in the same individuals when they were taking SSRIs compared to when they were not taking SSRIs, thus adjusting for all factors that were constant within the individual. However, this approach cannot fully account for time-varying risk factors, such as increased drug or alcohol use during periods of SSRI medication, worsening of symptoms, or a general psychosocial decline. We attempted to address the first of these by investigating substance-related convictions, one proxy for problem substance use, and recorded rates of emergency treatment for alcohol-related problems. Although we did not find an association of SSRI treatment with substance-related convictions, this was a crude outcome with an incidence of 1.7%. An alternative marker of alcohol use was the rate of emergency inpatient or outpatient treatment for alcohol intoxication or misuse, where we found some support for an increasing rate during SSRI medication, which is in keeping with one case series [44]. Although emergency treatment for alcohol intoxication or misuse is a more sensitive measure than alcohol-related crimes, it needs further clarification using prospective clinical designs. Symptom severity may moderate the association between SSRIs and the adverse outcomes reported in this study, and younger people on SSRIs may be less adherent than others, and may have more residual symptoms, such as impulsivity and hostility, which are risk factors for violence [45]. This is underscored by recent epidemiological work that suggested that depression and bipolar disorder are independent risk factors for violent crime [46,47]. If these underlying conditions are partially treated—especially in bipolar patients who are not also prescribed a mood stabiliser [27]—then residual symptoms may partly explain any association. This is further suggested by our finding that the increased hazard for violent crime conviction in younger people was not found in individuals with therapeutic SSRI exposures (≥1 DDD/day). However, the finding that there was a risk increase for non-violent crime arrests and non-violent crime convictions with SSRI use suggests a non-specificity in our findings that could be explained by time-varying confounders, or that the links may be mediated by factors that increase the risk of both violent and non-violent crime. The risk increases for non-violent crime outcomes were smaller than those for violent crime outcomes, which suggests some complexity to the possible mechanisms involved. A final possibility is that non-specific treatment factors, such as contact with health care staff, may partly explain the relationship. As most of the individuals in the sample were outpatients, and unlikely to see health care staff regularly once treatment was initiated, these factors may not be strong. In addition, the finding that violent crime conviction was inversely associated with a group of non-psychotropic medications (diuretics) suggests that, if anything, non-specific treatment effects would reduce any observed association. Another possible challenge to the results is reverse causality—that the observed association was due to individuals taking SSRIs after being arrested for a crime (for various reasons, including coping with the anxiety and stress of arrest or that taking SSRIs might mitigate their criminal sanction). In order to address reverse causality, we excluded all persons who received SSRIs within 7, 14, 30, or 60 d after committing a violent crime, and the association between SSRI treatment and violent crime convictions remained significant, with no material change in risk. Differences between individual SSRIs were examined. The increased association between paroxetine and citalopram use and violent crime arrests could be due to their poorer efficacy and/or shorter half-life compared with other SSRIs [34,48]. Shorter half-lives are linked with withdrawal effects on discontinuation, with increased agitation and possible hostility [34,49]. Further, the single-dose and mean steady-state half-life of SSRIs with short half-lives are shorter in adolescents than in adults, and aggression thus could be a withdrawal rather than side effect [34]. In support of this, venlafaxine and, non-significantly, the heterocyclics were also linked with higher risks of violent crime convictions than SSRIs, and these medications have shorter half-lives and poorer efficacy [34,48]. Moreover, we found an increased risk of violent crime conviction for low SSRI exposure only, which is consistent with the reported links with antidepressants with shorter half-lives and case reports of increased hostility and aggression in children and adolescents at low starting doses in the first weeks of SSRI treatment [50,51]. Escitalopram, with a half-life similar to that of citalopram, was also associated with increased violent crime convictions in younger persons (S1 Table). However, any increased risks of post-cessation withdrawal for violent crime would not be included as related to SSRIs using the design in the current study if the medication was discontinued as planned, and therefore our estimates may be conservative. Our sensitivity analyses to measure post-cessation withdrawal (considering the treatment period to continue up to 3 wk or 12 wk after the date the last SSRI prescription was dispensed) nevertheless showed no material difference in the increased risk of violent crime conviction. An alternative explanation could be that the increased risk of some SSRIs is confounded by psychiatric morbidity; citalopram, escitalopram, and paroxetine are not recommended as first-line treatment for children and adolescents by the Swedish National Board of Health and Welfare, and thus are reserved for treatment-resistant patients with more severe problems [52]. Further work will need to validate differences between SSRIs and dosing strategies, and investigate underlying mechanisms in younger populations. One potentially important explanatory factor will be the timing of doses, which requires further examination. There are two principal clinical implications arising from this study. First, no association between SSRIs and violent crime convictions was found for the majority of people who were prescribed these medications, including individuals aged 25 y and older. Second, the risk increase we report in young people is not insignificant, and hence warrants further examination. If our findings related to young people are validated in other designs, samples, and settings, warnings about an increased risk of violent behaviours while being treated with SSRIs may be needed. Any such changes to the advice given to young persons prescribed SSRIs will need to be carefully considered, as the public health benefit from decreases in violence following restrictions in SSRI use may be countered by increases in other adverse outcomes (such as more disability, rehospitalisation, or suicides) [53]. From a public health perspective, this worsening of overall morbidity and mortality might argue against restrictions on the primary care prescribing of SSRIs as long as potential risks are disclosed [54]. The present study was characterised by several strengths. The study included a large population-based cohort with longitudinal data retrieved from national registers. Information on SSRIs was complete, as each prescription that is dispensed is registered in the Swedish Prescribed Drug Register. Using a within-individual design allowed us to adjust for many unobserved factors that may bias estimates. Although a marginal structural model would have been desirable because of its ability to handle time-varying confounding factors that are also predicted by treatment history, such a model could be used only in a between-individual design that includes measures of all confounding factors. Since many confounders are most likely unobserved, we used a within-individual design as our principal approach. This allowed us to adjust for both measured and unmeasured time-invariant confounding factors, as well as for some measured time-varying confounders that are not predicted by exposure history (like sex and age). Limitations of the study include the use of diagnoses from the national patient register, which only includes diagnoses from specialists. Also, the use of official sources of data for crime outcomes is likely to underestimate true rates of crime and possibly involve selection effects. However, we tried to address such biases by using arrests with preliminary investigations in addition to convictions and also by examining accidents. It is not clear whether these findings will translate to less severe forms of violence or those not reported to the police, and triangulating the findings with information on self- or informant-reported violence will be an important future research direction. Another limitation is that detailed information about the actual prescriptions was not available. Although our data are an improvement over prescription data—as they reflect prescriptions that are dispensed by pharmacies to individuals—we were unable to account for lack of, or variations in, adherence. This problem is parallel to non-adherence in randomised controlled trials, and our within-individual estimate is comparable to the intention-to-treat analysis used in randomised controlled trials. If individuals consumed SSRIs during periods when we assumed that they were not, then this should reduce the hazards reported and would suggest that our estimates are underestimates. A possible source of underestimation is that we excluded persons who were prescribed SSRIs on only one occasion, who may have discontinued the medications due to adverse effects that were not included. We thus carried out analyses where individuals with a single prescription were included, and found no material differences in hazard of violent crime conviction. Another possible source of underestimation is that we used a conservative approach to measure the end of a treatment period (we defined this as the date the last SSRI prescription in a treatment period was dispensed), which could result in slightly lower sensitivity (i.e., individuals classified as unmedicated when truly medicated). However, sensitivity analyses using less conservative approaches to measure the end of a treatment period (3 wk and 12 wk after the last dispensed SSRI prescription in a treatment period) resulted in a similarly increased risk of violent crime conviction. Sweden has prescription rates of SSRIs that are higher than the average for Europe (5-y mean DDD/1,000 individuals/day: Sweden = 70.1; across 29 European countries = 40.0) [14] and similar to the US (10.8% treated in our cohort between 2006–2009 compared to 10.1% treated in the US in 2005) [55]. In relation to criminality, Sweden has similar police-reported assault rates as the US [56]. Finally, there might be residual confounding for the within-individual estimates due to unmeasured time-varying confounders. However, we are not aware of any statistical method that also allows adjustment for unmeasured time-varying confounders. In summary, we demonstrated associations between SSRIs and violent crime that vary by age group. The clinical and public health implications of this require careful consideration, and validation in other settings. Supporting Information S1 STROBE. STROBE checklist. https://doi.org/10.1371/journal.pmed.1001875.s001 (DOCX) S1 Methods. Definitions of SSRIs and other psychotropic medications, crimes, and hospitalisations. https://doi.org/10.1371/journal.pmed.1001875.s002 (DOCX) S1 Table. Violent crime convictions and alternative outcomes in individuals treated with SSRI medication compared to non-treatment periods in the same person, stratified by age using stratified Cox regression models. https://doi.org/10.1371/journal.pmed.1001875.s003 (DOCX) S2 Table. Violent crime convictions and arrests in individuals treated with SSRI medication compared to non-treatment periods in the same person using conditional Poisson regression, stratified Cox regression models for alternative treatment periods, and analyses excluding individuals who received SSRIs after committing a crime. https://doi.org/10.1371/journal.pmed.1001875.s004 (DOCX)

doi: 10.1371/journal.pmed.1001874pmid: 26348214

Background Multistage stepwise HIV testing and treatment initiation procedures can result in lost opportunities to provide timely antiretroviral therapy (ART). Incomplete patient engagement along the continuum of HIV care translates into high levels of preventable mortality. We aimed to evaluate the ability of a simplified test and treat structural intervention to reduce mortality. Methods and Findings In the “pre-intervention 2010” (from January 2010 to December 2010) and “pre-intervention 2011” (from January 2011 to December 2011) phases, patients who screened HIV-positive at health care facilities in Zhongshan and Pubei counties in Guangxi, China, followed the standard-of-care process. In the “post-intervention 2012” (from July 2012 to June 2013) and “post-intervention 2013” (from July 2013 to June 2014) phases, patients who screened HIV-positive at the same facilities were offered a simplified test and treat intervention, i.e., concurrent HIV confirmatory and CD4 testing and immediate initiation of ART, irrespective of CD4 count. Participants were followed for 6–18 mo until the end of their study phase period. Mortality rates in the pre-intervention and post-intervention phases were compared for all HIV cases and for treatment-eligible HIV cases. A total of 1,034 HIV-positive participants (281 and 339 in the two pre-intervention phases respectively, and 215 and 199 in the two post-intervention phases respectively) were enrolled. Following the structural intervention, receipt of baseline CD4 testing within 30 d of HIV confirmation increased from 67%/61% (pre-intervention 2010/pre-intervention 2011) to 98%/97% (post-intervention 2012/post-intervention 2013) (all p < 0.001 [i.e., for all comparisons between a pre- and post-intervention phase]), and the time from HIV confirmation to ART initiation decreased from 53 d (interquartile range [IQR] 27–141)/43 d (IQR 15–113) to 5 d (IQR 2–12)/5 d (IQR 2–13) (all p < 0.001). Initiation of ART increased from 27%/49% to 91%/89% among all cases (all p < 0.001) and from 39%/62% to 94%/90% among individuals with CD4 count ≤ 350 cells/mm3 or AIDS (all p < 0.001). Mortality decreased from 27%/27% to 10%/10% for all cases (all p < 0.001) and from 40%/35% to 13%/13% for cases with CD4 count ≤ 350 cells/mm3 or AIDS (all p < 0.001). The simplified test and treat intervention was significantly associated with decreased mortality rates compared to pre-intervention 2011 (adjusted hazard ratio [aHR] 0.385 [95% CI 0.239–0.620] and 0.380 [95% CI 0.233–0.618] for the two post-intervention phases, respectively, for all newly diagnosed HIV cases [both p < 0.001], and aHR 0.369 [95% CI 0.226–0.603] and 0.361 [95% CI 0.221–0.590] for newly diagnosed treatment-eligible HIV cases [both p < 0.001]). The unit cost of an additional patient receiving ART attributable to the intervention was US$83.80. The unit cost of a death prevented because of the intervention was US$234.52. Conclusions Our results demonstrate that the simplified HIV test and treat intervention promoted successful engagement in care and was associated with a 62% reduction in mortality. Our findings support the implementation of integrated HIV testing and immediate access to ART irrespective of CD4 count, in order to optimize the impact of ART. Background Every year, about 2.1 million people (mostly living in resource-limited countries) are newly infected with HIV, the virus that causes AIDS and that has killed 39 million people over the past three decades. HIV, which is usually transmitted through unprotected sex with an infected individual, gradually destroys CD4 lymphocytes and other immune system cells, leaving HIV-positive individuals susceptible to other infections. Early in the AIDS epidemic, most HIV-positive individuals died within ten years of infection. Then, in 1996, effective antiretroviral therapy (ART) became available, and, for people living in high-income countries, HIV became a chronic condition. But ART was expensive, so HIV/AIDS remained largely untreated and fatal in resource-limited countries. In 2003, the international community began to work towards achieving universal ART coverage. By 2013, about 12.9 million people living with HIV (a third of all HIV-positive people) had access to ART, and the rate of AIDS-related deaths had fallen by a third from its 2005 peak. Why Was This Study Done? Unfortunately, in many countries, late diagnosis of HIV infection, incomplete linkage to care after diagnosis, and high rates of loss to follow-up before and after ART initiation remain major barriers to effective HIV/AIDS treatment and to maximization of the preventative benefits of ART: as well as keeping HIV-positive people healthy, ART also reduces their chances of transmitting HIV to a sexual partner. Here, the researchers investigate whether a simplified “test and treat” intervention can reduce HIV/AIDS mortality (death) rates in China by reducing these barriers. Currently, CD4 testing is offered to people in China only after an initial HIV diagnosis has been confirmed using a second type of test. This standard-of-care policy introduces a delay into ART initiation because a CD4 count below 350 cells/μl blood is the primary determinant of eligibility for treatment through the Chinese national free ART program. By contrast, the simplified test and treat intervention, which is designed to be completed within a week of the patient’s first positive HIV test result, incorporates immediate HIV confirmatory testing, pre-ART CD4 testing, pre-treatment counseling, and ART initiation regardless of CD4 count. What Did the Researchers Do and Find? The researchers followed about 1,000 patients who tested positive for HIV at health care facilities in two counties in Guangxi (one of the Chinese provinces most heavily affected by HIV/AIDS) in two 12-month pre-intervention phases, during which patients followed the standard-of-care process, and two 12-month post-intervention phases, during which patients were offered the simplified test and treat intervention. About 65% and 97% of the patients received baseline CD4 testing during the pre-intervention and post-intervention phases, respectively. Following the structural intervention, the time from HIV confirmation to ART initiation decreased from around 50 days to five days, and the proportion of individuals who initiated ART increased from below 36% to above 90% among all the patients and from below 47% to above 93% among patients eligible for treatment under the standard-of-care policy. Notably, the mortality rate decreased from about 26% to about 10% among all the study participants following the intervention, and from about 37% to about 13% among the participants eligible for ART under the standard-of-care policy. Finally, the researchers estimated that the cost of each death prevented by the intervention was about US$234.52 over the study period; importantly, most of this cost was accrued during the initial year of the intervention. What Do These Findings Mean? These findings indicate that, in the two Chinese counties involved in this study, the simplified test and treat intervention—which incorporated a streamlined, standardized time frame for HIV diagnosis and expanded access to ART—promoted successful engagement in care among HIV-positive individuals and was associated with a 62% reduction in mortality. Moreover, the intervention required very little further investment once it had been set up and should, therefore, be sustainable. Because the design of the simplified test and treat intervention took into account the characteristics of the HIV epidemic and the health care structure in China, these findings may not be fully generalizable to other countries. In addition, reliance on a pre-intervention/post-intervention study design, rather than a controlled trial, may limit the accuracy of these findings. Nevertheless, these results suggest that the implementation of integrated HIV testing and immediate access to ART regardless of CD4 count has the potential to optimize the individual and public health impacts of ART by ensuring that fewer patients are lost along the multistage continuum of HIV testing and treatment. Additional Information This list of resources contains links that can be accessed when viewing the PDF on a device or via the online version of the article at http://dx.doi.org/10.1371/journal.pmed.1001874. Information is available from the US National Institute of Allergy and Infectious Diseases on HIV infection and AIDS NAM/aidsmap provides basic information about HIV/AIDS, summaries of recent research findings on HIV care and treatment, and personal stories about living with HIV/AIDS Information is available from Avert, an international AIDS charity, on many aspects of HIV/AIDS, including information on universal access to ART and on HIV/AIDS in China; Avert also provides personal stories about living with HIV/AIDS The World Health Organization provides information on all aspects of HIV/AIDS (in several languages), including its Consolidated Guidelines on the Use of Antiretroviral Therapy for Treating and Preventing HIV Infection, its recently released consolidated guidelines on HIV testing, and information on the WHO/UNAIDS Treatment 2.0 strategy, an initiative to expand access to HIV testing and ART The UNAIDS Fast-Track Strategy to End