Background Prevention of persistent pain following breast cancer surgery, via early identification of patients at high risk, is a clinical need. Supervised machine-learning was used to identify parameters that predict persistence of significant pain. Methods Over 500 demographic, clinical and psychological parameters were acquired up to 6 months after surgery from 1,000 women (aged 28–75 years) who were treated for breast cancer. Pain was assessed using an 11-point numerical rating scale before surgery and at months 1, 6, 12, 24, and 36. The ratings at months 12, 24, and 36 were used to allocate patents to either “persisting pain” or “non-persisting pain” groups. Unsupervised machine learning was applied to map the parameters to these diagnoses. Results A symbolic rule-based classifier tool was created that comprised 21 single or aggregated parameters, including demographic features, psychological and pain-related parameters, forming a questionnaire with “yes/no” items (decision rules). If at least 10 of the 21 rules applied, persisting pain was predicted at a cross-validated accuracy of 86% and a negative predictive value of approximately 95%. Conclusions The present machine-learned analysis showed that, even with a large set of parameters acquired from a large cohort, early identification of these patients is only partly successful. This indicates that more parameters are needed for accurate prediction of persisting pain. However, with the current parameters it is possible, with a certainty of almost 95%, to exclude the possibility of persistent pain developing in a woman being treated for breast cancer. Keywords Pain · Bioinformatics · Data science · Chronification Introduction Breast cancer is the most common cancer in women in the developed countries and its prevalence is steadily increas- ing [3, 15, 17]. Improved management of breast cancer has Eija Kalso and Alfred Ultsch have equally contributed to this work. made survivorship issues important , including moder- ate to severe pain reported with a prevalence of about 15% Electronic supplementary material The online version of this at 1 year from surgery . About 34% of these patients article (https ://doi.org/10.1007/s1054 9-018-4841-8) contains supplementary material, which is available to authorized users. * Jörn Lötsch Pain Clinic, Department of Anaesthesiology, Intensive email@example.com Care and Pain Medicine, Helsinki University Hospital and University of Helsinki, Helsinki, Finland Institute of Clinical Pharmacology, Goethe - University, Breast Surgery Unit, Comprehensive Cancer Center, Helsinki Theodor – Stern - Kai 7, 60590 Frankfurt am Main, University Hospital and University of Helsinki, Helsinki, Germany Finland Fraunhofer Institute for Molecular Biology and Applied DataBionics Research Group, University of Marburg, Hans – Ecology IME, Project Group Translational Medicine Meerwein - Straße, 35032 Marburg, Germany and Pharmacology TMP, Theodor – Stern - Kai 7, 60596 Frankfurt am Main, Germany Vol.:(0123456789) 1 3 400 Breast Cancer Research and Treatment (2018) 171:399–411 have signs of neuropathic pain  and the pain can last for treated at the Helsinki University Hospital between 2006 several years , significantly impairing their quality of and 2010 with breast-conserving surgery or mastectomy, life . Prevention of persistent post-surgical pain requires and sentinel node biopsy and/or axillary clearance. Exclu- early identification of patients at the highest risk to initiate sion criteria were neoadjuvant therapy, i.e., administration of appropriate medical and psychosocial interventions [10, 43]. chemotherapy to shrink the tumor before the main surgical On the other hand, identifying patients in whom the possibil- treatment , and immediate breast reconstruction surgery. ity of persistent pain can be dismissed with high confidence Of the 1,536 consecutive eligible patients, 1,149 patients may be similarly desirable as this could prevent unnecessary were invited to participate, of whom 126 patients declined therapeutic interventions. and 23 were withdrawn. The prediction of persistent pain after a surgical interven- The cohort has previously been described in detail [24, tion is therefore an active research topic that has already led 37]. Briefly, perioperative analgesia was standardized, con- to several proposed solutions, as indicated by 160 hits in sisting of oral acetaminophen and intravenous oxycodone, a PubMed search at https ://www.ncbi.nlm.nih.gov/pubme titrated first by a research nurse in the post-anesthesia care d for “(chronic or persistent) and pain and prediction and unit followed by patient-controlled analgesia on the ward. *surgery” on December 6, 2017. However, diagnostic tools No regional anesthesia was used. Analgesia at home during providing such a prediction have remained an unmet clinical the first postoperative week consisted of ibuprofen, aceta- need. Considering that pain has a highly complex patho- minophen or a combination of acetaminophen and codeine. physiology [22, 47] and is triggered by several different Adjuvant treatments were given according to international causes including cancer  and surgery , such tools guidelines . may require a combination of different parameters. However, in addition to proposals of biochemical or genetic markers, Data acquisition clinical and psychological parameters have remained a basis of predictive assessments of the development of pain after Acquisition of data on pain during the follow-up surgery [4, 11, 30, 37]. In the present analysis, more than 500 parameters were A full listing of the acquired data is provided in Supplemen- available from a 3-year follow-up of 1,000 women operated tary Table 1. Pain was assessed using a standard 11-point on for breast cancer . This provided a robust basis for Numerical Rating Scale (NRS) ranging from 0 to 10, 0 indi- clinical judgment of pain persistence and, using parameters cating “no pain” and 10 the “worst imaginable pain” . acquired during the perioperative period, allowed