the AIDS Epidemic by 2030 provides up-to-date information about the AIDS epidemic and efforts to halt it, including progress towards universal access to antiretroviral therapy Introduction In June 2010, the World Health Organization (WHO) and the Joint United Nations Programme on HIV/AIDS (UNAIDS) launched the Treatment 2.0 strategy, an initiative to expand access to HIV testing and antiretroviral therapy (ART) and to maximize the individual and public health benefits of modern HIV treatment [1]. Global experience over the past decade has confirmed the lifesaving benefits of ART for treating HIV-positive patients. However, late diagnosis, incomplete linkage to care, and loss to follow-up (LTFU) remain major clinical and public health challenges [2–4]. High rates of LTFU, before and after ART initiation, are relatively common in both high- and low-resource settings [5–9]. Moreover, late initiation of ART and high LTFU rates are significant drivers of mortality [10–13]. Despite nationwide scale-up of HIV programs in China over the last decade, the proportion of ART-eligible HIV-positive patients who receive treatment remains low. In 2009, it was estimated that 53.3% of HIV-positive individuals received a baseline CD4 count test within 6 mo of diagnosis [14], and among ART-eligible patients, treatment coverage was 63.4%, and the mortality rate was 14.2 per 100 person-years [12]. In China, as in many settings worldwide, patients are lost at each step along the continuum of HIV testing and care. This includes patients lost after not meeting ART eligibility criteria at the time of diagnosis. Other patients may meet treatment criteria but fail to initiate ART. Complicated HIV testing policies may be contributing to early LTFU. According to the current Chinese standard-of-care policies [15], CD4 testing is offered only after the HIV diagnosis has been confirmed through Western blot (WB) testing. This results in a structural delay to initiating ART because CD4 count is the primary measure used to determine eligibility for the Chinese National Free Antiretroviral Treatment Program (NFATP). Although published studies have described the HIV care continuum, most studies have been either observational or focused on an intervention targeting a limited section of the HIV care cascade, such as increasing the proportion of participants receiving CD4 testing or the proportion initiating ART [16,17]. Strategic interventions to streamline HIV testing and treatment procedures should be designed to decrease LTFU and mortality. We designed a structurally simplified test and treat intervention, to be completed within a week of the first positive HIV screening test result, incorporating immediate HIV confirmatory testing, pre-ART CD4 testing, pretreatment counseling, and ART initiation regardless of CD4 count. The aim of this pilot study was to evaluate the effectiveness of the simplified test and treat intervention in reducing delays to treatment and decreasing mortality. Methods Study Design We used a pre- and post-intervention study design to evaluate the ability of a simplified HIV test and treat intervention to reduce mortality among newly diagnosed HIV/AIDS cases. The original design was one pre-intervention and one post-intervention phase. The design was modified to have two pre-intervention and two post-intervention phases (Fig 1) based on suggestions from peer review. Data from the “pre-intervention 2010” phase, the period from 1 January 2010 to 31 December 2010, and the “pre-intervention 2011” phase, the period from 1 January 2011 to 31 December 2011, were analyzed as the control arm, in comparison to the “post-intervention 2012” phase, the period from 1 July 2012 to 30 June 2013, and the “post-intervention 2013” phase, the period from 1 July 2013 to 30 June 2014. The period from 1 January 2012 to 30 June 2012 was treated as the “intervention transition period.” Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. Study design of the simplified HIV test and treat intervention. Pre-intervention consisted of the standard of care; there were six steps from the enzyme immunoassay (EIA) screen to ART initiation, and the eligibility for ART was at CD4 count ≤ 350 cells/mm3. The simplified test and treat intervention comprised three steps from the enzyme immunoassay screen to ART initiation, regardless of CD4 level. In all four phases, participants were followed for a period of 6 to 18 mo, from the date of their WB confirmation results until 6 mo after the end of the recruitment phase in each study phase. https://doi.org/10.1371/journal.pmed.1001874.g001 Study Site Guangxi Zhuang Autonomous Region is one of the provinces in China most heavily affected by HIV/AIDS. In 2011, the overall mortality of HIV-positive individuals in Guangxi was 6.8% compared to a national average mortality of 5%. The mortality in Guangxi was in the medium-to-high range among the 31 provinces in China. However, Guangxi reported the highest absolute number of HIV-related deaths, accounting for 22% of the deaths in all 31 provinces in China. Late diagnosis (defined as diagnosis at CD4 count ≤ 200 cells/mm3 or no CD4 count but clinical AIDS at the time of diagnosis of HIV infection) accounted for approximately one-third of cases in Guangxi from 2007 to 2011. In 2011, nearly 70% of newly diagnosed cases had an initial CD4 count ≤ 350 cells/mm3, and among the cases who died in 2011, 48% had been diagnosed in the same calendar year [18]. About 79% of individuals who died of HIV-related causes had never received ART, indicating that linkage to care was suboptimal. Zhongshan County and Pubei County were selected as study sites because they had previously reported high proportions of deaths occurring within the same calendar year of HIV/AIDS diagnosis. In 2011, the respective cumulative numbers of HIV cases in Zhongshan and Pubei were 625 and 986, and the main route of transmission was heterosexual contact. The 2011 mortality was 33.9% and 31.1% of newly diagnosed HIV/AIDS cases in Zhongshan and Pubei, respectively. Study Participants Eligibility criteria were the following: (1) participants were newly diagnosed HIV-positive adults (≥ 18 y), (2) participants received a positive confirmation test at a study site (defined as a WB test that met the national laboratory standards [19]), and (3) participants resided within a study clinic’s catchment area. Participants were followed from the date of HIV confirmation to the end of the study phase, and the follow-up duration for study participants ranged from 6 to 18 mo. Standard-of-Care (Pre-Intervention) Procedures HIV screening in Guangxi is available through health care facilities at the township level and above (i.e., in ascending order, the levels township, county, city, provincial, and national). Screening is available through self-referral for testing and provider-initiated testing (which is routine in surgery departments, sexually transmitted infection clinics, and maternal care clinics). Under China’s national policies, patients receive at least two screening tests in succession (ELISA [enzyme-linked immunosorbent assay] or rapid test) and one WB test to confirm a diagnosis of HIV infection. In the two study counties, patients are screened for HIV using a rapid test at the initiating facility. If the initial screening test is positive, the patient is asked to return to the same facility to give a second blood sample, which is sent to the local county Center for Disease Control and Prevention (CDC) laboratory. This second screening test is analyzed using an ELISA test; if positive, the sample is sent to the city-level CDC laboratory for confirmatory WB testing. About 29%–36% of patients who screen HIV-positive at health care facilities fail to present for the second blood draw, marking the first drop-off in the care cascade. WB results are usually available within 7–18 d, and if the WB test is positive, the patient is asked to provide a third blood sample at the county CDC, which is transferred to the city CDC laboratory for CD4 testing. After the CD4 results are available (typically within 7–18 d), patients eligible for ART (CD4 count ≤ 350 cells/mm3) are asked to seek treatment at a separate facility designated to provide ART, which is usually based at the county general hospital [15]. Before treatment initiation, patients are expected to receive education and counseling on ART and adherence, and a physical exam. A fourth blood sample is collected for baseline pre-ART assessment, including kidney function, liver function, and other routine assessments. Patients who are not eligible for ART at diagnosis are advised to undergo CD4 testing every 6 mo for ongoing reassessment of ART eligibility. In many counties, the diagnostic and treatment initiation process requires patients to independently navigate multiple clinic visits, often at different facilities. For patients who are ART-eligible at diagnosis, the usual timeline from the initial screening to ART initiation is 2 to 4 mo. Treatment medication is provided free of charge through the NFATP. Data for the pre-intervention phase were collected through retrospective records review. Simplified Test and Treat Intervention Procedures The simplified test and treat intervention incorporated (1) a streamlined, standardized time frame for diagnosis and (2) expanded access to ART. These changes were enacted under the leadership of the National Center for AIDS/STD Control and Prevention (NCAIDS) at the Chinese Center for Disease Control and Prevention (China CDC), and included reorganization of services [20], with an emphasis on linkage and integration of services. Other aspects of the intervention included provider training and guidelines that accelerated treatment provision [21]. Standardized framework for diagnosis: organizational structure and integration of services. The simplified test and treat intervention framework designated activities to take place on a predetermined schedule. Following a positive HIV screening test, the patient received post-screening counseling (e.g., basic education on HIV/AIDS knowledge, the benefits of ART, and the importance of encouraging partners to receive HIV testing) on the same day at the same facility. At this time, the patient was scheduled for a visit at the HIV/AIDS clinic of the county general hospital on the following Wednesday. The HIV/AIDS epidemiologist from the county CDC and the chief physician of the county general hospital’s HIV/AIDS clinic were proactively notified regarding the newly diagnosed patients that were expected to visit the HIV/AIDS clinic the following Wednesday. Every Wednesday, an infectious disease physician and nurses from the HIV/AIDS clinic of the county general hospital carried out medical consultations for ART, blood sample collections for WB confirmation tests and CD4 cell count tests, pretreatment physical examinations, liver and kidney function tests, treatment of opportunistic infections and other co-morbid conditions, and other services as appropriate. The county CDC HIV/AIDS epidemiologist was based at the county general hospital each Wednesday to coordinate the delivery of blood samples from the county general hospital to the city CDC laboratory for same-day WB and CD4 cell count testing and to confirm that patients who screened HIV-positive within the past week presented for their scheduled visit. If a patient failed to arrive at the Wednesday HIV/AIDS clinic visit, the epidemiologist informed the referring clinician to follow-up with the patient and to reschedule the HIV/AIDS clinic visit for the following Wednesday. Every Friday, the city CDC delivered the HIV confirmation and CD4 cell count results from the previous Wednesday’s samples to the county CDC HIV/AIDS epidemiologist, who transferred the results to the county general hospital. The county general hospital was responsible for patient notification and all subsequent follow-up care. The county CDC HIV/AIDS epidemiologist prepared a monthly report on all newly diagnosed HIV cases, testing results, and treatment referral rates. Expanded access to ART: training and guidelines. ART eligibility was expanded to patients with a confirmed HIV diagnosis, irrespective of CD4 cell count, thereby superseding the previous primary criterion for ART initiation of CD4 count ≤ 350 cells/mm3. The simplified test and treat intervention explicitly included a minimum of three counseling sessions prior to treatment initiation. The first ART counseling session was provided by the staff of the originating facility immediately after screening. The first counseling session emphasized the importance of presenting to follow-up visits and ensured that the patients had the directions and contact information for the county general hospital designated for providing ART services. The second ART counseling session was provided by the county general hospital physician at the time of collecting blood for HIV confirmatory and CD4 cell count testing. The third ART counseling session was provided when patients received their HIV confirmatory and CD4 cell count results. During each counseling session, all patients were strongly encouraged to initiate ART promptly. Treatment medication was provided free of charge through the NFATP, which is consistent with the standard-of-care procedure. In order to implement the structural intervention described above, an initial planning meeting was held with key stakeholders, including research staff from NCAIDS/China CDC, officials from the Guangxi Health Department, technical experts from Guangxi CDC, health officials from the Pubei and Zhongshan county health departments, hospital directors and HIV/AIDS clinicians from the two county general hospitals, county CDC directors, and county CDC HIV/AIDS epidemiologists. A consensus was reached on the structural intervention components and the corresponding implementation plan for the two study sites. Policy papers on implementing the simplified test and treat intervention were issued by the two local county health departments and distributed to all health facilities that provide HIV screening in the two counties. NCAIDS carried out training workshops for providers to review the new HIV care guidelines, practice mock exercises, strengthen communication skills, and reinforce professional expectations. Clinicians who conduct HIV screening tests were also intensively trained on providing post-screening counseling. A handbook was issued to all providers on HIV testing, treatment, and prevention, and educational materials were prepared for distribution to newly diagnosed HIV-positive patients. To monitor the intervention, site supervisors from NCAIDS were stationed in the two study sites for the first 2 mo of the intervention. Afterwards, site visits were conducted monthly to monitor adherence to the intervention protocol. The total cost for the study intervention was approximately US$14,525.75. Data Management The local county CDC HIV/AIDS epidemiologists are legally responsible for following up on all HIV-positive patients to record demographic information, eligibility for ART, and present status (i.e., in regular HIV care, lost to follow-up, migrated out of the county, or deceased) every 6 mo. Physicians providing ART are required to collect ART-related information, including medication side effects and ART status (i.e., engaged on ART, dropped out of ART, lost to follow-up, migrated out of the county, or deceased) every 3 mo and success or failure of ART based on viral load testing once a year. In the pre-intervention phase, per the standard of care, the CDC HIV/AIDS epidemiologists collected data independently of the ART-providing clinicians. In the intervention phase, the HIV/AIDS epidemiologists collaborated with clinicians to record patient data using standardized case report forms. Data were subsequently entered into the national HIV/AIDS case reporting and NFATP databases. These two databases are subsystems of China’s Comprehensive Response Information Management System (CRIMS), a national web-based real-time data system for HIV care and follow-up, which has been previously described [22]. An HIV/AIDS dataset for Pubei and Zhongshan covering 1 January 2010 to 31 December 2014 was downloaded from CRIMS. The study participants in the pre-intervention 2010 phase comprised patients newly diagnosed as HIV-positive between 1 January 2010 and 31 December 2010 and were followed until 30 June 2011. The study participants in the pre-intervention 2011 phase comprised patients newly diagnosed as HIV-positive between 1 January 2011 and 31 December 2011 and were followed until 30 June 2012. The study participants in the post-intervention 2012 phase were patients newly diagnosed as HIV-positive between 1 July 2012 and 30 June 2013 and were followed until 31 December 2013. The study participants in the post-intervention 2013 phase were patients newly diagnosed as HIV-positive between 1 July 2013 and 30 June 2014 and were followed until 31 December 2014. In all four phases, each participant was followed for a duration between 6 and 18 mo. We used data from the 12 mo preceding the pre-intervention 2011 phase and from the 12 mo after the post-intervention 2012 phase to assess potential secular time trends independent of the intervention, based on suggestions from peer review. Statistical Analysis Survival times were calculated as the time from HIV confirmation until death or the last follow-up (at which point survival times were censored). The number of days (median, interquartile range [IQR]) from HIV-positive screening to HIV confirmatory testing, the total number of deaths, and mortality rates were tabulated, stratified by study phase. The overall proportions receiving ART and receiving ART within 30 d of HIV confirmation in each phase were compared by chi-square tests. Kaplan-Meier curves describing survival and ART initiation were compared by log-rank tests. Cox proportional hazards multivariate regressions were used to calculate hazard ratios (HRs) for the effect of the study phase on mortality after adjustment for baseline demographic risk factors. We estimated adjusted hazard ratios (aHRs) and 95% confidence intervals (CIs) from the Cox regressions, using the pre-intervention 2011 phase as a reference group. In addition, we performed an analysis where we further adjusted for variables that could have been modified by the intervention (baseline CD4 or clinical status, and ART initiation). ART initiation was treated as a time-dependent covariate, which took the value zero until the time that a patient first received ART. The proportional hazards assumption for all covariates was assessed by testing the Martingale residuals, and we found no evidence that this assumption was violated. Because the treatment eligibility criteria changed between the pre-intervention and post-intervention study phases, we conducted additional analyses among the subset of individuals with CD4 count ≤ 350 cells/mm3 or with clinical AIDS so that the treatment-related outcomes were comparable between the two pre-intervention and two post-intervention phases. Statistical analyses were performed using SAS (version 9.1.3, SAS Institute). We estimated the actual cost of implementing the intervention, including the cost of the training of local health workers, incentives for health workers for timely referral and initiation of patients on ART, monitoring of the implementation, and educational materials. We excluded any costs that were specific to research. The number of additional patients receiving ART because of the intervention was estimated for the initial year of the intervention (post-intervention 2012 phase) and for the second year of the intervention (post-intervention 2013 phase), based on differences in the ART coverage rate compared to that in the pre-intervention 2011 phase. The number of deaths prevented because of the intervention was estimated for the post-intervention 2012 phase and for the post-intervention 2013 phase, based on the differences in overall mortality compared to that in the pre-intervention 2011 phase. We estimated the cost both as a total for the program and as an incremental cost per incremental patient achieving the primary outcome. Among the above, the approach to stratifying the proportional hazard assumptions, assessing potential secular trends, and conducting the cost analysis emerged from the review process. Ethics The study protocol was reviewed and approved by the Institutional Review Board of NCAIDS, China CDC (#X120717220). All patients who screened HIV-positive were requested to give their written informed consent for the use of their de-personalized data in future epidemiological analysis. No additional study-specific written informed consent was obtained. None of the participants opted out. There was no incentive for study participants. Study Design We used a pre- and post-intervention study design to evaluate the ability of a simplified HIV test and treat intervention to reduce mortality among newly diagnosed HIV/AIDS cases. The original design was one pre-intervention and one post-intervention phase. The design was modified to have two pre-intervention and two post-intervention phases (Fig 1) based on suggestions from peer review. Data from the “pre-intervention 2010” phase, the period from 1 January 2010 to 31 December 2010, and the “pre-intervention 2011” phase, the period from 1 January 2011 to 31 December 2011, were analyzed as the control arm, in comparison to the “post-intervention 2012” phase, the period from 1 July 2012 to 30 June 2013, and the “post-intervention 2013” phase, the period from 1 July 2013 to 30 June 2014. The period from 1 January 2012 to 30 June 2012 was treated as the “intervention transition period.” Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. Study design of the simplified HIV test and treat intervention. Pre-intervention consisted of the standard of care; there were six steps from the enzyme immunoassay (EIA) screen to ART initiation, and the eligibility for ART was at CD4 count ≤ 350 cells/mm3. The simplified test and treat intervention comprised three steps from the enzyme immunoassay screen to ART initiation, regardless of CD4 level. In all four phases, participants were followed for a period of 6 to 18 mo, from the date of their WB confirmation results until 6 mo after the end of the recruitment phase in each study phase. https://doi.org/10.1371/journal.pmed.1001874.g001 Study Site Guangxi Zhuang Autonomous Region is one of the provinces in China most heavily affected by HIV/AIDS. In 2011, the overall mortality of HIV-positive individuals in Guangxi was 6.8% compared to a national average mortality of 5%. The mortality in Guangxi was in the medium-to-high range among the 31 provinces in China. However, Guangxi reported the highest absolute number of HIV-related deaths, accounting for 22% of the deaths in all 31 provinces in China. Late diagnosis (defined as diagnosis at CD4 count ≤ 200 cells/mm3 or no CD4 count but clinical AIDS at the time of diagnosis of HIV infection) accounted for approximately one-third of cases in Guangxi from 2007 to 2011. In 2011, nearly 70% of newly diagnosed cases had an initial CD4 count ≤ 350 cells/mm3, and among the cases who died in 2011, 48% had been diagnosed in the same calendar year [18]. About 79% of individuals who died of HIV-related causes had never received ART, indicating that linkage to care was suboptimal. Zhongshan County and Pubei County were selected as study sites because they had previously reported high proportions of deaths occurring within the same calendar year of HIV/AIDS diagnosis. In 2011, the respective cumulative numbers of HIV cases in Zhongshan and Pubei were 625 and 986, and the main route of transmission was heterosexual contact. The 2011 mortality was 33.9% and 31.1% of newly diagnosed HIV/AIDS cases in Zhongshan and Pubei, respectively. Study Participants Eligibility criteria were the following: (1) participants were newly diagnosed HIV-positive adults (≥ 18 y), (2) participants received a positive confirmation test at a study site (defined as a WB test that met the national laboratory standards [19]), and (3) participants resided within a study clinic’s catchment area. Participants were followed from the date of HIV confirmation to the end of the study phase, and the follow-up duration for study participants ranged from 6 to 18 mo. Standard-of-Care (Pre-Intervention) Procedures HIV screening in Guangxi is available through health care facilities at the township level and above (i.e., in ascending order, the levels township, county, city, provincial, and national). Screening is available through self-referral for testing and provider-initiated testing (which is routine in surgery departments, sexually transmitted infection clinics, and maternal care clinics). Under China’s national policies, patients receive at least two screening tests in succession (ELISA [enzyme-linked immunosorbent assay] or rapid test) and one WB test to confirm a diagnosis of HIV infection. In the two study counties, patients are screened for HIV using a rapid test at the initiating facility. If the initial screening test is positive, the patient is asked to return to the same facility to give a second blood sample, which is sent to the local county Center for Disease Control and Prevention (CDC) laboratory. This second screening test is analyzed using an ELISA test; if positive, the sample is sent to the city-level CDC laboratory for confirmatory WB testing. About 29%–36% of patients who screen HIV-positive at health care facilities fail to present for the second blood draw, marking the first drop-off in the care cascade. WB results are usually available within 7–18 d, and if the WB test is positive, the patient is asked to provide a third blood sample at the county CDC, which is transferred to the city CDC laboratory for CD4 testing. After the CD4 results are available (typically within 7–18 d), patients eligible for ART (CD4 count ≤ 350 cells/mm3) are asked to seek treatment at a separate facility designated to provide ART, which is usually based at the county general hospital [15]. Before treatment initiation, patients are expected to receive education and counseling on ART and adherence, and a physical exam. A fourth blood sample is collected for baseline pre-ART assessment, including kidney function, liver function, and other routine assessments. Patients who are not eligible for ART at diagnosis are advised to undergo CD4 testing every 6 mo for ongoing reassessment of ART eligibility. In many counties, the diagnostic and treatment initiation process requires patients to independently navigate multiple clinic visits, often at different facilities. For patients who are ART-eligible at diagnosis, the usual timeline from the initial screening to ART initiation is 2 to 4 mo. Treatment medication is provided free of charge through the NFATP. Data for the pre-intervention phase were collected through retrospective records review. Simplified Test and Treat Intervention Procedures The simplified test and treat intervention incorporated (1) a streamlined, standardized time frame for diagnosis and (2) expanded access to ART. These changes were enacted under the leadership of the National Center for AIDS/STD Control and Prevention (NCAIDS) at the Chinese Center for Disease Control and Prevention (China CDC), and included reorganization of services [20], with an emphasis on linkage and integration of services. Other aspects of the intervention included provider training and guidelines that accelerated treatment provision [21]. Standardized framework for diagnosis: organizational structure and integration of services. The simplified test and treat intervention framework designated activities to take place on a predetermined schedule. Following a positive HIV screening test, the patient received post-screening counseling (e.g., basic education on HIV/AIDS knowledge, the benefits of ART, and the importance of encouraging partners to receive HIV testing) on the same day at the same facility. At this time, the patient was scheduled for a visit at the HIV/AIDS clinic of the county general hospital on the following Wednesday. The HIV/AIDS epidemiologist from the county CDC and the chief physician of the county general hospital’s HIV/AIDS clinic were proactively notified regarding the newly diagnosed patients that were expected to visit the HIV/AIDS clinic the following Wednesday. Every Wednesday, an infectious disease physician and nurses from the HIV/AIDS clinic of the county general hospital carried out medical consultations for ART, blood sample collections for WB confirmation tests and CD4 cell count tests, pretreatment physical examinations, liver and kidney function tests, treatment of opportunistic infections and other co-morbid conditions, and other services as appropriate. The county CDC HIV/AIDS epidemiologist was based at the county general hospital each Wednesday to coordinate the delivery of blood samples from the county general hospital to the city CDC laboratory for same-day WB and CD4 cell count testing and to confirm that patients who screened HIV-positive within the past week presented for their scheduled visit. If a patient failed to arrive at the Wednesday HIV/AIDS clinic visit, the epidemiologist informed the referring clinician to follow-up with the patient and to reschedule the HIV/AIDS clinic visit for the following Wednesday. Every Friday, the city CDC delivered the HIV confirmation and CD4 cell count results from the previous Wednesday’s samples to the county CDC HIV/AIDS epidemiologist, who transferred the results to the county general hospital. The county general hospital was responsible for patient notification and all subsequent follow-up care. The county CDC HIV/AIDS epidemiologist prepared a monthly report on all newly diagnosed HIV cases, testing results, and treatment referral rates. Expanded access to ART: training and guidelines. ART eligibility was expanded to patients with a confirmed HIV diagnosis, irrespective of CD4 cell count, thereby superseding the previous primary criterion for ART initiation of CD4 count ≤ 350 cells/mm3. The simplified test and treat intervention explicitly included a minimum of three counseling sessions prior to treatment initiation. The first ART counseling session was provided by the staff of the originating facility immediately after screening. The first counseling session emphasized the importance of presenting to follow-up visits and ensured that the patients had the directions and contact information for the county general hospital designated for providing ART services. The second ART counseling session was provided by the county general hospital physician at the time of collecting blood for HIV confirmatory and CD4 cell count testing. The third ART counseling session was provided when patients received their HIV confirmatory and CD4 cell count results. During each counseling session, all patients were strongly encouraged to initiate ART promptly. Treatment medication was provided free of charge through the NFATP, which is consistent with the standard-of-care procedure. In order to implement the structural intervention described above, an initial planning meeting was held with key stakeholders, including research staff from NCAIDS/China CDC, officials from the Guangxi Health Department, technical experts from Guangxi CDC, health officials from the Pubei and Zhongshan county health departments, hospital directors and HIV/AIDS clinicians from the two county general hospitals, county CDC directors, and county CDC HIV/AIDS epidemiologists. A consensus was reached on the structural intervention components and the corresponding implementation plan for the two study sites. Policy papers on implementing the simplified test and treat intervention were issued by the two local county health departments and distributed to all health facilities that provide HIV screening in the two counties. NCAIDS carried out training workshops for providers to review the new HIV care guidelines, practice mock exercises, strengthen communication skills, and reinforce professional expectations. Clinicians who conduct HIV screening tests were also intensively trained on providing post-screening counseling. A handbook was issued to all providers on HIV testing, treatment, and prevention, and educational materials were prepared for distribution to newly diagnosed HIV-positive patients. To monitor the intervention, site supervisors from NCAIDS were stationed in the two study sites for the first 2 mo of the intervention. Afterwards, site visits were conducted monthly to monitor adherence to the intervention protocol. The total cost for the study intervention was approximately US$14,525.75. Standardized framework for diagnosis: organizational structure and integration of services. The simplified test and treat intervention framework designated activities to take place on a predetermined schedule. Following a positive HIV screening test, the patient received post-screening counseling (e.g., basic education on HIV/AIDS knowledge, the benefits of ART, and the importance of encouraging partners to receive HIV testing) on the same day at the same facility. At this time, the patient was scheduled for a visit at the HIV/AIDS clinic of the county general hospital on the following Wednesday. The HIV/AIDS epidemiologist from the county CDC and the chief physician of the county general hospital’s HIV/AIDS clinic were proactively notified regarding the newly diagnosed patients that were expected to visit the HIV/AIDS clinic the following Wednesday. Every Wednesday, an infectious disease physician and nurses from the HIV/AIDS clinic of the county general hospital carried out medical consultations for ART, blood sample collections for WB confirmation tests and CD4 cell count tests, pretreatment physical examinations, liver and kidney function tests, treatment of opportunistic infections and other co-morbid conditions, and other services as appropriate. The county CDC HIV/AIDS epidemiologist was based at the county general hospital each Wednesday to coordinate the delivery of blood samples from the county general hospital to the city CDC laboratory for same-day WB and CD4 cell count testing and to confirm that patients who screened HIV-positive within the past week presented for their scheduled visit. If a patient failed to arrive at the Wednesday HIV/AIDS clinic visit, the epidemiologist informed the referring clinician to follow-up with the patient and to reschedule the HIV/AIDS clinic visit for the following Wednesday. Every Friday, the city CDC delivered the HIV confirmation and CD4 cell count results from the previous Wednesday’s samples to the county CDC HIV/AIDS epidemiologist, who transferred the results to the county general hospital. The county general hospital was responsible for patient notification and all subsequent follow-up care. The county CDC HIV/AIDS epidemiologist prepared a monthly report on all newly diagnosed HIV cases, testing results, and treatment referral rates. Expanded access to ART: training and guidelines. ART eligibility was expanded to patients with a confirmed HIV diagnosis, irrespective of CD4 cell count, thereby superseding the previous primary criterion for ART initiation of CD4 count ≤ 350 cells/mm3. The simplified test and treat intervention explicitly included a minimum of three counseling sessions prior to treatment initiation. The first ART counseling session was provided by the staff of the originating facility immediately after screening. The first counseling session emphasized the importance of presenting to follow-up visits and ensured that the patients had the directions and contact information for the county general hospital designated for providing ART services. The second ART counseling session was provided by the county general hospital physician at the time of collecting blood for HIV confirmatory and CD4 cell count testing. The third ART counseling session was provided when patients received their HIV confirmatory and CD4 cell count results. During each counseling session, all patients were strongly encouraged to initiate ART promptly. Treatment medication was provided free of charge through the NFATP, which is consistent with the standard-of-care procedure. In order to implement the structural intervention described above, an initial planning meeting was held with key stakeholders, including research staff from NCAIDS/China CDC, officials from the Guangxi Health Department, technical experts from Guangxi CDC, health officials from the Pubei and Zhongshan county health departments, hospital directors and HIV/AIDS clinicians from the two county general hospitals, county CDC directors, and county CDC HIV/AIDS epidemiologists. A consensus was reached on the structural intervention components and the corresponding implementation plan for the two study sites. Policy papers on implementing the simplified test and treat intervention were issued by the two local county health departments and distributed to all health facilities that provide HIV screening in the two counties. NCAIDS carried out training workshops for providers to review the new HIV care guidelines, practice mock exercises, strengthen communication skills, and reinforce professional expectations. Clinicians who conduct HIV screening tests were also intensively trained on providing post-screening counseling. A handbook was issued to all providers on HIV testing, treatment, and prevention, and educational materials were prepared for distribution to newly diagnosed HIV-positive patients. To monitor the intervention, site supervisors from NCAIDS were stationed in the two study sites for the first 2 mo of the intervention. Afterwards, site visits were conducted monthly to monitor adherence to the intervention protocol. The total cost for the study intervention was approximately US$14,525.75. Data Management The local county CDC HIV/AIDS epidemiologists are legally responsible for following up on all HIV-positive patients to record demographic information, eligibility for ART, and present status (i.e., in regular HIV care, lost to follow-up, migrated out of the county, or deceased) every 6 mo. Physicians providing ART are required to collect ART-related information, including medication side effects and ART status (i.e., engaged on ART, dropped out of ART, lost to follow-up, migrated out of the county, or deceased) every 3 mo and success or failure of ART based on viral load testing once a year. In the pre-intervention phase, per the standard of care, the CDC HIV/AIDS epidemiologists collected data independently of the ART-providing clinicians. In the intervention phase, the HIV/AIDS epidemiologists collaborated with clinicians to record patient data using standardized case report forms. Data were subsequently entered into the national HIV/AIDS case reporting and NFATP databases. These two databases are subsystems of China’s Comprehensive Response Information Management System (CRIMS), a national web-based real-time data system for HIV care and follow-up, which has been previously described [22]. An HIV/AIDS dataset for Pubei and Zhongshan covering 1 January 2010 to 31 December 2014 was downloaded from CRIMS. The study participants in the pre-intervention 2010 phase comprised patients newly diagnosed as HIV-positive between 1 January 2010 and 31 December 2010 and were followed until 30 June 2011. The study participants in the pre-intervention 2011 phase comprised patients newly diagnosed as HIV-positive between 1 January 2011 and 31 December 2011 and were followed until 30 June 2012. The study participants in the post-intervention 2012 phase were patients newly diagnosed as HIV-positive between 1 July 2012 and 30 June 2013 and were followed until 31 December 2013. The study participants in the post-intervention 2013 phase were patients newly diagnosed as HIV-positive between 1 July 2013 and 30 June 2014 and were followed until 31 December 2014. In all four phases, each participant was followed for a duration between 6 and 18 mo. We used data from the 12 mo preceding the pre-intervention 2011 phase and from the 12 mo after the post-intervention 2012 phase to assess potential secular time trends independent of the intervention, based on suggestions from peer review. Statistical Analysis Survival times were calculated as the time from HIV confirmation until death or the last follow-up (at which point survival times were censored). The number of days (median, interquartile range [IQR]) from HIV-positive screening to HIV confirmatory testing, the total number of deaths, and mortality rates were tabulated, stratified by study phase. The overall proportions receiving ART and receiving ART within 30 d of HIV confirmation in each phase were compared by chi-square tests. Kaplan-Meier curves describing survival and ART initiation were compared by log-rank tests. Cox proportional hazards multivariate regressions were used to calculate hazard ratios (HRs) for the effect of the study phase on mortality after adjustment for baseline demographic risk factors. We estimated adjusted hazard ratios (aHRs) and 95% confidence intervals (CIs) from the Cox regressions, using the pre-intervention 2011 phase as a reference group. In addition, we performed an analysis where we further adjusted for variables that could have been modified by the intervention (baseline CD4 or clinical status, and ART initiation). ART initiation was treated as a time-dependent covariate, which took the value zero until the time that a patient first received ART. The proportional hazards assumption for all covariates was assessed by testing the Martingale residuals, and we found no evidence that this assumption was violated. Because the treatment eligibility criteria changed between the pre-intervention and post-intervention study phases, we conducted additional analyses among the subset of individuals with CD4 count ≤ 350 cells/mm3 or with clinical AIDS so that the treatment-related outcomes were comparable between the two pre-intervention and two post-intervention phases. Statistical analyses were performed using SAS (version 9.1.3, SAS Institute). We estimated the actual cost of implementing the intervention, including the cost of the training of local health workers, incentives for health workers for timely referral and initiation of patients on ART, monitoring of the implementation, and educational materials. We excluded any costs that were specific to research. The number of additional patients receiving ART because of the intervention was estimated for the initial year of the intervention (post-intervention 2012 phase) and for the second year of the intervention (post-intervention 2013 phase), based on differences in the ART coverage rate compared to that in the pre-intervention 2011 phase. The number of deaths prevented because of the intervention was estimated for the post-intervention 2012 phase and for the post-intervention 2013 phase, based on the differences in overall mortality compared to that in the pre-intervention 2011 phase. We estimated the cost both as a total for the program and as an incremental cost per incremental patient achieving the primary outcome. Among the above, the approach to stratifying the proportional hazard assumptions, assessing potential secular trends, and conducting the cost analysis emerged from the review process. Ethics The study protocol was reviewed and approved by the Institutional Review Board of NCAIDS, China CDC (#X120717220). All patients who screened HIV-positive were requested to give their written informed consent for the use of their de-personalized data in future epidemiological analysis. No additional study-specific written informed consent was obtained. None of the participants opted out. There was no incentive for study participants. Results As shown in Fig 2, a total of 281 and 339 newly diagnosed HIV-positive patients were included in the pre-intervention 2010 and pre-intervention 2011 phases, respectively, and 215 and 199 patients in the post-intervention 2012 and post-intervention 2013 phases, respectively. Among the 281 patients in the pre-intervention 2010 phase, 76 (27.0%) enrolled in ART, and 75 (26.7%) died, including two deaths following ART initiation. Among the 339 participants in the pre-intervention 2011 phase, 165 (48.7%) enrolled in ART, and 90 (26.5%) died, including 15 deaths following ART initiation. During the post-intervention 2012 phase, 196 out of 215 (91.2%) individuals enrolled in ART, and 21 (9.8%) individuals died, including 12 deaths following ART initiation. During the post-intervention 2013 phase, 177 out of 199 (89%) patients enrolled in ART, and 20 (10.0%) patients died, including 13 deaths following ART initiation. Median participant follow-up was 8.77 mo (IQR 6.0–12.2) for the post-intervention 2013 phase and 9.2 mo (IQR 5.9–12.2) for the post-intervention 2012 phase, compared to 7.9 mo (IQR 5.6–11.2) for the pre-intervention 2011 phase and 7.33 mo (IQR 4.1–11.1) for the pre-intervention 2010 phase. In the intervention transition period (not included in Fig 2), 111 patients were diagnosed with HIV infection, 77 (69.4%) enrolled in ART, and 22 died (19.8%), including seven deaths following ART initiation. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. Number of HIV diagnoses, ART initiations, and deaths in the pre-intervention 2010, pre-intervention 2011, post-intervention 2012, and post-intervention 2013 phases in Guangxi, China. https://doi.org/10.1371/journal.pmed.1001874.g002 Of the total 1,034 participants (Table 1), 746 (72%) were male, and the median age was 48 y (IQR 36–60). Nearly all participants had a middle school education or below (95.6%) and were employed as farmers (90.6%). Over half of the individuals were married or partnered (56.5%). The participants who were enrolled during the four phases were not statistically significantly different on any baseline demographic characteristics. The predominant mode of HIV transmission was heterosexual contact, and this remained consistent throughout the study. Among 111 patients diagnosed in the intervention transition period (not included in Table 1), 87 (78.4%) were male, median age was 51 y (IQR 38–60), and 109 cases were infected by heterosexual transmission, and two cases via injecting drug use. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 1. Characteristics of newly diagnosed HIV cases in the pre-intervention 2010, pre-intervention 2011, post-intervention 2012, and post-intervention 2013 phases in Guangxi, China. https://doi.org/10.1371/journal.pmed.1001874.t001 In the pre-intervention 2010 and pre-intervention 2011 phases, 67% (median CD4 count = 243 cells/mm3, IQR 75–384) and 60.5% (median CD4 count = 219 cells/mm3, IQR 77–403) of participants had a baseline CD4 cell count within 30 d of HIV confirmation, respectively. This value was statistically significantly higher in the post-intervention 2012 phase (97.7%, median CD4 count = 220 cells/mm3, IQR 69–379, all p < 0.001[i.e., for all comparisons between a pre- and post-intervention phase]) and in the post-intervention 2013 phase (97%, median CD4 count = 178 cells/mm3, IQR 53–330, all p < 0.001). In the pre-intervention 2010 and pre-intervention 2011 phases, 34% and 16% of patients failed to obtain CD4 testing at any time point, compared to 2% and 0% in the post-intervention 2012 and post-intervention 2013 phases (all p < 0.001), respectively. The median time from HIV confirmatory testing to CD4 testing was 28 d (IQR 11–145) in the pre-intervention 2010 phase and 14 d (IQR 6–42) in the pre-intervention 2011 phase, compared to 1 d (IQR 0–4) in the post-intervention 2012 phase (all p < 0.001) and 0 d (IQR 0–1) in the post-intervention 2013 phase (all p < 0.001). In the intervention transition period (not included in Table 1), 93 (83.8%, median CD4 count = 264 cells/mm3, IQR 75–433) had a baseline CD4 cell count within 30 d of HIV confirmation, and 11 (10%) of patients failed to obtain CD4 testing at any time point. The median time from HIV confirmatory testing to CD4 testing was 6 d (IQR 4–13). In the pre-intervention 2010 and pre-intervention 2011 phases, 27% and 48.7% of the total participants initiated ART, respectively, compared to 91.2% and 89% in the post-intervention 2012 and post-intervention 2013 phases, respectively (all p < 0.001). Among individuals with CD4 count ≤ 350 cells/mm3 or missing their CD4 count but reported as diagnosed with AIDS, 39.4% in the pre-intervention 2010 phase and 61.6% in the pre-intervention 2011 phase initiated ART, compared to 93.5% in the post-intervention 2012 phase and 89.9% in post-intervention 2013 phase (all p < 0.001). For patients who initiated ART, the median time from HIV screening to confirmatory testing was 11 d (IQR 6–22) in the pre-intervention 2010 phase and 11 d (IQR 7–18) in the pre-intervention 2011 phase, compared to 7 d (IQR 3–10) in the post-intervention 2012 phase and 6 d (IQR 3–9) in the post-intervention 2013 phase (all p < 0.001). The median time from HIV confirmatory testing to treatment initiation was 53 d (IQR 27–141) in the pre-intervention 2010 phase and 43 d (IQR 15–113) in the pre-intervention 2011 phase, compared to 5 d (IQR 2–12) in the post-intervention 2012 phase and 5 d (IQR 2–13) in the post-intervention 2013 phase (all p < 0.001). In the intervention transition period (not included in Table 1), 77 cases (69.4%) initiated ART, and the median time from HIV confirmatory testing to treatment initiation was 25 d (IQR 7–97). A total of 228 deaths occurred over the entire study observation period, with a crude mortality of 26.7% in the pre-intervention 2010 phase and 26.5% in the pre-intervention 2011 phase, compared to 9.8% in the post-intervention 2012 phase and 10.1% in the post-intervention 2013 phase (all p < 0.001) and 19.8% in the intervention transition period (not included in Table 1). Among individuals with CD4 count ≤ 350 cells/mm3 or missing CD4 count but diagnosed with AIDS, 198 deaths occurred, with mortality of 39.9% in the pre-intervention 2010 phase and 35.0% in the pre-intervention 2011 phase, compared to 13.0% in the post-intervention 2012 phase and 12.7% in the post-intervention 2013 phase (all p < 0.001). Reassuringly, the primary outcome variable of overall mortality was quite consistent within the two pre-intervention phases, at 26.7% and 26.5%, respectively. Mortality was 9.8% and 10.1% in the two post-intervention phases, respectively. We observed 22 deaths among 113 HIV cases newly diagnosed during the 6 mo of the intervention transition period, with a crude mortality of 19.5% (22/113), which falls in between the values of the two pre-intervention phases and the two post-intervention phases. We also calculated the mortality rates for the two pre-intervention phases, the intervention transition period, and the two post-intervention phases, and there were 3.5, 3.4, 2.5, 1.2, and 1.2 deaths/100 person-years, respectively. The HIV care cascades by study phase are shown in Fig 3. Our study focused on the period between HIV-positive confirmation and death. Among all patients in the pre-intervention 2010 and pre-intervention 2011 phases, 33.5% and 15.3% of patients were lost to follow-up between HIV diagnosis and CD4 testing, respectively. A further 39.5% and 36.0% of patients were lost before ART initiation, respectively, and 26.7% and 26.5% of patients died, respectively. In contrast, in the post-intervention 2012 and post-intervention 2013 phases, only 1.9% and 0.0% of patients were lost to follow-up between HIV diagnosis and CD4 testing, respectively; an additional 7.0% and 11.1% of patients were lost before ART initiation, respectively, and 9.8% and 10.1% of patients died, respectively. Among individuals with CD4 count ≤ 350 cells/mm3 or missing CD4 but diagnosed with AIDS, 30.9% of patients in the pre-intervention 2010 phase and 18.1% of patients in the pre-intervention 2011 phase were lost to follow-up between HIV diagnosis and CD4 testing, a further 29.8% and 20.3% were lost before ART initiation, respectively, and 39.9% and 35.0% died, respectively. Among comparable patients in the post-intervention 2012 and post-intervention 2013 phases, 1.3% and 0.0% were lost to follow-up between HIV diagnosis and CD4 testing, respectively, a further 5.2% and 10.1% were lost before ART initiation, respectively, and 13.0% and 12.7% died, respectively. In the intervention transition period (not included in Fig 3), among all newly diagnosed cases, 9.9% of patients were lost to follow-up between HIV diagnosis and CD4 testing, a further 20.7% were lost before ART initiation, and 19.8% died. Among individuals with CD4 count ≤ 350 cells/mm3 or missing CD4 but diagnosed with AIDS, 12.9% of patients were lost to follow-up between HIV diagnosis and CD4 testing, a further 11.4% were lost before ART initiation, and 31.4% died. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. Cascade of confirmed HIV diagnosis, CD4 testing, ART initiation, and mortality during the pre-intervention 2010, pre-intervention 2011, post-intervention 2012, and post-intervention 2013 phases in Guangxi, China. (A) For all HIV cases. (B) For individuals with CD4 count ≤ 350 cells/mm3 or missing CD4 but reported as AIDS cases. https://doi.org/10.1371/journal.pmed.1001874.g003 Patients in the two post-intervention phases had significantly higher survival rates over the follow-up period than those in the two pre-intervention phases (all p < 0.001; Fig 4). As shown in Fig 5, within 90 d of HIV confirmation, ART coverage in the post-intervention 2012 and post-intervention 2013 phases was similar and was 2.5 times the coverage in the pre-intervention 2011 phase and 4.8 times the coverage in the pre-intervention 2010 phase for all newly diagnosed patients, and 2.0 times and 3.3 times, respectively, for patients with CD4 count ≤ 350 cells/mm3 or missing CD4 but reported as AIDS cases. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 4. Kaplan-Meier survival curves for newly diagnosed HIV cases in the pre-intervention 2010, pre-intervention 2011, post-intervention 2012, and post-intervention 2013 phases in Guangxi, China. (A) For all HIV cases. (B) For individuals with CD4 count ≤ 350 cells/mm3 or missing CD4 but reported as AIDS cases. https://doi.org/10.1371/journal.pmed.1001874.g004 Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 5. Kaplan-Meier curves for ART initiation for newly diagnosed HIV cases in the pre-intervention 2010, pre-intervention 2011, post-intervention 2012, and post-intervention 2013 phases in Guangxi, China. (A) For all HIV cases. (B) For individuals with CD4 count ≤ 350 cells/mm3 or missing CD4 but reported as AIDS cases. https://doi.org/10.1371/journal.pmed.1001874.g005 Table 2 describes factors associated with mortality for study participants pooled from the two pre-intervention and the two post-intervention phases. In univariate analysis, factors associated with death were being male (HR 2.604, 95% CI 1.758–3.856, p < 0.001), every 10-y increase in age (HR 1.258, 95% CI 1.155–1.371), CD4 count ≤ 200 cells/mm3 or AIDS (HR 15.344, 95% CI 8.918–26.402, p < 0.001), and receiving ART (HR 0.231, 95% CI 0.162–0.328, p < 0.001). The simplified test and treat intervention was statistically strongly associated with a decreased risk for mortality (HR 0.361, 95% CI 0.224–0.580, p < 0.001, for post-intervention 2012; HR 0.367, 95% CI 0.226–0.603, p < 0.001, for post-intervention 2013). Download: PPT PowerPoint slide PNG larger image TIFF original image Table 2. Mortality among all newly diagnosed HIV cases during the pre-intervention 2010, pre-intervention 2011, post-intervention 2012, and post-intervention 2013 phases, based on Cox model analysis. https://doi.org/10.1371/journal.pmed.1001874.t002 In the multivariate proportional hazards model, after adjusting for age and gender, the simplified test and treat intervention was significantly associated with decreased mortality compared to pre-intervention 2011 (aHR 0.385, 95% CI 0.239–0.620, p < 0.001, for post-intervention 2012; aHR 0.380, 95% CI 0.233–0.618, p < 0.001, for post-intervention 2013). When the variables that were potentially modified by the intervention were also included in the regression model (i.e., CD4 testing, CD4 cell counts, and ART initiation), we found that the intervention was no longer significantly associated with mortality (aHR 1.034, 95% CI 0.611–1.752, p = 0.900, for post-intervention 2012; aHR 0.882, 95% CI 0.517–1.504, p = 0.644, for post-intervention 2013). This result suggests that the impact of the intervention on mortality was meditated through changes in these other variables. We repeated the analyses for the subset of individuals with CD4 count ≤ 350 cells/mm3 or clinical AIDS (Table 3). Under the multivariable model that controlled for gender and age, the simplified test and treat intervention was significantly associated with reduced mortality compared to pre-intervention 2011 (aHR 0.369, 95% CI 0.226–0.603, p < 0.001, for post-intervention 2012; aHR = 0.361, 95% CI 0.221–0.590, p < 0.001, for post-intervention 2013). As was the case for all HIV cases, after adjustment for variables that were specifically changed by the intervention, the intervention’s effect on mortality was no longer statistically significant (aHR 1.033, 95% CI 0.601–1.774, p = 0.907, for post-intervention 2012; aHR 0.915, 95% CI 0.534–1.567, p = 0.746, for post-intervention 2013). Download: PPT PowerPoint slide PNG larger image TIFF original image Table 3. Mortality among newly diagnosed treatment-eligible HIV cases (with CD4 count ≤ 350 cells/mm3 or missing CD4 but reported as AIDS cases) during the pre-intervention 2010, pre-intervention 2011, post-intervention 2012, and post-intervention 2013 phases, based on Cox model analysis. https://doi.org/10.1371/journal.pmed.1001874.t003 The cost analysis results are shown in Table 4. The total cost of the simplified test and treat intervention implementation was US$14,525.75, with a cost of US$13,750.00 and US$775.75 for the first and second year, respectively. The unit cost for an additional patient receiving ART attributable to the intervention was US$83.80, with a cost of US$147.46 and US$9.69 for the first and second year, respectively (Table 5). The unit cost of a death prevented because of the intervention was US$234.52, with US$420.08 in the initial year and US$26.56 in the second year. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 4. Cost analysis of implementation of the intervention in Guangxi, China, 2012–2014. https://doi.org/10.1371/journal.pmed.1001874.t004 Download: PPT PowerPoint slide PNG larger image TIFF original image Table 5. Unit cost per patient receiving ART, unit cost per additional patient receiving ART because of the intervention, and unit cost per death prevented because of intervention in Guangxi, China, 2012–2014. https://doi.org/10.1371/journal.pmed.1001874.t005 Discussion Our results show that a simplified HIV test and treat intervention incorporating a streamlined, standardized time frame for diagnosis and expanded access to ART, irrespective of CD4 cell count, resulted in a significant increase in ART coverage within 90 d of diagnosis of HIV infection, from below 36% to over 90% among all newly diagnosed HIV/AIDS cases, and from under 47% to over 93% among newly diagnosed cases with CD4 count ≤ 350 cells/mm3 or missing CD4 but reported as AIDS cases. The simplified HIV test and treat intervention was also associated with significantly reduced overall mortality, from about 26% to fewer than 10%. Our results also show that the cost of the intervention was quite low, and mostly accrued in the initial year, as we set up the intervention. The unit cost per additional patient receiving ART declined to US$9.69 in the second year. Similarly, the unit cost per death prevented attributable to the intervention was US$234.52 over the study period, and it declined to US$26.56 in the second year of the study. Based on these results, we feel that the simplified test and treat intervention represents an effective and sustainable structural intervention that requires very little further investment once it is set up. China is committed to providing universal access to HIV testing and treatment in concordance with the WHO Treatment 2.0 strategy [1]. However, optimizing the cascade of HIV care, particularly during pre-treatment follow-up and monitoring, remains a key challenge. To address this issue, NCAIDS, China CDC, prioritized the identification, implementation, and assessment of new strategies for improved engagement in HIV testing and treatment. Few other studies have addressed the test and treat strategy in Asia. Our findings show that the implementation of a simplified test and treat intervention was associated with a significant decrease in mortality rate. Furthermore, this is the first study to our knowledge to show that a structural intervention to streamline HIV testing procedures and to expand treatment eligibility can have a substantial impact on the full HIV care cascade and patient mortality. We present evidence that a reconceptualization of HIV testing and treatment policies can have a dramatic effect on patient outcomes, even in the absence of new diagnostic technologies. Our findings provide program-based evidence to support widespread implementation of this intervention in China and to support similar test and treat strategies elsewhere. Given the size of China’s HIV-positive population, the simplified test and treat intervention at scale could streamline the diagnostic process for hundreds of thousands of patients. Over the pre-intervention phases, 33%–40% of patients failed to complete CD4 testing within 30 d of HIV confirmation. This is concerning because the CD4 count is the primary determinant of ART eligibility under the standard of care. While the proportion of patients who had CD4 count > 350 cells/mm3 remained consistent throughout the study (27%), we noted a higher proportion of patients who had CD4 count ≤ 200 cells/mm3 or who were reported as AIDS cases during the pre-intervention phases. This suggests that under the standard-of-care practice (i.e., the pre-intervention study phases), patients who failed to complete CD4 testing were more likely to have lower CD4 cell counts. These results reinforce previous findings that late HIV diagnosis remains a crucial challenge for HIV care providers in China and globally [14,23–25]. Complex testing procedures increase the risk for LTFU. Under the standard of care, the steps of HIV testing and treatment initiation were not co-located. Rather, patients often had to attend several medical facilities, such as the original facility that provided screening (e.g., township clinic), the local CDC, and the county hospital, in order to confirm the diagnosis and to begin treatment. This promotes structural delays in the initiation of treatment and increases the likelihood of attrition. The simplified test and treat intervention provided an opportunity to streamline the procedures from a patient-centered perspective. After the initial screening, all care visits occurred at a single location, the county general hospital. Furthermore, providing concurrent CD4 and WB confirmatory testing reduced the total number of visits, and thus delays, before beginning ART. Our data suggest that the simplified test and treat intervention had significant overall success in expanding access to CD4 testing and promoting initiation of ART, with the consequent favorable impact of decreased mortality. A small proportion of patients (8.8%–11%) in the two post-intervention phases still failed to initiate treatment, which may be due to stigma, competing needs, insufficient understanding of ART’s benefits, or voluntary refusal of treatment [6,26–29]. To minimize this proportion, the simplified test and treat intervention has multiple built-in opportunities during the pre-treatment process to deliver pre-ART counseling and to encourage ART engagement. In our study, the overwhelming majority of patients were willing to initiate ART, irrespective of CD4 cell count. The initiation of ART for all HIV/AIDS cases is likely to be facilitated in the future as emerging global guidelines evolve to embrace the recently released results of the TEMPRANO [30] and START [31] studies, which definitively confirmed that immediate initiation of ART, irrespective of CD4 cell count, is associated with significant reductions in clinical disease progression [32]. We noted that the number of newly diagnosed patients in the two pre-intervention phases (281 and 339) was higher than the number in the two post-intervention