the pre- The mean daily post-surgical pain intensity was recorded diction of either persistent pain at later stages or its unlike- during the first postoperative week with patient-rated paper liness. Previous analyses indicated the suitability of these diaries. At 1, 6, 12, 24, and 36 months after surgery, ques- parameters for the prediction of persistent pain [32, 49]; tionnaires were posted to all patients asking about the pres- however, the present analysis used a data-driven approach, ence, location, and intensity of pain, pain interference, and without prior hypotheses on the most important parameters, mood. The patients reported, as a single rating, the worst to create a predictive diagnostic tool for persistent pain, or pain, either at rest or during movement, in any of the sur- its absence, following breast cancer surgery and adjuvant gery-related locations (breast, axilla, upper arm) during the therapies. previous week. The questionnaires about pain interference with daily life and sleep were developed for this study. The presence or absence of persistent pain was established from Methods these questions at 12–36 months. Patients Acquisition of candidate parameters for prediction of pain persistence The study followed the Declaration of Helsinki, and both the Coordinating Ethics Committee (journal number 136/ To predict the development or absence of persistent pain E6/2006) and the Ethics Committee of the Department of after breast cancer surgery, the so-called “input space” of Surgery (148/E6/05) of the Hospital District of Helsinki and 542 different variables was acquired from the data collected Uusimaa approved the study protocol. Written informed con- before surgery, during the perioperative phase, and at follow- sent for data acquisition and publication in an anonymized up until 6 months after surgery. The acquired variables, in manner was obtained from each participating patient. One the present context of machine-learning  called param- thousand women with unilateral non-metastasized breast eters or “features,” included the patients’ medical history cancer, aged 28–75 years (Fig. 1) were enrolled during the (diseases, medications, number of previous surgeries), and preoperative visit before breast cancer surgery. They were demographic data. In addition, pain-related data included 1 3 Breast Cancer Research and Treatment (2018) 171:399–411 401 Fig. 1 Flow chart showing the classification of the patients 1,000 paents on the basis of the 3-year development of pain following breast cancer surgery. A total of 853 women fell into the two main groups of persisting or NRS ≥ 4 and non-persisting pain, accord- month36 NRS < 4and month36 ing to the criteria displayed in NRS > 0and month12..month36 NRS < 4 the gray-shaded frames. This month12..month36 (NRSmonth − NRS >0 36 month24) was the main cohort that was analyzed. The remaining 143 women in whom the criteria for Other Non- class assignment applied only Persistent 74 779 persistent 147 me courses partly were therefore excluded pain pain of pain from machine-learned classifier establishment but they were used as an exploratory short- Excluded ened “test” data set. Incomplete (reconstrucon Data returns of pain questionnaires surgery within 1 4 4 incomplete month before pain were dealt with by imputa- query) tion as detailed in the methods section Other Non- Persistent 70 779 persistent 143 me courses pain pain of pain Main data setExploratory (shortened) “test” data set the presence of persistent pain of any kind, preoperative pain All data were entered manually into Excel files and double- in the operated area (breast, axilla, arm), opioid (remifen- checked by two investigators. tanil) consumption during surgery, immediate postopera- tive pain intensity ratings (at rest and during movement), Data analysis amount of oxycodone needed for satisfactory pain relief for the first time after surgery, oxycodone consumption during Data analysis was performed using the MATLAB numeri- 20 h after surgery, pain intensity, and analgesic consump- cal computing environment (version 220.127.116.112, MathWorks, tion during the first postoperative week. The parameters also Natick, MS, USA) and R software package (version 3.2.3 for included surgical data such as type of surgery (mastectomy Linux; http://CRAN.R-proje ct.org/ ). or breast-conserving surgery, axillary clearance, sentinel Application of the previously proposed six-factor risk node biopsy), complications of surgery and re-operations, index for persisting pain was performed according to the pathology data such as tumor and lymph node character- instructions in Table A1 in the appendix of . In brief, istics. Finally, data on adjuvant therapies (chemotherapy, the six items “age,” “chronic pain of any kind,” “number of radiation therapy, endocrine therapy) were acquired. previous operations,” “body mass index,” “preoperative pain In the follow-up questionnaires, detailed pain-related in the area to be operated on,” and “smoking” were weighted parameters about specific pain locations were sought and according to these instructions. For example, age ≤ 39 years the patients were asked whether the pain had disturbed was given a weight of 0, age 40–69 years a weight of 8, and their sleep or otherwise affected their life (11-points NRS age ≥ 70 years a weight of 16. The respective weights for ranging from 0 to 10, 0 indicating “not at all” and 10 “very the six parameters were added from an index of 20, patients much”), and what analgesics they used, if any. Psychologi- were predicted as potentially developing persisting pain. cal data were acquired with questionnaires including Beck’s Given the broader data basis available in the present Depression Inventory (BDI)  and Spielberger’s State-Trait analysis, a novel and independently designed analysis was Anxiety Inventory (STAI)  before surgery and at the performed to assess whether the classification performance follow-up times. The State-Trait Anger Expression Inventory could be improved. This was