phases (215 and 199) and that the mortality in the two pre-intervention phases (26.7% and 26.5%) was also higher than in the two post-intervention phases (9.8% and 10.1%). The number of people who are diagnosed per year depends on the regional HIV incidence, the performance of HIV testing programs, and individual willingness to be tested. In 2010 (immediately prior to the pre-intervention 2010 phase), Guangxi launched a province-wide 5-y public campaign to promote HIV testing. This may have led to a sharp increase in the number of people being tested and diagnosed in 2010 and 2011 (i.e., the two pre-intervention phases). The later stages of the HIV testing promotion campaign in 2012–2013 coincided with the intervention phase. We suspect that, by this time, there may have been a lower proportion of undiagnosed HIV cases, which led to fewer newly diagnosed individuals in the intervention phase compared to the pre-intervention phase. Compared to the post-intervention 2012 phase (n = 215), the post-intervention 2013 phase experienced an additional but modest decrease in the number of new HIV cases (n = 199). The concomitant testing promotion campaign in Guangxi may have potentially uncovered the sickest individuals first, who would have been a part of the pre-intervention outcomes, and gradually identified healthier patients, who went into the post-intervention phases. This could have contributed to the observed reduction in mortality in our study. However, the fact that the observed mortality went from above 35% before to below 13% after the implementation of the simplified test and treat intervention among the subset of patients with CD4 count ≤ 350 cells/mm3 or missing CD4 but reported as AIDS cases is reassuring in this regard. We noticed that from the pre-intervention 2010 phase to the pre-intervention 2011 phase the proportion of patients initiating ART within 30 d of HIV confirmation increased from 9% to 20.1% and that overall ART initiation increased from 27% to 48.7%; in other words, ART initiation doubled, but mortality remained unchanged from the pre-intervention 2010 phase to the pre-intervention 2011 phase. This might be explained if the pre-intervention increase in ART initiation comprised mainly healthier patients, while the sickest patients continued to die at the same rate. The simplified test and treat intervention is likely to speed up treatment initiation effectively but may also change who gets treated. Please also note that though ART initiation doubled from the first to the second pre-intervention phase, ART coverage was still very low and therefore not sufficient to favorably impact mortality. The improved ART coverage from the first to the second pre-intervention phase should not undermine the effect of the intervention—although it may change in part the “mechanism” of the effect and the nature of the intervention. The fact that ART coverage within 30 d of HIV diagnosis increased sharply to 81.4% in the post-intervention 2012 phase and remained at 80% in the post-intervention 2013 phase is unlikely to be due solely to time secular trends; rather, we believe that the intervention was a major contributor to the increase in ART coverage. In 2013, WHO changed its ART guidelines to recommend ART initiation for patients with CD4 count ≤ 500 cells/mm3, rather than at the previously recommended threshold of 350 cells/mm3 [33]. It is estimated that these revised guidelines have increased the total number of ART-eligible people in low- and middle-income countries from 16.7 million to 25.8 million [34]. In resource-limited settings, using CD4 testing as the basis for HIV staging increases the demands on clinic and laboratory staff compared to clinical staging. However, some jurisdictions in China have already embraced universal ART eligibility, irrespective of CD4 cell count, as proposed in our study [35]. Such a policy maximizes the benefits of ART in reducing HIV disease progression to AIDS and premature death and at the same time reduces HIV transmission and thereby maximizes the potential of the treatment as prevention strategy [36–39]. The HIV epidemic in China presents unique challenges. Because of the country’s size, a low national prevalence of <1% still translates into a very large HIV-positive population of approximately 780,000 [40]. In Guangxi, the epidemic is characterized by a high proportion of late-diagnosed cases, and delayed ART initiation due to the additional time required for progressing through the sequential steps of the cascade of care. With a simplified cascade, both diagnosis and ART initiation occurred earlier, resulting in enhanced engagement in ART and decreased mortality. We designed the simplified test and treat intervention specifically for the context of China’s HIV epidemic and health care structure, and, as a result, the specifics of the intervention procedures and results may not be fully generalizable to other countries. Nevertheless, the intervention targets the issues of late diagnosis and procedural barriers that delay care, which are common challenges in HIV programs worldwide. We believe that our experience is beneficial to the global community in that it provides evidence in support of the feasibility and clinical benefits of the test and treat strategy, particularly for settings where late diagnosis and treatment are common. Although the pre-intervention/post-intervention study design allowed for control of some hospital-based characteristics, assumptions about the simplified test and treat intervention’s causal effects must be treated with caution. Internal validity is affected by history, maturation, and Hawthorne threats. Implementing a package of interventions also prevents the evaluation of the relative contribution of each of the individual components on the overall observed effect. Late diagnosis (CD4 count ≤ 200 cells/mm3) of HIV is a key barrier to reducing mortality and morbidity, which we could not evaluate fully in this study. In the intervention phase, we used data collected over the course of standardized care, while outcomes in the two pre-intervention phases were assessed retrospectively. The pre-intervention data monitoring was less thorough than the data monitoring in the two post- intervention phases, which may have had an impact on data quality. During the two pre-intervention phases, reporting of outcomes such as mortality may have been delayed, resulting in a lower reported baseline mortality and an underestimation of the impact of the intervention. China has begun expanding the simplified test and treat intervention to 12 additional counties in nine provinces. Further evaluation of the feasibility and acceptability of this intervention will provide additional evidence for national and international policymakers. Early diagnosis in conjunction with prompt linkage to care could be a game changer for the HIV epidemic in China. Our results demonstrate that a simplified HIV testing approach combined with expanded access to ART, irrespective of CD4 count, can lead to a substantial reduction in mortality. Our findings support increased integration of HIV testing and treatment to optimize the potential individual and public health benefits of ART. Supporting Information S1 Checklist. STROBE checklist. https://doi.org/10.1371/journal.pmed.1001874.s001 (DOC) S1 Text. Clinical trial registration receipt. https://doi.org/10.1371/journal.pmed.1001874.s002 (PDF) S2 Text. Study proposal. https://doi.org/10.1371/journal.pmed.1001874.s003 (PDF) S3 Text. Institutional review board approval letter. https://doi.org/10.1371/journal.pmed.1001874.s004 (PDF) Acknowledgments The authors would like to thank Hui Wei, Qiuying Zhu, Xiaoai Qian, and Peili Wu for their contributions to the implementation of this project in the field. We also thank Willa Dong and Jonas Tillman for their help in editing the manuscript.

doi: 10.1371/journal.pmed.1001869pmid: 26327271

Background HIV impacts heavily on the operating costs of companies in sub-Saharan Africa, with many companies now providing antiretroviral therapy (ART) programmes in the workplace. A full cost–benefit analysis of workplace ART provision has not been conducted using primary data. We developed a dynamic health-state transition model to estimate the economic impact of HIV and the cost–benefit of ART provision in a mining company in South Africa between 2003 and 2022. Methods and Findings A dynamic health-state transition model, called the Workplace Impact Model (WIM), was parameterised with workplace data on workforce size, composition, turnover, HIV incidence, and CD4 cell count development. Bottom-up cost analyses from the employer perspective supplied data on inpatient and outpatient resource utilisation and the costs of absenteeism and replacement of sick workers. The model was fitted to workforce HIV prevalence and separation data while incorporating parameter uncertainty; univariate sensitivity analyses were used to assess the robustness of the model findings. As ART coverage increases from 10% to 97% of eligible employees, increases in survival and retention of HIV-positive employees and associated reductions in absenteeism and benefit payments lead to cost savings compared to a scenario of no treatment provision, with the annual cost of HIV to the company decreasing by 5% (90% credibility interval [CrI] 2%–8%) and the mean cost per HIV-positive employee decreasing by 14% (90% CrI 7%–19%) by 2022. This translates into an average saving of US$950,215 (90% CrI US$220,879–US$1.6 million) per year; 80% of these cost savings are due to reductions in benefit payments and inpatient care costs. Although findings are sensitive to assumptions regarding incidence and absenteeism, ART is cost-saving under considerable parameter uncertainty and in all tested scenarios, including when prevalence is reduced to 1%—except when no benefits were paid out to employees leaving the workforce and when absenteeism rates were half of what data suggested. Scaling up ART further through a universal test and treat strategy doubles savings; incorporating ART for family members reduces savings but is still marginally cost-saving compared to no treatment. Our analysis was limited to the direct cost of HIV to companies and did not examine the impact of HIV prevention policies on the miners or their families, and a few model inputs were based on limited data, though in sensitivity analysis our results were found to be robust to changes to these inputs along plausible ranges. Conclusions Workplace ART provision can be cost-saving for companies in high HIV prevalence settings due to reductions in healthcare costs, absenteeism, and staff turnover. Company-sponsored HIV counselling and voluntary testing with ensuing treatment of all HIV-positive employees and family members should be implemented universally at workplaces in countries with high HIV prevalence. Background Every year, more than 2 million people become newly infected with HIV, the virus that causes AIDS, usually by having unprotected sex with an infected partner. People in the early stages of HIV infection rarely have any symptoms, but, over time, HIV destroys CD4 lymphocytes and other immune system cells, and, eventually, HIV-positive individuals become susceptible to numerous other infections. Because many of these infections are extremely serious, early in the AIDS epidemic, most HIV-infected individuals died within ten years of infection. Then, in 1996, effective antiretroviral therapy (ART)—cocktails of drugs that stop HIV replicating—became available. For people living in affluent countries, HIV/AIDS became a chronic condition, but because ART was expensive, HIV/AIDS remained fatal in low- and middle-income countries. In 2003, the international community began to work towards achieving universal access to ART. By 2013, nearly 13 million HIV-positive people—more than a third of the global HIV-infected population—had access to ART. Why Was This Study Done? HIV disease hits individuals in the prime of their working lives, thereby increasing absenteeism, the turnover of labor, and the operating costs of companies working in countries where HIV infection is common (high HIV prevalence). To reduce the economic impact of HIV/AIDS, some companies provide their workforces in such countries with comprehensive HIV services that include counseling and testing, and ART. For example, mining companies in South Africa (where nearly 20% of the working-age population is HIV-positive) provide HIV services to their workforces. However, although there is strong evidence that HIV disease increases the cost of doing business, a full cost–benefit analysis (the quantification of both the costs and benefits of a business strategy or medical intervention) of ART provision in the workplace based on real-world data has not been undertaken. Here, the researchers use a mathematical model to estimate the economic impact of HIV and the costs and benefits of company-level ART provision by a South African mining company between 2003 and 2022. What Did the Researchers Do and Find? The researchers developed a mathematical model to evaluate the past and future impact and costs to the employer of an ART program provided since 2002 by a coal mining company operating at a number of South African colleries. They fed data on the workforce’s characteristics, the annual number of new HIV infections in the workforce, the CD4 cell counts of HIV-positive employees, healthcare resource utilization, and the costs of absenteeism and labor turnover into the model. The model estimated that, as ART coverage increased from 10% to 97% of eligible employees, increases in the survival and retention of HIV-positive employees and reductions in absenteeism and benefit payments would lead to overall cost savings compared to a scenario of no ART provision. Specifically, the annual cost of HIV to the company would decrease by 5% and the average cost per HIV-positive employee would decrease by 14% by 2022. These changes in costs (which mainly accrue from reductions in benefit payments for death and ill-health retirement and in employee healthcare costs) translate into average savings of nearly US$1 million per year. Finally, scaling up ART coverage through a universal test and treat strategy would double savings, whereas providing ART for family members as well as employees would reduce savings but remain marginally cost-saving compared to no ART. What Do These Findings Mean? These findings suggest that workplace ART provision can be cost-saving for companies operating in settings with a high HIV prevalence because of reductions in healthcare costs, absenteeism, and staff turnover. That is, the costs to the employer of providing ART can be less than the costs saved by reducing healthcare use, absenteeism, and worker turnover. The accuracy of these findings depends on the quality of the data used to run the model. However, additional analyses indicate that ART provision is likely to be cost-saving unless people receive no benefits on leaving the workforce or the absenteeism rate is considerably lower than the available data suggest. Thus, the researchers propose that company-sponsored counseling and voluntary HIV testing with treatment of all HIV-positive employees and family members should be implemented universally at workplaces in countries with a high HIV prevalence. Such a strategy should be cost-saving for employers and might also take some pressure off resource-limited public sector ART programs. Additional Information This list of resources contains links that can be accessed when viewing the PDF on a device or via the online version of the article at http://dx.doi.org/10.1371/journal.pmed.1001869. Information is available from the US National Institute of Allergy and Infectious Diseases on HIV infection and AIDS NAM/aidsmap provides basic information about HIV/AIDS, summaries of recent research findings on HIV care and treatment, and personal stories about living with HIV/AIDS Information is available from Avert, an international AIDS charity, on many aspects of HIV/AIDS, including information on ART, universal access to ART, and HIV/AIDS in South Africa; Avert also provides personal stories about living with HIV/AIDS The World Health Organization provides information on all aspects of HIV/AIDS (in several languages), including its guidelines on the use of ART for treating and preventing HIV infection The UNAIDS Fast-Track Strategy to End the AIDS Epidemic by 2030 provides up-to-date information about the AIDS epidemic and efforts to halt it; UNAIDS also provides detailed information about HIV/AIDS in South Africa Wikipedia has a page about cost–benefit analysis (note that Wikipedia is a free online encyclopedia that anyone can edit; available in several languages) The HIV Modelling Consortium has a database of models used in analyzing the impact of HIV and ART, including models such as the one used in this analysis The International AIDS Economics Network has a library of research on the economics of HIV around the world Introduction HIV disease hits adults in the prime of their working lives. Companies therefore take a heavy toll in countries with high HIV prevalence [1,2]. To counter this, some companies provide their workforce with a number of HIV services, ranging from prevention activities to HIV testing and antiretroviral therapy (ART). While several companies in sub-Saharan Africa started ART programmes from 2002 onwards [3–5], quantifying these programmes’ costs and benefits has proven difficult [3]. Even in sophisticated in-house medical programmes, longitudinal data collection is fraught with difficulty, and the relationship between costs and benefits, such as regained productivity, can be hard to establish [3]. This makes it hard for companies to plan and budget for additional HIV-specific health programmes, and impossible to ascertain the programme’s impact on the company’s operations and profits. HIV disease increases rates of absenteeism, labour force turnover, and, ultimately, the costs of company operations in sub-Saharan Africa. A number of studies have quantified the impact of HIV on labour forces in the region, with the cost of HIV ranging from 0.7% of wages [6] or 1% of labour cost [7] to 1%–9% of profits [8]. Only one study, amongst Kenyan tea pluckers, has estimated the impact of HIV on the productivity of a single worker, finding an 18% decrease in earnings in the year before termination amongst HIV-positive workers [9], in a setting where earnings are directly related to productivity. South Africa is the sub-Saharan African country with the largest number of people living with HIV [10,11], with 18.8% of the working-age population (15–49 y old) being HIV infected [12]. In the last large-scale survey of 22 companies in South Africa, between 1999 and 2005, the workforce HIV prevalence in a non-representative sample averaged 11% [13], though estimates varied over time and between industries [3,13]. Similarly, the costs of HIV vary, with the estimated increase due to HIV in the cost of doing business (termed AIDS “tax” [1]) ranging from 0.4% to 5.9% of the annual wage bill of six South African companies in 2001 [1,2], or a 0.6%–10.8% increase in labour costs amongst companies from six countries in sub-Saharan Africa [3]. The cost per employee also varies considerably by skill level [2]. None of these studies, however, included the impact of workplace ART provision. HIV care, including ART, has been provided by mining companies in South Africa since 2002, predating ART provision in the public sector [4,5]. While there are numerous estimates of the cost [14–27] and cost-effectiveness [28–45] of public sector ART provision in South Africa, the cost and impact of private sector ART provision at the workplace level have not yet been established. And while some aspects of this impact have been estimated in other countries, such as Kenya [46–50], Botswana [50], and Uganda [51], none of these estimates included productivity as well as healthcare costs, and none was a full cost–benefit analysis based on real-world programme data. In order to provide evidence for company management and policy-makers alike, we evaluated the impact and cost of both HIV and ART in a mining company in South Africa, and analysed the incremental cost–benefit balance of the company’s ART programme compared to no ART provision. Methods Workplace under Study We report on the ART programme of a coal mining company operating at a number of collieries in Mpumalanga province since 2002. The programme is run from the mines’ own clinics and hospitals and provides care for employees, contractors, and employees’ dependants. Annual anonymous HIV counselling and testing (HCT) campaigns in the mines provide easy access to testing. HIV-positive employees are enrolled in an HIV wellness programme that provides CD4 cell count testing every 3 mo and interventions, such as isoniazid and cotrimoxazole prophylaxis, for the prevention and treatment of opportunistic infections. Employees were initiated on ART once their CD4 cell count was at or below 250 cells/mm3 during the period 2003–2007, or at or below 350 cells/mm3 during 2008–2010, or if presenting with WHO stage 3/4 disease, and their CD4 cell count and viral load (VL) were monitored twice annually thereafter. By the end of 2010, out of 9,252 employees, 1,149 had tested HIV positive in confirmatory tests and had been enrolled in the company’s wellness programme. Since 2002, 629 employees have been initiated on ART, with 555 employees retained on ART by the end of 2010. Model Description A dynamic Markov health-state transition model, the Workplace Impact Model (WIM), was developed to evaluate both the past and future impact and costs of introducing ART into the workforce from the perspective of the employer. The model is run twice, under a scenario of no ART provision (no ART scenario) and again under a scenario representing the scale-up of ART in the workforce (ART scenario). Both scenarios also include the cost and impact of other components of HIV healthcare such as HIV testing, wellness care, and other outpatient and inpatient care for HIV. The model projects the HIV-positive and-negative workforce over 20 y from 2003, taking into account planned changes to the workforce size as well as ageing and promotions. This time period is necessary to capture the full impact of the gradual scale-up of ART. The model calculates, in 3-mo time steps, employees’ HIV prevalence, their HIV test uptake and coverage with and loss from wellness and ART care, the number of employees leaving the workforce as a result of mortality and morbidity due to HIV (separations), the number of recruits to the workforce (some of which are HIV infected) that are required to offset this loss, the change in CD4 cell count (an indicator of immune system function) in HIV-positive employees, and the incremental costs of the ART programme itself, of additional outpatient and inpatient healthcare, and of absenteeism and workforce turnover (Fig 1). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. Population model of changes within the workforce. Recruits join the susceptible or infected (I) workforce depending on their HIV status at first employment. Employees move from the susceptible to the infected population according to prevalence and incidence. In the infected population, employees change between sub-populations representing different types of care (not tested, tested but not yet in care, wellness care, successful first- or second-line ART, and first-line or second-line treatment failure) according to coverage rates and, in case of treatment failure, to failure rates. Employees can drop out of care, i.e., be lost to retention, at any time and go back to the no care sub-population according to loss-to-retention rates; they can also leave the workforce for reasons related or unrelated to HIV (separations). Within each of the sub-populations, additional unidirectional changes due to ageing and promotion rates apply (not shown here); within each of the infected sub-populations, additional bi-directional changes due to transitions between CD4-cell-count-defined health states apply. https://doi.org/10.1371/journal.pmed.1001869.g001 In order to capture important differences in survival and/or in healthcare and absenteeism costs, the HIV-infected workforce is divided into two genders, three age groups, six job grades, and five CD4-cell-count-defined health states, although not every parameter is differentiated by all four categories. Table 1 summarises the population categories used in the model; Table 2 gives more detail on the stratification levels. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 1. Job grade, health state, and age group categories used in model. https://doi.org/10.1371/journal.pmed.1001869.t001 Download: PPT PowerPoint slide PNG larger image TIFF original image Table 2. Details of parameter estimation, level of stratification, and data sources. https://doi.org/10.1371/journal.pmed.1001869.t002 Due to the difficulty in capturing the programme’s benefit to dependants, this analysis is limited to employees. The model incorporates HIV incidence in the workforce but does not model HIV transmission from the workforce or the effect of ART on HIV transmission. Separations, i.e., losses to the workforce other than through retirement or retrenchment, most often due to ill-health or death, are differentiated into three categories (death, ill-health/disability, and other) in the model and are further differentiated by HIV status, job grade for HIV-negative employees, and CD4 cell count stratum for HIV-positive employees. More details on the methods used in estimating each parameter are given in Tables 2–4 and in S1 Text, which also gives the model equations. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 3. Values and sources of main model inputs and assumptions. https://doi.org/10.1371/journal.pmed.1001869.t003 Download: PPT PowerPoint slide PNG larger image TIFF original image Table 4. Values and sources of main model inputs and assumptions (HIV-related separations only). https://doi.org/10.1371/journal.pmed.1001869.t004 Model Parameterisation The model was parameterised with company data on the size, composition, and turnover of the workforce at the mines obtained from the company employee database of 9,211 employees, covering the period January 2003 to December 2010 and including job grade, gender, engagement and termination dates, and the coverage and results of the serial HCT campaigns. Annual coverage with linked workplace HCT campaigns increased from 40% of all employees in 2003 to 86% in 2008, enabling a reliable estimation of HIV incidence in later years. A separate database documenting the 1,149 employees who tested HIV positive and were enrolled in the company’s HIV care programme over the same period of time provided inputs regarding coverage of wellness care and ART, retention in care, development of treatment failure, and employees’ CD4 cell counts over time. The two databases were anonymously linked for this analysis. We parameterised the model with annual HCT and ART coverage, HIV prevalence in new employees joining the workforce, as well as the incidence of treatment failure and loss to retention in the programme as reported in these databases. Based on these data, HCT coverage was set to reach 92% by 2010 and to remain constant thereafter. The HCT data were also used to estimate the HIV incidence and prevalence amongst all employees. Incidence was estimated for those employees with two or more HIV tests, with HIV conversion assumed to be at the midpoint between the first positive and the last prior negative HIV test [53]. These data suggested that HIV incidence varied between 1.2 and 2.6 per 100 employee-years in the workforce throughout and that prevalence increased from 11% in 2005 to 16% in 2010. ART coverage of those eligible was calibrated to increase from 11% in 2003 to 68% in 2010, as suggested by the workforce data, and was modelled to reach 88% by 2013 and 100% by 2022. First-line treatment failure was set to vary between 8% and 11% per year, and loss to follow-up between 6% and 12% per year, likely including some migration to ART programmes outside the workforce. The values of important model parameters are summarised in Tables 3 and 4; the remainder of the parameters and their 95% confidence intervals are available in S1 Text. Transition Probabilities A detailed electronic register including the results of all CD4 cell count measurements (every 3 mo) from all HIV-positive employees for the same period as the workforce database (January 2003–December 2010) was used to estimate the transition probabilities between CD4-cell-count-defined health states for the wellness care and ART populations (Table 5). The database contained a total of 10,972 CD4 cell count test results, with a mean patient follow-up of 961 d (maximum 2,822 d). Since almost all employees who test HIV positive in the workplace testing programme immediately enter care, we used historic data from the South African public sector to parameterise the transitions for the undiagnosed and no care populations [52]. Because of insufficient data, these transitions were also applied to the treatment failure population. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 5. Model 3-mo transition probabilities between CD4-cell-count-defined health states by type of care. https://doi.org/10.1371/journal.pmed.1001869.t005 Each employee’s available CD4 cell count data were allocated to each type of care in 3-mo time periods from the start date for this type of care up until the time period including the stop date for this type of care. If CD4 cell counts were missing for one or two consecutive time periods, they were linearly interpolated from the CD4 cell counts of the two adjacent time periods. These CD4 cell counts were then allocated to five different CD4 cell count strata, which in turn defined the model health states (see Table 1). For the calculation of transition probabilities, in order to differentiate between patients in wellness care and those accessing ART outside the company healthcare system, CD4 cell counts were considered to be wellness care CD4 cell counts only if any VL measured during the same 3-mo time period was unsuppressed (>50 copies/ml). If a suppressed VL count was found before the date of ART initiation in the workforce programme, the patient was deleted from the wellness care CD4 analysis. In order to exclude patients in treatment failure, CD4 cell counts were considered to be ART CD4 cell counts only if any VL measured during the same time period was suppressed (≤50 copies/ml), though the patient could still contribute other (i.e., earlier or later) CD4 cell counts to the ART CD4 population if they coincided with a suppressed VL. Cost Data A bottom-up patient-level analysis of economic costs from the employer perspective was conducted in 2006 to quantify all costs of HIV/AIDS to the company. The analysis, which has been described in detail elsewhere [54,55], included the cost of the ART programme, including the cost of antiretroviral drugs, ART-specific laboratory tests such as CD4 cell count and VL, and management and training costs within and above the facility level, as well as any other HIV-related cost such as inpatient and outpatient resource utilisation and costs, and the costs of absenteeism and replacing a sick or deceased worker, including the benefits paid to the worker or his/her family and the costs of recruiting and training a replacement. Healthcare resource use, quantified as the number of inpatient days and outpatient visits, was abstracted from record systems at the company health centres and averaged by CD4 cell count stratum, based on the employee’s most recent CD4 cell count. Absenteeism was calculated as the median number of days of sick leave of patients in wellness care and on ART by CD4 cell count stratum, based on the company’s payroll data. Both healthcare and absenteeism costs were calculated incrementally to that of HIV-negative employees. Due to the choice of an employer perspective, costs to the employee and the broader society were excluded, but since most employees of the mining company seek care at the workplace clinics and hospitals, resource use captured for this analysis is unusually complete. Cost inputs are summarised in Table 6. Cost data were collected in South African rands (ZAR) during 2006/2007, adjusted for inflation to 2010, and converted to US dollars (USD) using the 2010 average conversion rate of 8 ZAR/1 USD (S1 Text contains an explanation of the time period for inflation adjustment). Costs are presented undiscounted and discounted at 5% per annum, the repurchase rate of the South African Reserve Bank during most of the analysis period [56]. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 6. Annual per employee cost and frequency of absenteeism by CD4 cell count category, incremental to that of HIV-negative employees. https://doi.org/10.1371/journal.pmed.1001869.t006 Model Calibration and Sensitivity Analysis Because sampling uncertainty surrounds many of the important model parameters, we defined probability distributions around the main inputs, with the distributions based on the primary workforce, absenteeism, and cost data used in this analysis. Some parameters were also stratified by CD4 cell count or job grade (separation rates) or were time dependent (treatment failure probability). Statistical distributions were assigned to these parameters based on standard practice in economic evaluations [57], with specific details included in S1 Text. To calibrate the model while accounting for this sampling uncertainty, 20,000 parameter sets were randomly sampled (using Latin hypercube sampling) from the parameter distributions, and the resulting model runs were compared to see if they fit within the uncertainty range for the observed HIV prevalence of the workforce in 2010 (12.8%–19.2%) and the average annual number of separations in HIV-positive (50–150) and HIV-negative (200–500) employees during 2005–2009. The 998 model runs that fit these data were then used to assess the uncertainty around our main outcomes (total costs, cost savings, and HIV prevalence), with medians and 90% credibility intervals (CrIs) being produced for each outcome. In addition, an analysis of co-variance was undertaken to quantify the contribution of different parameters to the uncertainty in the projected undiscounted savings due to ART. Additionally, we undertook univariate sensitivity analyses on selected parameters, examining the impact of the following: reducing all absenteeism by half; assuming the same absenteeism on ART as off ART; assuming the same ART cost and health-state transition probabilities as found in analyses of public sector ART provision in South Africa using similar methodology [58,59]; changing inpatient and outpatient costs by ±50% (note that in each instance only the extremes of the range were considered); changing the number of annual salary equivalents paid out as benefits to 0, 1, or 2 y instead of 3; changing HIV-dependent separation rates by ±20%; changing incidence by ±50%; and, in order to examine the generalisability of results to a setting with low HIV prevalence, reducing incidence to an extremely low value of 0.0001 and prevalence in the start population and amongst new recruits each to a tenth of the baseline values. For each of these sensitivity analyses, the effect of the parameter change was evaluated on all the baseline model fits so that an average effect could be estimated. Lastly, in order to analyse the future impact of changes in treatment policies, we parameterised the model for two additional scenarios to be implemented from 2013 onwards. First, we considered a universal test and treat scenario in which HCT coverage was 100% each year, and 100% of employees who tested HIV-positive initiated ART within 6 mo, regardless of CD4 cell count or clinical status. We conservatively assumed no impact of this high-level ART coverage on HIV incidence since the intervention would cover only employees and not their sexual partners. In a second scenario (“family treatment”), we incorporated the extension of ART to those family members of employees who were eligible for ART, with an assumed average of one ART-eligible dependant per HIV-positive employee on ART. Ethics Approval The study was reviewed and approved by the following ethics committees: the London School of Hygiene & Tropical Medicine Ethics Committee (application number 962), the Anglogold Health Service Research Ethics Committee (AHS REC 004/02), and the University of KwaZulu-Natal Biomedical Research Ethics Committee (BE093/08). Employees’ consent to participation in this study was waived as we used only data that were collected for routine care purposes and, as in most other routine care settings, employees did not give written consent for this care. Data Availability The fully parameterised model that incorporates all data and that was used to produce all projections within this paper can be downloaded from OpenBU via http://hdl.handle.net/2144/10817. Workplace under Study We report on the ART programme of a coal mining company operating at a number of collieries in Mpumalanga province since 2002. The programme is run from the mines’ own clinics and hospitals and provides care for employees, contractors, and employees’ dependants. Annual anonymous HIV counselling and testing (HCT) campaigns in the mines provide easy access to testing. HIV-positive employees are enrolled in an HIV wellness programme that provides CD4 cell count testing every 3 mo and interventions, such as isoniazid and cotrimoxazole prophylaxis, for the prevention and treatment of opportunistic infections. Employees were initiated on ART once their CD4 cell count was at or below 250 cells/mm3 during the period 2003–2007, or at or below 350 cells/mm3 during 2008–2010, or if presenting with WHO stage 3/4 disease, and their CD4 cell count and viral load (VL) were monitored twice annually thereafter. By the end of 2010, out of 9,252 employees, 1,149 had tested HIV positive in confirmatory tests and had been enrolled in the company’s wellness programme. Since 2002, 629 employees have been initiated on ART, with 555 employees retained on ART by the end of 2010. Model Description A dynamic Markov health-state transition model, the Workplace Impact Model (WIM), was developed to evaluate both the past and future impact and costs of introducing ART into the workforce from the perspective of the employer. The model is run twice, under a scenario of no ART provision (no ART scenario) and again under a scenario representing the scale-up of ART in the workforce (ART scenario). Both scenarios also include the cost and impact of other components of HIV healthcare such as HIV testing, wellness care, and other outpatient and inpatient care for HIV. The model projects the HIV-positive and-negative workforce over 20 y from 2003, taking into account planned changes to the workforce size as well as ageing and promotions. This time period is necessary to capture the full impact of the gradual scale-up of ART. The model calculates, in 3-mo time steps, employees’ HIV prevalence, their HIV test uptake and coverage with and loss from wellness and ART care, the number of employees leaving the workforce as a result of mortality and morbidity due to HIV (separations), the number of recruits to the workforce (some of which are HIV infected) that are required to offset this loss, the change in CD4 cell count (an indicator of immune system function) in HIV-positive employees, and the incremental costs of the ART programme itself, of additional outpatient and inpatient healthcare, and of absenteeism and workforce turnover (Fig 1). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 1. Population model of changes within the workforce. Recruits join the susceptible or infected (I) workforce depending on their HIV status at first employment. Employees move from the susceptible to the infected population according to prevalence and incidence. In the infected population, employees change between sub-populations representing different types of care (not tested, tested but not yet in care, wellness care, successful first- or second-line ART, and first-line or second-line treatment failure) according to coverage rates and, in case of treatment failure, to failure rates. Employees can drop out of care, i.e., be lost to retention, at any time and go back to the no care sub-population according to loss-to-retention rates; they can also leave the workforce for reasons related or unrelated to HIV (separations). Within each of the sub-populations, additional unidirectional changes due to ageing and promotion rates apply (not shown here); within each of the infected sub-populations, additional bi-directional changes due to transitions between CD4-cell-count-defined health states apply. https://doi.org/10.1371/journal.pmed.1001869.g001 In order to capture important differences in survival and/or in healthcare and absenteeism costs, the HIV-infected workforce is divided into two genders, three age groups, six job grades, and five CD4-cell-count-defined health states, although not every parameter is differentiated by all four categories. Table 1 summarises the population categories used in the model; Table 2 gives more detail on the stratification levels. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 1. Job grade, health state, and age group categories used in model. https://doi.org/10.1371/journal.pmed.1001869.t001 Download: PPT PowerPoint slide PNG larger image TIFF original image Table 2. Details of parameter estimation, level of stratification, and data sources. https://doi.org/10.1371/journal.pmed.1001869.t002 Due to the difficulty in capturing the programme’s benefit to dependants, this analysis is limited to employees. The model incorporates HIV incidence in the workforce but does not model HIV transmission from the workforce or the effect of ART on HIV transmission. Separations, i.e., losses to the workforce other than through retirement or retrenchment, most often due to ill-health or death, are differentiated into three categories (death, ill-health/disability, and other) in the model and are further differentiated by HIV status, job grade for HIV-negative employees, and CD4 cell count stratum for HIV-positive employees. More details on the methods used in estimating each parameter are given in Tables 2–4 and in S1 Text, which also gives the model equations. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 3. Values and sources of main model inputs and assumptions. https://doi.org/10.1371/journal.pmed.1001869.t003 Download: PPT PowerPoint slide PNG larger image TIFF original image Table 4. Values and sources of main model inputs and assumptions (HIV-related separations only). https://doi.org/10.1371/journal.pmed.1001869.t004 Model Parameterisation The model was parameterised with company data on the size, composition, and turnover of the workforce at the mines obtained from the company employee database of 9,211 employees, covering the period January 2003 to December 2010 and including job grade, gender, engagement and termination dates, and the coverage and results of the serial HCT campaigns. Annual coverage with linked workplace HCT campaigns increased from 40% of all employees in 2003 to 86% in 2008, enabling a reliable estimation of HIV incidence in later years. A separate database documenting the 1,149 employees who tested HIV positive and were enrolled in the company’s HIV care programme over the same period of time provided inputs regarding coverage of wellness care and ART, retention in care, development of treatment failure, and employees’ CD4 cell counts over time. The two databases were anonymously linked for this analysis. We parameterised the model with annual HCT and ART coverage, HIV prevalence in new employees joining the workforce, as well as the incidence of treatment failure and loss to retention in the programme as reported in these databases. Based on these data, HCT coverage was set to reach 92% by 2010 and to remain constant thereafter. The HCT data were also used to estimate the HIV incidence and prevalence amongst all employees. Incidence was estimated for those employees with two or more HIV tests, with HIV conversion assumed to be at the midpoint between the first positive and the last prior negative HIV test [53]. These data suggested that HIV incidence varied between 1.2 and 2.6 per 100 employee-years in the workforce throughout and that prevalence increased from 11% in 2005 to 16% in 2010. ART coverage of those eligible was calibrated to increase from 11% in 2003 to 68% in 2010, as suggested by the workforce data, and was modelled to reach 88% by 2013 and 100% by 2022. First-line treatment failure was set to vary between 8% and 11% per year, and loss to follow-up between 6% and 12% per year, likely including some migration to ART programmes outside the workforce. The values of important model parameters are summarised in Tables 3 and 4; the remainder of the parameters and their 95% confidence intervals are available in S1 Text. Transition Probabilities A detailed electronic register including the results of all CD4 cell count measurements (every 3 mo) from all HIV-positive employees for the same period as the workforce database (January 2003–December 2010) was used to estimate the transition probabilities between CD4-cell-count-defined health states for the wellness care and ART populations (Table 5). The database contained a total of 10,972 CD4 cell count test results, with a mean patient follow-up of 961 d (maximum 2,822 d). Since almost all employees who test HIV positive in the workplace testing programme immediately enter care, we used historic data from the South African public sector to parameterise the transitions for the undiagnosed and no care populations [52]. Because of insufficient data, these transitions were also applied to the treatment failure population. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 5. Model 3-mo transition probabilities between CD4-cell-count-defined health states by type of care. https://doi.org/10.1371/journal.pmed.1001869.t005 Each employee’s available CD4 cell count data were allocated to each type of care in 3-mo time periods from the start date for this type of care up until the time period including the stop date for this type of care. If CD4 cell counts were missing for one or two consecutive time periods, they were linearly interpolated from the CD4 cell counts of the two adjacent time periods. These CD4 cell counts were then allocated to five different CD4 cell count strata, which in turn defined the model health states (see Table 1). For the calculation of transition probabilities, in order to differentiate between patients in wellness care and those accessing ART outside the company healthcare system, CD4 cell counts were considered to be wellness care CD4 cell counts only if any VL measured during the same 3-mo time period was unsuppressed (>50 copies/ml). If a suppressed VL count was found before the date of ART initiation in the workforce programme, the patient was deleted from the wellness care CD4 analysis. In order to exclude patients in treatment failure, CD4 cell counts were considered to be ART CD4 cell counts only if any VL measured during the same time period was suppressed (≤50 copies/ml), though the patient could still contribute other (i.e., earlier or later) CD4 cell counts to the ART CD4 population if they coincided with a suppressed VL. Cost Data A bottom-up patient-level analysis of economic costs from the employer perspective was conducted in 2006 to quantify all costs of HIV/AIDS to the company. The analysis, which has been described in detail elsewhere [54,55], included the cost of the ART programme, including the cost of antiretroviral drugs, ART-specific laboratory tests such as CD4 cell count and VL, and management and training costs within and above the facility level, as well as any other HIV-related cost such as inpatient and outpatient resource utilisation and costs, and the costs of absenteeism and replacing a sick or deceased worker, including the benefits paid to the worker or his/her family and the costs of recruiting and training a replacement. Healthcare resource use, quantified as the number of inpatient days and outpatient visits, was abstracted from record systems at the company health centres and averaged by CD4 cell count stratum, based on the employee’s most recent CD4 cell count. Absenteeism was calculated as the median number of days of sick leave of patients in wellness care and on ART by CD4 cell count stratum, based on the company’s payroll data. Both healthcare and absenteeism costs were calculated incrementally to that of HIV-negative employees. Due to the choice of an employer perspective, costs to the employee and the broader society were excluded, but since most employees of the mining company seek care at the workplace clinics and hospitals, resource use captured for this analysis is unusually complete. Cost inputs are summarised in Table 6. Cost data were collected in South African rands (ZAR) during 2006/2007, adjusted for inflation to 2010, and converted to US dollars (USD) using the 2010 average conversion rate of 8 ZAR/1 USD (S1 Text contains an explanation of the time period for inflation adjustment). Costs are presented undiscounted and discounted at 5% per annum, the repurchase rate of the South African Reserve Bank during most of the analysis period [56]. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 6. Annual per employee cost and frequency of absenteeism by CD4 cell count category, incremental to that of HIV-negative employees. https://doi.org/10.1371/journal.pmed.1001869.t006 Model Calibration and Sensitivity Analysis Because sampling uncertainty surrounds many of the important model parameters, we defined probability distributions around the main inputs, with the distributions based on the primary workforce, absenteeism, and cost data used in this analysis. Some parameters were also stratified by CD4 cell count or job grade (separation rates) or were time dependent (treatment failure probability). Statistical distributions were assigned to these parameters based on standard practice in economic evaluations [57], with specific details included in S1 Text. To calibrate the model while accounting for this sampling uncertainty, 20,000 parameter sets were randomly sampled (using Latin hypercube sampling) from the parameter distributions, and the resulting model runs were compared to see if they fit within the uncertainty range for the observed HIV prevalence of the workforce in 2010 (12.8%–19.2%) and the average annual number of separations in HIV-positive (50–150) and HIV-negative (200–500) employees during 2005–2009. The 998 model runs that fit these data were then used to assess the uncertainty around our main outcomes (total costs, cost savings, and HIV prevalence), with medians and 90% credibility intervals (CrIs) being produced for each outcome. In addition, an analysis of co-variance was undertaken to quantify the contribution of different parameters to the uncertainty in the projected undiscounted savings due to ART. Additionally, we undertook univariate sensitivity analyses on selected parameters, examining the impact of the following: reducing all absenteeism by half; assuming the same absenteeism on ART as off ART; assuming the same ART cost and health-state transition probabilities as found in analyses of public sector ART provision in South Africa using similar methodology [58,59]; changing inpatient and outpatient costs by ±50% (note that in each instance only the extremes of the range were considered); changing the number of annual salary equivalents paid out as benefits to 0, 1, or 2 y instead of 3; changing HIV-dependent separation rates by ±20%; changing incidence by ±50%; and, in order to examine the generalisability of results to a setting with low HIV prevalence, reducing incidence to an extremely low value of 0.0001 and prevalence in the start population and amongst new recruits each to a tenth of the baseline values. For each of these sensitivity analyses, the effect of the parameter change was evaluated on all the baseline model fits so that an average effect could be estimated. Lastly, in order to analyse the future impact of changes in treatment policies, we parameterised the model for two additional scenarios to be implemented from 2013 onwards. First, we considered a universal test and treat scenario in which HCT coverage was 100% each year, and 100% of employees who tested HIV-positive initiated ART within 6 mo, regardless of CD4 cell count or clinical status. We conservatively assumed no impact of this high-level ART coverage on HIV incidence since the intervention would cover only employees and not their sexual partners. In a second scenario (“family treatment”), we incorporated the extension of ART to those family members of employees who were eligible for ART, with an assumed average of one ART-eligible dependant per HIV-positive employee on ART. Ethics Approval The study was reviewed and approved by the following ethics committees: the London School of Hygiene & Tropical Medicine Ethics Committee (application number 962), the Anglogold Health Service Research Ethics Committee (AHS REC 004/02), and the University of KwaZulu-Natal Biomedical Research Ethics Committee (BE093/08). Employees’ consent to participation in this study was waived as we used only data that were collected for routine care purposes and, as in most other routine care settings, employees did not give written consent for this care. Data Availability The fully parameterised model that incorporates all data and that was used to produce all projections within this paper can be downloaded from OpenBU via http://hdl.handle.net/2144/10817. Results Patient-Level Cost and Resource Use and Absenteeism of Employees on and off ART The results of our bottom-up cost analyses in HIV-positive employees show that regardless of ART status, average annual outpatient and inpatient employee costs both increase with decreasing CD4 cell count, and, in contrast to analyses of the cost of public sector ART provision in South Africa [26–29], inpatient costs are higher than outpatient costs per patient-year (Table 6). Once employees initiate ART, these costs of care decrease dramatically across all CD4 cell count strata. However, when considering the healthcare cost of the HIV programme only, and excluding other HIV-related costs such as absenteeism and the cost of staff turnover, the addition of ART renders the HIV programme more expensive than without ART. HIV-positive employees not on ART have between 11 and 40 sick leave days annually over and above the average number of sick leave days in HIV-negative employees (Table 6). For specific CD4 strata, the level of absenteeism decreases by 16%–42% after ART initiation, except in employees with a CD4 cell count of <50 cells/mm3. As with healthcare costs, the most absenteeism is seen in the lowest CD4 cell count stratum, whether on or off ART. Coverage with Care, Survival in Employment, and HIV Prevalence Fig 2 shows the distribution of employees into types of care over the model projection period. While the proportion of untested HIV-positive employees falls with increasing HCT coverage, the proportion in wellness care first increases and then drops slightly as the proportion of employees on ART increases. From 2010, the proportion of employees in each type of care remains relatively stable, with newly tested HIV-positive employees moving quickly through wellness care and, if eligible, onto ART, and the proportion of employees on second-line ART slowly increasing. From 2012, only 35%–44% of HIV-positive employees are on ART, because many are not eligible for ART; however, 75%–97% of employees with CD4 cell count < 350 cells/mm3 are on ART. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. Distribution of HIV-positive employees into types of HIV care, 2003–2022 (ART scenario). https://doi.org/10.1371/journal.pmed.1001869.g002 Across all available model fits, projections suggest that an HIV-infected employee with a current CD4 cell count > 350 cells/mm3 will have a 39% (90% CrI 35%–43%), 57% (50%–62%), or 78% (73%–82%) probability of surviving the following 10 y if they are in no care, in wellness care, or on ART, respectively. (Note that this survival does not take into account deaths in employees once they have left the workforce.) However, survival in the workforce at 10 y is much lower, as a result of death as well as disability and other separations: 16% (90% CrI 13%–19%), 23% (20%–27%), and 35% (31%–39%) for employees in no care, in wellness care, and on ART, respectively. Without ART, these survival rates lead to a total of 22,274 (90% CrI 20,887–24,086) HIV-positive employee-years (or life-years in employment) at the mines between 2003 and 2022, with HIV prevalence increasing from 13.3% (90% CrI 12.8%–14.4%) in 2010 to 14.3% (13.0%–15.9%) in 2022. With ART coverage increasing from 10% of eligible HIV-positive employees in 2003 to 97% in 2020, the number of deaths amongst employees due to HIV over 20 y decreases by 16% (90% CrI 11%–21%) from 1,583 (90% CrI 1,406–1,791) without ART to 1,336 (1,183–1,497) with ART. Survival in employment increases by 8% (90% CrI 6%–12%) to 24,134 (90% CrI 22,848–25,841) HIV-positive life-years. This increase is not larger because on average only 34% of HIV-infected employees are on ART at any given time (since only a fraction of HIV-infected employees are eligible for ART), only a portion of these would have left the workforce or died in absence of ART over this period, and some leave the workforce before realising the full benefit of treatment. The increase in survival leads to an increase in HIV prevalence from 14.3% (90% CrI 13.0%–15.9%) in 2022 without ART to 16.3% (14.9%–17.8%) with ART. HIV prevalence is always higher in the lower job grades: 21% (90% CrI 19%–22%) in job grade 1 and 21% (18%–24%) in job grade 2 in 2022 with ART (Fig 3). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. Prevalence by job grade, 2003–2022, with workplace ART provision. Job grade 1: unskilled worker; grades 2 and 3: semi-skilled worker; grades 4 and 5: skilled worker; grade 6: management. https://doi.org/10.1371/journal.pmed.1001869.g003 Changes in Workforce Turnover, Absenteeism, and Separations With workplace ART provision, other changes are experienced by the workforce between 2003 and 2022. The total number of absent days due to HIV are estimated to be reduced by 8% (90% CrI 6%–10%), from 330,172 (90% CrI 297,729–367,723) to 303,897 (277,147–335,776) days, with 33% (90% CrI 26%–40%) fewer absenteeism days amongst employees with CD4 cell counts below 100 (Fig 4). The number of employees leaving employment for HIV-related reasons is estimated to decrease by 5% (90% CrI 3%–7%) to 3,626 (90% CrI 3,403–3,815) over 20 y, and the number of recruits is estimated to decrease by 2% (1%–3%) to 17,201 (16,454–17,912). Recruitment does not decrease further because of the large expansion of the company over this period (from 5,247 to 9,252 employees) and the considerable separations in the HIV-uninfected workforce. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 4. Total number of days absent due to HIV per CD4-cell-count-defined health state, 2003–2022. https://doi.org/10.1371/journal.pmed.1001869.g004 Total and Average Cost with and without ART Without workplace ART provision, the undiscounted total cost of HIV to the company (including all healthcare, absenteeism, and turnover costs) over 20 y is estimated at US$296 million (90% CrI US$274–US$320 million) (Table 7), with the mean annual cost estimated to increase from US$13 million (90% CrI US$12–US$15 million) in the first 10 y to US$15 million (US$14–US$16 million) over 20 y, mostly due to increasing HIV prevalence. This translates to a mean annual cost per HIV-positive employee of US$13,271 (90% CrI US$12,101–US$14,522) over 20 y. With ART, over 98% of model projections suggest that these costs decrease: the total and mean annual costs are estimated to decrease by 5% (90% CrI 2%–8%) over 20 y, and the mean annual cost per HIV-positive employee by 9% (5%–13%). These savings are estimated to accrue from the first year of the ART programme onwards and to increase as the average CD4 cell count of HIV-positive employees on ART rises. Similar changes are seen with the discounted cost (S1 Fig). Moreover, ART is estimated to be cost-saving at even the lowest coverage level, as each employee on ART saves absenteeism, healthcare, and turnover costs that are greater than the per employee cost of ART. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 7. Total cost of HIV to company with and without ART programme and cost savings due to ART—main results and sensitivity analysis. https://doi.org/10.1371/journal.pmed.1001869.t007 Average Cost and Savings by Item Without ART provision, the largest components of the mean undiscounted annual cost of HIV to the company over 20 y are estimated to be benefit payments (53% of mean annual cost) and medical care costs (24%), followed by absenteeism (15%), and training and recruitment (8%) (Table 8). The cost of medical care is dominated by inpatient care (78% of medical care costs). Once ART is introduced, we estimate that benefit payments and medical care costs remain the largest contributors to the annual HIV costs (46% and 21%, respectively), whereas the cost of the ART programme itself is estimated to be comparatively small, at just 7% of the total. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 8. Annual undiscounted cost and savings by cost item, 2003–2022. https://doi.org/10.1371/journal.pmed.1001869.t008 Overall, the average undiscounted annual savings from scaling up ART coverage over 20 y are estimated to be US$950,215 (90% CrI US$220,879–US$1,616,104). The largest contribution to these estimated savings (52% of total savings) is the 13% decrease in benefit payments, followed by the 15% decrease in medical care costs (27% of total savings) (Table 8). Although the cost of training and recruitment is estimated to fall by 15% with ART, this makes up only 9% of annual savings, whilst the cost of absenteeism, which falls by 11%, is estimated to contribute 12% of savings. Without ART, the total undiscounted annual cost of HIV to the company is estimated to make up 3.6% (90% CrI 3.3%–3.9%) of total company payroll between 2003 and 2022, whereas with ART, this falls to 3.4% (3.1%–3.7%). Sensitivity and Uncertainty Analysis and Additional Scenarios The univariate sensitivity analysis showed that total costs over 20 y are very sensitive to reductions in benefits paid for death and disability (−33%/66%) and changes in HIV incidence (±50%), as well as to using public sector data for CD4 cell count transition probabilities, reductions in absenteeism (−50%), and changes in inpatient cost (±50%) (Table 7). However, total costs do not change much if absenteeism by CD4 cell count category are assumed to be the same with and without ART or if the HIV-dependent separation rates (±20%) or outpatient costs (±50%) are changed. Equally, there is little change when ART costs from recent analyses of public sector ART provision are used [59]. Importantly, the only assumptions under which ART provision stops being cost-saving are if absenteeism is reduced by 50% (over both 10 and 20 y) or if no benefits are paid out (over 20 y only); under all other assumptions tested, ART still saves between 3% and 12% of total costs over 20 y. Finally, reducing HIV incidence as well as HIV prevalence in the starting population and recruits to low levels results in a much reduced HIV prevalence (1%) by 2022, representative of a low prevalence setting; in this scenario, the cost of HIV to the company reduces by 95% both without and with ART, with ART still saving 4% of costs. The overall findings of the probabilistic sensitivity analysis agreed with the findings of the univariate sensitivity analysis, despite the wide ranges assigned to many model parameters, with over 98% of all model fits predicting that ART provision was cost-saving (Table 7). The analysis also reinforced the relative contribution of individual cost items to total cost (Table 8). The analysis of co-variance revealed that 69% of the variability in the total savings achieved with ART in the probabilistic sensitivity analysis (after 20 y and undiscounted) were explained by uncertainty in the costs of ART (64%), as well as in the difference between the upwards CD4-cell-count-defined health-state transition probabilities on ART compared to with wellness care (21%) (see S2 and S3 Figs), and in the outpatient costs on ART (15%). Interestingly, although the cost of ART is always a relatively small component of the total cost of HIV (5%–11%), it can contribute significantly to offsetting the cost savings achieved with ART, with the cost of ART cancelling out 53% (90% CrI 32%–87%) of all potential savings. Importantly, the model projections suggest ART will always be cost-saving if it costs less than US$2,057 per patient-year. The large dependence of the estimated cost savings on the difference between the ART and wellness care health-state transition probabilities suggests that ART will not be cost-saving if it has little benefit for disease progression on top of what is already achieved with wellness care. The cost of HIV in the test and treat sensitivity scenario over 10 y (2013–2022) increases only marginally, by 0.2%, because of increased savings in terms of inpatient care, absenteeism, and benefit payments, which almost offsets the cost of the additional treatment occurring (Table 7). In the family treatment scenario, total cost with ART provision between 2013 and 2022 increases by 9%, but ART provision is still marginally cost-saving. Patient-Level Cost and Resource Use and Absenteeism of Employees on and off ART The results of our bottom-up cost analyses in HIV-positive employees show that regardless of ART status, average annual outpatient and inpatient employee costs both increase with decreasing CD4 cell count, and, in contrast to analyses of the cost of public sector ART provision in South Africa [26–29], inpatient costs are higher than outpatient costs per patient-year (Table 6). Once employees initiate ART, these costs of care decrease dramatically across all CD4 cell count strata. However, when considering the healthcare cost of the HIV programme only, and excluding other HIV-related costs such as absenteeism and the cost of staff turnover, the addition of ART renders the HIV programme more expensive than without ART. HIV-positive employees not on ART have between 11 and 40 sick leave days annually over and above the average number of sick leave days in HIV-negative employees (Table 6). For specific CD4 strata, the level of absenteeism decreases by 16%–42% after ART initiation, except in employees with a CD4 cell count of <50 cells/mm3. As with healthcare costs, the most absenteeism is seen in the lowest CD4 cell count stratum, whether on or off ART. Coverage with Care, Survival in Employment, and HIV Prevalence Fig 2 shows the distribution of employees into types of care over the model projection period. While the proportion of untested HIV-positive employees falls with increasing HCT coverage, the proportion in wellness care first increases and then drops slightly as the proportion of employees on ART increases. From 2010, the proportion of employees in each type of care remains relatively stable, with newly tested HIV-positive employees moving quickly through wellness care and, if eligible, onto ART, and the proportion of employees on second-line ART slowly increasing. From 2012, only 35%–44% of HIV-positive employees are on ART, because many are not eligible for ART; however, 75%–97% of employees with CD4 cell count < 350 cells/mm3 are on ART. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 2. Distribution of HIV-positive employees into types of HIV care, 2003–2022 (ART scenario). https://doi.org/10.1371/journal.pmed.1001869.g002 Across all available model fits, projections suggest that an HIV-infected employee with a current CD4 cell count > 350 cells/mm3 will have a 39% (90% CrI 35%–43%), 57% (50%–62%), or 78% (73%–82%) probability of surviving the following 10 y if they are in no care, in wellness care, or on ART, respectively. (Note that this survival does not take into account deaths in employees once they have left the workforce.) However, survival in the workforce at 10 y is much lower, as a result of death as well as disability and other separations: 16% (90% CrI 13%–19%), 23% (20%–27%), and 35% (31%–39%) for employees in no care, in wellness care, and on ART, respectively. Without ART, these survival rates lead to a total of 22,274 (90% CrI 20,887–24,086) HIV-positive employee-years (or life-years in employment) at the mines between 2003 and 2022, with HIV prevalence increasing from 13.3% (90% CrI 12.8%–14.4%) in 2010 to 14.3% (13.0%–15.9%) in 2022. With ART coverage increasing from 10% of eligible HIV-positive employees in 2003 to 97% in 2020, the number of deaths amongst employees due to HIV over 20 y decreases by 16% (90% CrI 11%–21%) from 1,583 (90% CrI 1,406–1,791) without ART to 1,336 (1,183–1,497) with ART. Survival in employment increases by 8% (90% CrI 6%–12%) to 24,134 (90% CrI 22,848–25,841) HIV-positive life-years. This increase is not larger because on average only 34% of HIV-infected employees are on ART at any given time (since only a fraction of HIV-infected employees are eligible for ART), only a portion of these would have left the workforce or died in absence of ART over this period, and some leave the workforce before realising the full benefit of treatment. The increase in survival leads to an increase in HIV prevalence from 14.3% (90% CrI 13.0%–15.9%) in 2022 without ART to 16.3% (14.9%–17.8%) with ART. HIV prevalence is always higher in the lower job grades: 21% (90% CrI 19%–22%) in job grade 1 and 21% (18%–24%) in job grade 2 in 2022 with ART (Fig 3). Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 3. Prevalence by job grade, 2003–2022, with workplace ART provision. Job grade 1: unskilled worker; grades 2 and 3: semi-skilled worker; grades 4 and 5: skilled worker; grade 6: management. https://doi.org/10.1371/journal.pmed.1001869.g003 Changes in Workforce Turnover, Absenteeism, and Separations With workplace ART provision, other changes are experienced by the workforce between 2003 and 2022. The total number of absent days due to HIV are estimated to be reduced by 8% (90% CrI 6%–10%), from 330,172 (90% CrI 297,729–367,723) to 303,897 (277,147–335,776) days, with 33% (90% CrI 26%–40%) fewer absenteeism days amongst employees with CD4 cell counts below 100 (Fig 4). The number of employees leaving employment for HIV-related reasons is estimated to decrease by 5% (90% CrI 3%–7%) to 3,626 (90% CrI 3,403–3,815) over 20 y, and the number of recruits is estimated to decrease by 2% (1%–3%) to 17,201 (16,454–17,912). Recruitment does not decrease further because of the large expansion of the company over this period (from 5,247 to 9,252 employees) and the considerable separations in the HIV-uninfected workforce. Download: PPT PowerPoint slide PNG larger image TIFF original image Fig 4. Total number of days absent due to HIV per CD4-cell-count-defined health state, 2003–2022. https://doi.org/10.1371/journal.pmed.1001869.g004 Total and Average Cost with and without ART Without workplace ART provision, the undiscounted total cost of HIV to the company (including all healthcare, absenteeism, and turnover costs) over 20 y is estimated at US$296 million (90% CrI US$274–US$320 million) (Table 7), with the mean annual cost estimated to increase from US$13 million (90% CrI US$12–US$15 million) in the first 10 y to US$15 million (US$14–US$16 million) over 20 y, mostly due to increasing HIV prevalence. This translates to a mean annual cost per HIV-positive employee of US$13,271 (90% CrI US$12,101–US$14,522) over 20 y. With ART, over 98% of model projections suggest that these costs decrease: the total and mean annual costs are estimated to decrease by 5% (90% CrI 2%–8%) over 20 y, and the mean annual cost per HIV-positive employee by 9% (5%–13%). These savings are estimated to accrue from the first year of the ART programme onwards and to increase as the average CD4 cell count of HIV-positive employees on ART rises. Similar changes are seen with the discounted cost (S1 Fig). Moreover, ART is estimated to be cost-saving at even the lowest coverage level, as each employee on ART saves absenteeism, healthcare, and turnover costs that are greater than the per employee cost of ART. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 7. Total cost of HIV to company with and without ART programme and cost savings due to ART—main results and sensitivity analysis. https://doi.org/10.1371/journal.pmed.1001869.t007 Average Cost and Savings by Item Without ART provision, the largest components of the mean undiscounted annual cost of HIV to the company over 20 y are estimated to be benefit payments (53% of mean annual cost) and medical care costs (24%), followed by absenteeism (15%), and training and recruitment (8%) (Table 8). The cost of medical care is dominated by inpatient care (78% of medical care costs). Once ART is introduced, we estimate that benefit payments and medical care costs remain the largest contributors to the annual HIV costs (46% and 21%, respectively), whereas the cost of the ART programme itself is estimated to be comparatively small, at just 7% of the total. Download: PPT PowerPoint slide PNG larger image TIFF original image Table 8. Annual undiscounted cost and savings by cost item, 2003–2022. https://doi.org/10.1371/journal.pmed.1001869.t008 Overall, the average undiscounted annual savings from scaling up ART coverage over 20 y are estimated to be US$950,215 (90% CrI US$220,879–US$1,616,104). The largest contribution to these estimated savings (52% of total savings) is the 13% decrease in benefit payments, followed by the 15% decrease in medical care costs (27% of total savings) (Table 8). Although the cost of training and recruitment is estimated to fall by 15% with ART, this makes up only 9% of annual savings, whilst the cost of absenteeism, which falls by 11%, is estimated to contribute 12% of savings. Without ART, the total undiscounted annual cost of HIV to the company is estimated to make up 3.6% (90% CrI 3.3%–3.9%) of total company payroll between 2003 and 2022, whereas with ART, this falls to 3.4% (3.1%–3.7%). Sensitivity and Uncertainty Analysis and Additional Scenarios The univariate sensitivity analysis showed that total costs over 20 y are very sensitive to reductions in benefits paid for death and disability (−33%/66%) and changes in HIV incidence (±50%), as well as to using public sector data for CD4 cell count transition probabilities, reductions in absenteeism (−50%), and changes in inpatient cost (±50%) (Table 7). However, total costs do not change much if absenteeism by CD4 cell count category are assumed to be the same with and without ART or if the HIV-dependent separation rates (±20%) or outpatient costs (±50%) are changed. Equally, there is little change when ART costs from recent analyses of public sector ART provision are used [59]. Importantly, the only assumptions under which ART provision stops being cost-saving are if absenteeism is reduced by 50% (over both 10 and 20 y) or if no benefits are paid out (over 20 y only); under all other assumptions tested, ART still saves between 3% and 12% of total costs over 20 y. Finally, reducing HIV incidence as well as HIV prevalence in the starting population and recruits to low levels results in a much reduced HIV prevalence (1%) by 2022, representative of a low prevalence setting; in this scenario, the cost of HIV to the company reduces by 95% both without and with ART, with ART still saving 4% of costs. The overall findings of the probabilistic sensitivity analysis agreed with the findings of the univariate sensitivity analysis, despite the wide ranges assigned to many model parameters, with over 98% of all model fits predicting that ART provision was cost-saving (Table 7). The analysis also reinforced the relative contribution of individual cost items to total cost (Table 8). The analysis of co-variance revealed that 69% of the variability in the total savings achieved with ART in the probabilistic sensitivity analysis (after 20 y and undiscounted) were explained by uncertainty in the costs of ART (64%), as well as in the difference between the upwards CD4-cell-count-defined health-state transition probabilities on ART compared to with wellness care (21%) (see S2 and S3 Figs), and in the outpatient costs on ART (15%). Interestingly, although the cost of ART is always a relatively small component of the total cost of HIV (5%–11%), it can contribute significantly to offsetting the cost savings achieved with ART, with the cost of ART cancelling out 53% (90% CrI 32%–87%) of all potential savings. Importantly, the model projections suggest ART will always be cost-saving if it costs less than US$2,057 per patient-year. The large dependence of the estimated cost savings on the difference between the ART and wellness care health-state transition probabilities suggests that ART will not be cost-saving if it has little benefit for disease progression on top of what is already achieved with wellness care. The cost of HIV in the test and treat sensitivity scenario over 10 y (2013–2022) increases only marginally, by 0.2%, because of increased savings in terms of inpatient care, absenteeism, and benefit payments, which almost offsets the cost of the additional treatment occurring (Table 7). In the family treatment scenario, total cost with ART provision between 2013 and 2022 increases by 9%, but ART provision is still marginally cost-saving. Discussion Using a dynamic health-state transition model, we conducted a cost–benefit analysis of an established ART programme operating in a number of coal mines in South Africa. Our analysis provides both a retrospective analysis of the programme between 2003 and 2010 and a projection of future developments based on the results of this retrospective analysis. When considering the impact of HIV on a company’s healthcare costs—as well as worker absenteeism, sickness and death benefits, and staff turnover—the introduction of ART to all eligible employees is cost-saving from the first year of the programme onwards. With ART provision, the total costs of HIV to the company over 20 y is estimated to be reduced by 6% (90% CrI 2%–11%), and the cost per HIV-positive employee is estimated to be reduced by 14% (7%–19%). Moreover, in our probabilistic sensitivity analysis, 98% of the 998 model fits (selected from amongst 20,000 model runs) confirm this cost savings. The biggest savings are due to reductions in the benefit payments for death and ill-health retirement, followed by a decrease in the cost of employee healthcare use. This finding that ART is cost-saving is robust to the uncertainty around the model parameters as well as to other changes in numerous parameters or assumptions, including if absenteeism is the same for employees on and off ART, if there are large reductions in benefit payments, and if HIV prevalence in the workforce is decreased to below 1%. The only instance where ART does not save costs over 20 y is if absenteeism in HIV-positive employees is reduced by 50% or if no benefits are paid out—though the latter strategy still saves costs over 10 y. In addition, a strategy of offering HIV testing to all employees and immediate ART to all HIV-positive employees also results in savings to the cost of the HIV programme, suggesting test and treat be recommended as a powerful intervention for companies trying to preserve their employees’ productivity. Offering ART to one family member for each HIV-positive employee, a generous assumption, reduces savings but is still cost-saving compared to no workplace ART provision. Previous work has shown a heterogeneous impact of HIV on absenteeism and replacement cost. In a study of nearly a thousand firms operating in Africa in 1997, the impact of HIV on staff turnover was minimal, probably because of the lower HIV prevalence at that time, with difficulties in replacing professional staff being the most significant problem companies were facing [60]. In another study, the total cost per HIV infection to South African companies was estimated at US$2,094 to US$15,000 for an unskilled worker (in 2001 prices) and US$8,736 to US$65,000 for a manager [2]. A study of a Natal sugar mill found that on average 28 d were lost in each of the 2 y preceding retirement on grounds of ill-health and estimated that the cost of each HIV infection was roughly three times the employee’s annual salary per year [61]. Similarly, a large part of the savings in our analysis were due to a policy of benefits being paid to the employee or their family in the case of disability or death, which might not apply to other workplaces and might limit the generalisability of the results across workplaces and countries. While our analysis adds to the body of knowledge on the economic impact of HIV and ART—through the use of detailed modelling incorporating a wealth of data on costs of HIV and ART outcomes from the same setting—our study nonetheless has limitations. First, it was limited to the direct cost of HIV to companies. In a previous study, the life insurer Metropolitan predicted that the indirect costs of HIV to business (including costs due to a loss in morale, legal costs, management costs, and labour consultation costs) could add up to 15% of the wage and salary budget by 2010 [62]. The provision of ART could improve morale and retention of skilled employees [5] as well as help safeguard the company’s license to mine [63]. Including this added indirect benefit of ART would have increased our savings from workplace ART provision. Second, we used an average drug cost for first-line and second-line ART that slightly underestimated the cost of ART in the later years of the projection, when more employees needed second-line ART, and did not stratify ART cost by time on treatment. However, since few HIV-positive employees were on second-line treatment throughout the projection period and the cost of ART was a small proportion of total costs, this underestimation is unlikely to change our findings. Third, data for some of the model inputs, such as transitions between certain CD4-cell-count-defined health states, was limited, resulting in uncertainty around some estimates. The effect of this uncertainty was included in our model projections as well as tested in our sensitivity analysis, and our results were found to be robust to changes along plausible ranges for these parameters. However, the deterministic nature of the model prevented it from capturing the full inherent variability present in this workforce. Lastly, we did not examine the impact of HIV prevention policies on the miners or their families. Further work could involve evaluating the effects of prevention and treatment interventions on HIV incidence, including in the areas around the mines and in miners’ families, and the cost of new policies such as providing pre-exposure prophylaxis or increasing the accommodation of miners’ families in the vicinity of the mines, in compliance with the mining charter [63]. Finally, given our finding of the importance of the cost of ART in influencing cost savings, further reductions in the private sector cost of antiretroviral drugs remain crucial. Conclusion Providing HIV care, including ART, in a workforce with high HIV prevalence and high resulting absenteeism and turnover can be cost-saving for the employer, with savings being greater at higher ART coverage, and might provide respite to the strained resources of large-scale public sector programmes. Beyond making good business sense, a company-level HIV care programme including ART could go a long way towards improving the strained labour relations in the South African mining sector, especially when improved access to healthcare extends to the entire community [64]. It is crucial that strategies such as those under study here are replicated in other companies in similar settings. Conclusion Providing HIV care, including ART, in a workforce with high HIV prevalence and high resulting absenteeism and turnover can be cost-saving for the employer, with savings being greater at higher ART coverage, and might provide respite to the strained resources of large-scale public sector programmes. Beyond making good business sense, a company-level HIV care programme including ART could go a long way towards improving the strained labour relations in the South African mining sector, especially when improved access to healthcare extends to the entire community [64]. It is crucial that strategies such as those under study here are replicated in other companies in similar settings. Supporting Information S1 Fig. Total annual cost with and without ART (discounted and undiscounted), 2003–2022 (2010 USD). https://doi.org/10.1371/journal.pmed.1001869.s001 (TIF) S2 Fig. Results of analysis of co-variance: yearly cost of ART. https://doi.org/10.1371/journal.pmed.1001869.s002 (TIF) S3 Fig. Results of analysis of co-variance: difference between wellness care and ART transition probabilities. https://doi.org/10.1371/journal.pmed.1001869.s003 (TIF) S1 Text. Details on parameter estimation, probabilistic sensitivity analysis, and model calculations. https://doi.org/10.1371/journal.pmed.1001869.s004 (PDF) Acknowledgments The authors are grateful to Salome Charalambous for her contribution to the clinical programme at the workplace, to Craig Innes for facilitating access to clinical datasets, to Thia Grobler, Sue Ingle, Margaret May, and Amy Huber for providing additional model inputs, and to the staff at the mines included in this study. This paper is dedicated to Wim Richter.