approached using supervised (STAXI2)  was used preoperatively and at 6 months. machine-learning  and feature-selection techniques 1 3 402 Breast Cancer Research and Treatment (2018) 171:399–411 , aimed at identifying parameters from the data acquired of the so-called output space, (i) feature pre-selection, (iii) before surgery and up to the sixth month after surgery, which feature selection, and (iv) classifier building, which are could predict the presence or absence of persisting pain in described briefly in the following; more detailed descrip- the area operated on during the 3-year follow-up available tions are provided in the Supplementary materials. for the present analysis. However, as before  an interpret- First, the so-called “output space” was established (Fig. 2 able and clinically immediately applicable diagnostic tool left part) by identifying the groups or “classes” of patients was desired and therefore, a rule-based classifier [ 59] was with respect to persistent pain. In the following, the term again chosen forming a questionnaire with “yes/no” items “persisting pain group” and “non-persisting pain group” (decision rules). The analysis was performed in four main or “classes” will be used when referring to these groups. steps as shown in Fig. 2 and comprising (i) establishment Subsequently, the input or feature space was analyzed to Output space Input (feature) space (pain chroniﬁcaon classes) (queried paent parameters) 3NRS queries 822 Parameters (features) (Month 12, 24, 36) Chroniﬁerclass 52 Interval 434 Categorial 336 Binary deﬁnion scaled Non-chronifying Other me Aggregaon / Chronifyingpain pain courses of pain combinaon Explorave Main data set shortened „test“ 21 Categorial data set Bayes decision Kullback-Leibler Accuracy, OR limit divergence 17 Interval 15 Categorial 0 Binary scaled ABC analysis ABC analysis 6 Interval scaled 15 Categorial Brute force combinaon tesng for maximum sensivity * speciﬁcity 21 Features for “predicve tool” Fig. 2 Flow chart of the data analysis. The figure provides an over - qualifying as component s of a diagnostic tool respectively classi- view on the applied machine-learning approach in four steps (indi- fier was stepwise reduced (initially 542, finally 21), forwarding to the cated in blue: output space preparation, input space feature pre-selec- next analytical step only those features that had passed the criteria of tion, feature selection and classifier building, including validation). the actual selection procedure. The Bayesian decision limit and Kull- The white frames show the variable flow; the gray frames depict the back–Leibler divergence refer to the respective standard procedure bioinformatics operation applied on the variables. During feature pre- presented elsewhere [28, 35] selection and feature selection, the number of candidate variables 1 3 Classiﬁer building Feature selecon Feature preselecon Data preprocessing Original data Breast Cancer Research and Treatment (2018) 171:399–411 403 identify those parameters that provided a valid assignment establishment (Fig. 1 right) but were used as an exploratory of a patient to the classes of persisting or non-persisting “test” data set. pain. This included the second analytical (Fig. 2 right part) step comprising a “feature” pre-selection, during which Classifier establishment all parameters were analyzed with respect to differences between the “persisting pain” and “non-persisting pain” Supervised machine-learning was applied to map groups, established in the output space. This was based on the “input space,” i.e., the acquired “features” (the the assessment of effect sizes between the two patient groups 542 parameters), x, to the “classes,” i.e., the out- depending on the numerical scaling of the parameters as puts, y, given the data subset of the input–output pairs continuous (interval-scaled), discrete or binary (“yes”/”no”). D = x , y x ∈ X, y ∈ Y, i = 1… n that comprised the i i i i Those parameters that had passed that step were taken n = 849 patients belonging to the two groups (persisting or into the third analytical step of feature selection (Fig. 2 non-persisting pain). Feature pre-selection (data analysis right part), which eliminated all parameters distinguishing step 2; Fig. 2 right part) identified 39 parameters for which between the persisting pain and non-persisting pain groups the analysis of probability of belonging to the “persisting but not providing additional relevant information justifying pain” group, i.e., p y = y x, D , was in principle possible. their inclusion in a tool (“classifier”) predictive of persist- During feature selection (data analysis step 3; Fig. 2 right ing pain. This applied a so-called calculated ABC analysis part), among these 39 candidates, ABC analysis picked . This is an inventory categorization technique originally 21 parameters (features) six continuous and 15 discrete developed for problems in economics to search for a subset variables (Fig. 2) that provided a substantial contribu- with the minimum possible effort that gives the maximum tion to the sensitivity and specificity of a classifier for a yield [23, 42]. The remaining parameters, in the following patient’s assignment to one of the groups (persisting pain called “features,” were taken into the fourth step of the data or non-persisting pain). The process of feature selection analysis (Fig. 2 bottom right) in which the predictive tool via 1,000 resampling and ABC analyses is shown in Fig. 3 or “classifier” was constructed and assessed with respect for the 17 continuous variables, of which only six passed to standard test performance measures. To obtain a robust the analytical step due to their consistent placement in set selection of parameters (features), Monte-Carlo resampling “A” comprising the parameters most suitable as predictors. was used repeatedly. A rule-based classifier  was obtained during analyti- cal step 4 (Fig. 2 bottom right). It took the shape of a diag- nostic tool shown in Table 2 consisting of a set of items Results presented as “yes” or “no” decisions. The diagnostic tool comprised three main categories of variables, i.e., demo- Persisting versus non‑persisting pain groups graphic, psychological, and pain-related parameters (fea- tures). For each item, the prediction of persistent pain, i.e., The recovery rate of the pain questionnaires was high with the assignment of a patient to the “persisting pain” group 95.3, 91.3, 90.2, 90.2, and 87.4% returned in months 1, 6, (output space), was likely at a specific threshold shown in 12, 24, and 36, respectively. This sufficed for a valid classi- the right column of Table 2. The number of positive deci- fication of the subjects into two main diagnostic groups that sions by which a patient was classified into the “persist- displayed temporal courses of pain for 3 years after breast ing pain” group was established testing all possible feature cancer surgery obtained in the first analytical step aimed at combinations. This analysis resulted in a threshold of 9.5 establishing the so-called output step (Fig. 2 left). Of the (Fig. 4), which indicates that when at least 10 of the 21 rules 1000 analyzed women (Fig. 1), 779 had a favorable time listed in Table 2 applied the patient was identified as likely course of postoperative pain and belonged to the “non-per- to develop persistent pain. The detailed use of the diagnostic sisting pain” group characterized by NRS ≤ 3 tool is described step-by-step in Textbox 1. month12...month36 (Fig. 1 left). By contrast, in 74 women, the pain followed the opposite path typical for the “persisting pain” group characterized by NRS ≥ 4 and NRS > 0 Textbox 1: Summary of the data acquisition month36 month12...month36 and (NRS − NRS ) ≥ 0 (Fig. 1 middle). Four of and calculation steps required to apply month36 month24 these women were excluded from the analysis due to breast the diagnostic tool, as presented in Table 2, reconstruction surgery within the previous month, which to a single patient. See Table 1 for details about the items that need to be averaged hampered the clear association of pain at month 36 to the original surgery. Finally, the criteria for class assignment were not completely met in the remaining 143 women who Observation period for application of the predictive tool were therefore excluded from machine-learned classifier defined in Table 2 is 6 months after breast cancer surgery. 1 3 404 Breast Cancer Research and Treatment (2018) 171:399–411 Fig. 3 Performance of the continuous variables with a Bayesian deci- BDI1 = BDI at 1 month after surgery, BDI2 = BDI at 6 months after sion boundary in 1,000 repeated cross-validations. The n = 17 contin- surgery, STAI0A, STAI1A, STAI2A = State anxiety (STAI) aquired uous variables were subjected to an ABC analysis (for ABC analysis, preoperatively and at 1 month and 6 months after surgery, respec- see ). The set A (best performers) was characterized by a sensitiv- tively, STAI0B, STAI1B, STAI2B = Trait anxiety (STAI) aquired ity · specificity > 40% (threshold; magenta line). The resulting 6 vari- preoperatively and at 1 and 6 months postoperatively, respectively, ables in set A were included in the classifier construction. Names of STAXI = Anger inhibition (STAXI) variables above the threshold: Age, BMI, BDI0 = preoperative BDI, • Query of demographic factors (age, BMI; items 1 and 2 are queried at week 1, 4, and 24 at various body in Table 1). locations and averaged for week 4 across all body • Query of complex psychological questionnaires (STAXI, locations and for each of six body locations (breast, preoperative), BDI, STAI state, and STAI trait at week axilla, upper arm, joints, lower arm, hand/fingers) 24 postoperative (items 3–6 in Table 1). averaged across weeks 1, 4 and 24. Query of pain-related ratings addressing how much the – How much pain has affected the patients sleep is pain has affected the patient’s life or sleep during the queried using an 11-point NRS ranging from 0, not last week, or querying the pain intensity, globally and at all, to 10, very much (items 13–16 in Table 1). also separately for various different body locations (items Items are queried at week 4 and 24, globally (items 7–21 in Table 1). 13 and 14 in Table 1), or asked specifically for the effect of pain at body locations breast or axilla. – How much pain has affected the patient’s life is que- Items are queried at week 4 and 24 for the two ried using an 11-point NRS ranging from 0, not at body locations are averaged per body location all, to 10, very much (items 7–12 in Table 1). Items (items 15 and 16 in Table 1). 1 3 Breast Cancer Research and Treatment (2018) 171:399–411 405 1 3 Table 1 Parameter list including original features along with the details of feature aggregation Number Parameters (features) Time Body location Preoperative Week 1 Week 4 Week 24 Breast Axilla Upper arm Joints Lower arm Hand/fingers Operated Somewhere side arm 1 Age X 2 BMI X 3 Depressive symptoms (BDI) X 4 State anxiety (STAI) X 5 Trait anxiety (STAI) X 6 Anger inhibition (STAXI-2) X 7 Have the pains in the extremities* affected your X X X X X X life? 8 Have the pains in the axilla affected your life? X X X X 9 Have the pains in the hand or fingers affected your X X X X life? 10 Have the pains in the joints affected your life? X X X X 11 Have the pains in the lower arm affected your life? X X X X 12 Have the pains in the upper arm affected your life? X X X X 13 How much has the pain disturbed your sleep? X 14 How much has the pain disturbed your sleep? X 15 How much has the pain in the axilla disturbed X X X your sleep? 16 How much has the pain in the breast disturbed X X X your sleep? 17 Pain intensity in the operated-side arm and axilla X X X X in the morning? 18 Worst pain intensity during the past week at one X X X X X X X month 19 Worst pain intensity during the past week at 6 X X X X X X X months 20 Worst pain intensity during the past week in the X X X operated breast? 21 Worst pain intensity during the past week some- X X X where? The table specifies the time points and, if applicable, the targeted body locations for several questions the patients are asked following surgery. Specifically, for the pain-related parameters #7 - #21 (according to the numbering in the left column), patients were asked about how the pain affected their lives or sleep. These questions were asked at several times (week 1, 4, and 24 after surgery) and for several specific body locations (breast, axilla, upper arm, joints, lower arm, hand, arm on operation side, elsewhere) indicated in the middle block of the table. The “X”s indicate which ratings were averaged to obtain the parameters *Extremities: aggregated parameters across axilla, upper arm, lower arm, hand/fingers, and the associated joints 406 Breast Cancer Research and Treatment (2018) 171:399–411 Table 2 Parameters (predictive factors) for persisting pain following breast cancer surgery Number Category Parameters (predictive factors) Threshold 1 Demographic factors Age > 62 2 BMI > 31.5 3 Psychological factors Depressive symptoms (BDI) > 11 4 State anxiety (STAI) > 35 5 Trait anxiety (STAI) > 37 6 Anger inhibition (STAXI) > 12 7 Pain-related factors Have the pains in the extremities *affected your life? > 1.5 8 Have the pains in the axilla affected your life? > 0.5 9 Have the pains in the hand or fingers affected your life? > 0.5 10 Have the pains in the joints affected your life? > 0.5 11 Have the pains in the lower arm affected your life? > 0.5 12 Have the pains in the upper arm affected your life? > 0.5 13 How much has the pain disturbed your sleep? > 0.5 14 How much has the pain disturbed your sleep? > 0.5 15 How much has the pain in the axilla disturbed your sleep? > 0.5 16 How much has the pain in the breast disturbed your sleep? > 0.5 17 Pain intensity in the operated-side arm and axilla in the morning? > 2.5 18 Worst pain intensity during the past week at one month > 1.5 19 Worst pain intensity during the past week at 6 months > 0.5 20 Worst pain intensity during the past week in the operated breast? > 1.5 21 Worst pain intensity during the past week somewhere? > 1.5 All “Persistent pain” class if sum of positively answered items ≥ 10 A patient is likely to develop persistent pain if at least 10 of the 21 items (rules) apply BMI body mass index, BDI Beck’s Depression Inventory, STAI Spielberger’s State-Trait Anxiety Inventory, STAXI Spielberger’s State-Trait Anger Expression Inventory, m months – Pain intensity in the operated-side arm and axilla Classifier performance (NRS from 0, no pain, to 10, worst imaginable pain), in the morning is queried at week 4 and 24 At the end of analytical step 4, the performance of the and averaged across the time points (item 17 in obtained classifiers to correctly assign a patient to the per - Table 1) sisting pain groups was tested and compared with that of – The worst pain intensity experienced during the the previously proposed six-factor risk index for persisting past week (items 18–21 in Table 1) is queried for pain  comprising the weighted items “age,” “chronic six body locations (breast, axilla, upper arm, joints, pain of any kind,” “number of previous operations,” “body lower arm, hand/fingers) and in addition, for any part mass index,” “preoperative pain in the area to be operated,” of the body (“somewhere”) using an 11-point NRS and “smoking,” This provided a performance of 70.6% sen- ranging from 0, not at all, to 10, very much. Items are sitivity and 45.4% specificity for correct assignment of a queried at week 4 and 24 averaged per week across patient to the persisting pain group. The accuracy for cor- all body locations (items 18 and 19 in Table 1). Rat- rect group prediction was 47.4%, and the balanced accuracy, ings for pain in the operated breast or “somewhere” which takes unequal group sizes into account, was 57.9%. are averaged across weeks 4 and 24 (items 20 and 21 However, the negative predictive value, which quantifies the in Table 1) probability that persisting pain will not develop when the test is negative, was high with 94.5% correctly excluding a Each of the 21 items (Table 1) obtained as described development toward persistent pain. above is assessed with respect to the Bayesian threshold The novel, more complex, diagnostic tool provided, via (Table 2, right column) and the sum of the “yes”/”no” 1000 Monte-Carlo random resamplings  of 50% of the answers to the question: “Above threshold?” are added. original data set, a cross-validated classifier performance Persistent pain is likely when the above sum equals 10 or of 79.3 ± 1.5% sensitivity and 51.4 ± 6% specificity for cor - higher. rect assignment of a patient to the persisting pain group. 1 3 Breast Cancer Research and Treatment (2018) 171:399–411 407 Fig. 4 Plot of the specificity versus the sensitivity of using all pos- the red dots indicate the number of conditions to be true according to sible combinations and thresholds for the 21 candidate predictors of the questions in Table 2. The maximum AUC, i.e., the best number of persistent pain after breast cancer surgery (classifier construction). conditions for a classifier, was obtained with at least 10 positive items The number of conditions for a positive classification into the “per - from Table 2, which was the result of the analysis shown in this figure sisting pain” groups ranges from n = 1–20 conditions. For all of these and the reason why the final predictive tool required 10 or more posi- positive conditions, the sensitivity, specificity, and the area under the tive items. The blue dots in the blue line indicate the corresponding curve (AUC = sensitivity · specificity) was calculated. The red dots in specificity (ordinate) versus sensitivity (abscissa) values. The lines the figure show AUC versus sensitivity. The black numbers close to have been drawn to enhance visibility and are spline interpolations The overall classification accuracy obtained in the 1,000 as having an unfavorable clinical outcome. These patients runs with resampled data was 86.2 ± 1.4%, and the bal- were thus defined as “presumably having persisting pain” anced accuracy was 65.4%. The negative predictive value and therefore belonging to the “persisting pain” group. was again high with 94.8% (binomial confidence interval: Taking the aforementioned unfavorable outcome as a sign 93.4–95.9%). of persisting pain, the 6-item classifier provided sensitivity Finally, the classifier performances were assessed on the and specificity of identifying potentially persisting pain of 143 patients who did not completely meet the strict criteria 94.1 and 43.7%, respectively, and an accuracy or balanced of persistent pain but nevertheless displayed an unsatisfac- accuracy of class assignment of 46.6 or 66.9%, respec- tory course of pain development and were therefore used as tively. The more complex novel classifier provided 14 true an exploratory “test” data set. Thus, when only the first cri- positive, 3 false negative, 56 false positive, and 70 true terion of the NRS-based classification into persisting and negative diagnoses of persistent pain. This translates into non-persisting pain groups was applied, i.e., patients with sensitivity and specificity of identifying a patient at risk NRS ≤ 3 at month 36 after surgery (NRS ≤ 3) were of 82.4 and 55.6%, respectively. The accuracy was 58.7%, month36 identified as belonging to the “non-persisting pain” group, the balanced accuracy was 69%, and a negative predictive while those with (NRS ≥ 4) had “persisting pain,” value of 95.9% was obtained (binomial confidence interval: month36 143 patients could be classified. Of these, 21 patients had 88.5–99.1%). at 3 years a pain score of ≥ 4 and were therefore considered 1 3 408 Breast Cancer Research and Treatment (2018) 171:399–411 The second major category of predictive features empha- Discussion sized the importance of psychological factors for the per- sistence of pain. This is in line with evidence that psycho- The main result from the present machine-learned analysis logical factors play an important role in pain persistence was that a more complex diagnostic tool than previously [4, 5, 11, 16, 27, 30, 31, 33, 34, 37]. Psychological factors proposed for a subgroup of 489 of the present patients  and pain have bidirectional influences, i.e., psychological improved the correct assignment of a patient to either the factors influence how the patient perceives and interprets persistent or non-persistent pain group, raising accuracy to pain and vice versa, constant pain may affect psychological 86% and balanced accuracy to 65.4%, while maintaining the factors and have an impact on mood either directly or via already high negative predictive value for persistent pain of its negative effects on sleep, functionality, social, and other 95%. activities [57, 58]. Thus, psychological factors may be at the The present criteria of persistent pain differed from the same time predictors, maintainers, changeable variables and proposal of the International Association for the Study of consequences of persistent pain . Therefore, psychologi- Pain in two aspects, i.e., (i) the time window was longer, six cal factors have been proposed as early predictive markers months instead of two, due to the particular clinical setting for the development of persisting pain (e.g., ). of breast cancer surgery, and (ii) the presence of relevant pre- In the present study, women at risk of developing per- operative pain jeopardizing the attribution of postoperative sistent pain reported more symptoms of depression and pain to the surgery. Eleven patients of the “persisting pain” state and trait anxiety. A higher tendency to inhibit anger, group had an NRS value > 3 for preoperative pain. In nine as assessed before surgery, was related to pain persistence. of the 74 patients assigned to the “persisting pain” group Anger inhibition has been suggested to be closely related to (12%), pain at 36 months was less although still > 3/10 NRS. general negative affect and depressive symptoms [6 ], and Thus, the definitive criteria for postsurgery pain were not thus to associate with chronic pain . Similar associa- met in these patients and the predictor correctly addressed tions of chronic pain with psychological factors have been persistent pain in a breast cancer setting without unequivo- reported in several clinical settings not restricted to can- cally implying a surgical procedure as a causal factor. This cer surgery [5, 16, 27, 31, 33, 34]. Feelings of anxiety and should be considered when interpreting and applying the lowered mood preoperatively are natural reactions to a seri- present diagnostic tool. ous disease, especially immediately after diagnosis. Inter- While in the original 6-item index “age,” “chronic pain estingly, the level of psychological wellbeing at 6 months of any kind,” “number of previous operations,” “body mass predicted persistent pain at 3 years, suggesting that better index,” “preoperative pain in the area to be operated,” and psychological coping after the acute phase might associate “smoking” were included, the novel selected features dif- with lower risk of pain persistence. It might therefore be fered partly from this earlier selection. This probably owes important to include routine assessment of psychological to the longer observation period of the present analysis, and coping at this time point and initiate appropriate interven- to the different techniques of feature selection. Nevertheless, tions as needed. some parameters agreed including a first major category of Finally, the third major category of predictive features predictive features comprised demographic parameters. An was related to pain. These features were present early on and increasing prevalence of chronic pain with age has been continued to persist for up to 3 years. Indeed, this is a rec- shown in several epidemiological studies . The preva- ognized observation and, due to this, early treatment of pain lence of neuropathic pain is also significantly higher in the has become a clinical routine. Interestingly, the early percep- elderly . Interestingly, many previous, mainly cross-sec- tion that pain impacted the patient’s life did not closely cor- tional studies, have reported that younger age would increase relate with pain intensity. This was indicated by the lack of the risk for persistent pain after breast cancer treatment . significant correlation between the ratings of the magnitude Previous studies have usually included mild pain intensities, of the impact on the life and the associated NRS ratings of whereas we used at least moderate pain as a cut-off. Younger the pain intensity at the first month after surgery (Kendall’s women may report more discomfort related to milder pain τ = 0.136) [26, 55]. due to a more active life style. Similarly, high body weight and chronic pain have been suggested to represent signifi- cant comorbidities, adversely impacting each other . The Conclusions association is not specific to a particular kind of pain but has been seen across several different aetiologies and clinical Using a data-driven approach in a cohort of 1,000 women conditions [8, 20, 38, 40, 46, 60]. Possible physiological operated on for breast cancer, a set of demographic, psycho- moderators include low-grade systemic inflammation and logical, and pain-related parameters available not later than metabolic disorders [7, 29]. at 6 months after surgery was associated with the persistence 1 3 Breast Cancer Research and Treatment (2018) 171:399–411 409 3. Bray F, Ren JS, Masuyer E, Ferlay J (2013) Global estimates of of pain in a 3-year follow-up. Applying machine-learning cancer prevalence for 27 sites in the adult population in 2008. Int techniques, a classifier was developed that finally took the J Cancer 132:1133–1145. https ://doi.org/10.1002/ijc.27711 shape of a questionnaire with “yes/no” decisions about clini- 4. Bruce J, Thornton AJ, Scott NW, Marfizo S, Powell R, Johnston cal features. Its main clinical strength lies in the exclusion M, Wells M, Heys SD, Thompson AM (2012) Chronic preopera- tive pain and psychological robustness predict acute postopera- of the possibility of developing persistent pain in a woman tive pain outcomes after surgery for breast cancer. Br J Cancer being treated for breast cancer with an accuracy of 95% 107:937–946. https ://doi.org/10.1038/bjc.2012.341 (negative predictive value). This provides a clinically reli- 5. Burke AL, Mathias JL, Denson LA (2015) Psychological func- able basis to release a patient early from multidisciplinary tioning of people living with chronic pain: a meta-analytic review. Br J Clin Psychol 54:345–360. https://doi.or g/10.1111/bjc.12078 therapy approaches . However, with not more than 80% 6. Burns JW, Quartana PJ, Bruehl S (2008) Anger inhibition and sensitivity and roughly 70% specificity, the performance of pain: conceptualizations, evidence and new directions. J Behav the novel classifier in predicting chronic pain in a patient Med 31:259–279. https ://doi.org/10.1007/s1086 5-008-9154-7 remained only moderate. This suggests that the demo- 7. Callaghan BC, Xia R, Reynolds E, Banerjee M, Rothberg AE, Burant CF, Villegas-Umana E, Pop-Busui R, Feldman EL (2016) graphic, psychological, and clinical parameters included in Association between metabolic syndrome components and poly- the present classifier are insufficient as predictors of persis- neuropathy in an obese population. JAMA Neurol 73:1468–1476. tent pain. Future improvements should be searched for in https ://doi.org/10.1001/jaman eurol .2016.3745 “omics,” i.e., features based on proteomics, lipidomics, or 8. Chou L, Brady SRE, Urquhart DM, Teichtahl AJ, Cicuttini FM, Pasco JA, Brennan-Olsen SL, Wluka AE (2016) The association genomics, to achieve the clinically desired early identifica- between obesity and low back pain and disability is affected by tion of patients who will develop persistent pain following mood disorders: a population-based, cross-sectional study of surgery for breast cancer. men. Medicine (Baltimore) 95:e3367. https ://doi.org/10.1097/ md.00000 00000 00336 7 Acknowledgements The authors thank Les Hearn for proofreading 9. Chou R, Gordon DB, de Leon-Casasola OA, Rosenberg JM, Bick- the manuscript. ler S, Brennan T, Carter T, Cassidy CL, Chittenden EH, Degen- hardt E, Griffith S, Manworren R, McCarberg B, Montgomery R, Murphy J, Perkal MF, Suresh S, Sluka K, Strassels S, Thirlby Author Contributions Conceived and designed the experiments: EK, R, Viscusi E, Walco GA, Warner L, Weisman SJ, Wu CL (2016) RS. Performed the study, collected the data: RS, TT, AME, TM. Ana- Management of postoperative pain: a clinical practice guideline lyzed the data: AU, JL. Wrote the paper: JL, EK, RS, AU, AME, TM, from the american pain society, the american society of regional DK. All authors have read and approved the manuscript. anesthesia and pain medicine, and the american society of anes- thesiologists’ committee on regional anesthesia, executive com- Funding The work has been supported by the European Union Sev- mittee, and administrative council. J Pain 17:131–157. https://doi. enth Framework Programme (FP7/2007–2013) under grant agreement org/10.1016/j.jpain .2015.12.008 no. 602919 (GLORIA, EK, JL), the Finnish Cancer Foundation (RS), 10. Dableh LJ, Yashpal K, Henry JL (2011) Neuropathic pain as a the Academy of Finland (EK), and the Helsinki University Hospital process: reversal of chronification in an animal model. J Pain Res Governmental Research funds (TYH2008225, TYH2010210, EK). The 4:315–323. https ://doi.org/10.2147/jpr.s1788 2 funders had no role in method design, data selection and analysis, deci- 11. Dimova V, Lötsch J, Hühne K, Winterpacht A, Heesen M, Par- sion to publish, or preparation of the manuscript. thum A, Weber PG, Carbon R, Griessinger N, Sittl R, Lauten- bacher S (2015) Association of genetic and psychological factors Compliance with ethical standards with persistent pain after cosmetic thoracic surgery. J Pain Res 8:829–844. https ://doi.org/10.2147/jpr.s9043 4 12. Dworkin RH, Turk DC, Farrar JT, Haythornthwaite JA, Jensen Conflict of interest The authors have declared no competing interests. MP, Katz NP, Kerns RD, Stucki G, Allen RR, Bellamy N, Carr DB, Chandler J, Cowan P, Dionne R, Galer BS, Hertz S, Jadad Open Access This article is distributed under the terms of the Crea- AR, Kramer LD, Manning DC, Martin S, McCormick CG, tive Commons Attribution 4.0 International License (http://creat iveco McDermott MP, McGrath P, Quessy S, Rappaport BA, Robbins mmons.or g/licenses/b y/4.0/), which permits unrestricted use, distribu- W, Robinson JP, Rothman M, Royal MA, Simon L, Stauffer JW, tion, and reproduction in any medium, provided you give appropriate Stein W, Tollett J, Wernicke J, Witter J, IMMPACT (2005) Core credit to the original author(s) and the source, provide a link to the outcome measures for chronic pain clinical trials: IMMPACT Creative Commons license, and indicate if changes were made. recommendations. Pain 113:9–19. https ://doi.or g/10.1016/j. pain.2004.09.012 13. Edwards RR, Dworkin RH, Sullivan MD, Turk DC, Wasan AD (2016) The Role of Psychosocial Processes in the Development References and Maintenance of Chronic Pain. J Pain 17:T70-92. https ://doi. org/10.1016/j.jpain .2016.01.001 14. Fayaz A, Croft P, Langford RM, Donaldson LJ, Jones GT (2016) 1. Andersen KG, Kehlet H (2011) Persistent pain after breast can- Prevalence of chronic pain in the UK: a systematic review and cer treatment: a critical review of risk factors and strategies for meta-analysis of population studies. BMJ Open 6:e010364. https prevention. J Pain 12:725–746. https ://doi.org/10.1016/j.jpain ://doi.org/10.1136/bmjop en-2015-01036 4 .2010.12.005 15. Ferlay J, Soerjomataram I, Dikshit R, Eser S, Mathers C, Rebelo 2. Beck AT, Ward CM, Mendelson M, Mock JE, Erbaugh JK (1961) M, Parkin DM, Forman D, Bray F (2015) Cancer incidence An inventory for measuring depression. Arch Gen Psychiat and mortality worldwide: sources, methods and major patterns 4:561–571 1 3 410 Breast Cancer Research and Treatment (2018) 171:399–411 in GLOBOCAN 2012. Int J Cancer 136:E359-386. https ://doi. for chronic pain in adults. J Behav Med 39:916–924. https ://doi. org/10.1002/ijc.29210 org/10.1007/s1086 5-016-9769-z 16. Fillingim RB, Ohrbach R, Greenspan JD, Knott C, Diatchenko 30. Lautenbacher S, Huber C, Schofer D, Kunz M, Parthum A, Weber L, Dubner R, Bair E, Baraian C, Mack N, Slade GD, Maixner PG, Roman C, Griessinger N, Sittl R (2010) Attentional and emo- W (2013) Psychological factors associated with development of tional mechanisms related to pain as predictors of chronic post- TMD: the OPPERA prospective cohort study. J Pain 14:T75-90. operative pain: a comparison with other psychological and physi- https ://doi.org/10.1016/j.jpain .2013.06.009 ological predictors. Pain 151:722–731. https ://doi.org/10.1016/j. 17. Fitzmaurice C, Allen C, Barber RM, Barregard L, Bhutta ZA, pain.2010.08.041 Brenner H, Dicker DJ, Chimed-Orchir O, Dandona R, Dandona L, 31. Linton SJ, Nicholas MK, MacDonald S, Boersma K, Bergbom S, Fleming T, Forouzanfar MH, Hancock J, Hay RJ, Hunter-Merrill Maher C, Refshauge K (2011) The role of depression and cata- R, Huynh C, Hosgood HD, Johnson CO, Jonas JB, Khubchan- strophizing in musculoskeletal pain. Eur J Pain 15:416–422. https dani J, Kumar GA, Kutz M, Lan Q, Larson HJ, Liang X, Lim ://doi.org/10.1016/j.ejpai n.2010.08.009 SS, Lopez AD, MacIntyre MF, Marczak L, Marquez N, Mokdad 32. Lötsch J, Ultsch A, Kalso E (2017) Prediction of persistent post- AH, Pinho C, Pourmalek F, Salomon JA, Sanabria JR, Sandar L, surgery pain by preoperative cold pain sensitivity: Biomarker Sartorius B, Schwartz SM, Shackelford KA, Shibuya K, Stanaway development with machine-learning-derived analysis. Br J J, Steiner C, Sun J, Takahashi K, Vollset SE, Vos T, Wagner JA, Anaesth aex236 Wang H, Westerman R, Zeeb H, Zoeckler L, Abd-Allah F, Ahmed 33. Margari F, Lorusso M, Matera E, Pastore A, Zagaria G, Bruno MB, Alabed S, Alam NK, Aldhahri SF, Alem G, Alemayohu MA, F, Puntillo F, Margari L (2014) Aggression, impulsivity, and sui- Ali R, Al-Raddadi R, Amare A, Amoako Y, Artaman A, Asayesh cide risk in benign chronic pain patients - a cross-sectional study. H, Atnafu N, Awasthi A, Saleem HB, Barac A, Bedi N, Bense- Neuropsych Dis Trea 10:1613–1620. https: //doi.org/10.2147/ndt. nor I, Berhane A, Bernabe E, Betsu B, Binagwaho A, Boneya D, s6620 9 Campos-Nonato I, Castaneda-Orjuela C, Catala-Lopez F, Chiang 34. McCracken LM, Gross RT, Aikens J, Carnrike CL Jr (1996) The P, Chibueze C, Chitheer A, Choi JY, Cowie B, Damtew S, das assessment of anxiety and fear in persons with chronic pain: a Neves J, Dey S, Dharmaratne S, Dhillon P, Ding E, Driscoll T, comparison of instruments. Behav Res Ther 34:927–933 Ekwueme D, Endries AY, Farvid M, Farzadfar F, Fernandes J, 35. McGrayne SB (2011) The Theory That Would Not Die: How Fischer F, TT GH, Gebru A, Gopalani S, Hailu A et al (2017) Bayes’ Rule Cracked the Enigma Code, Hunted Down Russian Global, regional, and national cancer incidence, mortality, years Submarines & Emerged Triumphant from Two Centuries of Con- of life lost, years lived with disability, and disability-adjusted life- troversy. Yale University Press, New Haven years for 32 cancer groups, 1990 to 2015: a systematic analysis 36. Mejdahl MK, Andersen KG, Gartner R, Kroman N, Kehlet H for the global burden of disease study. JAMA Oncol 3:524–548. (2013) Persistent pain and sensory disturbances after treatment https ://doi.org/10.1001/jamao ncol.2016.5688 for breast cancer: six year nationwide follow-up study. Bmj 18. Good PI (2006) Resampling methods: a practical guide to data 346:f1865. https ://doi.org/10.1136/bmj.f1865 analysis. Birkhäuser, Boston 37. Meretoja TJ, Leidenius MH, Tasmuth T, Sipila R, Kalso E (2014) 19. Guyon I, Andr E (2003) An introduction to variable and feature Pain at 12 months after surgery for breast cancer. JAMA 311:90– selection. J Mach Learn Res 3:1157–1182 92. https ://doi.org/10.1001/jama.2013.27879 5 20. Hussain SM, Urquhart DM, Wang Y, Dunstan D, Shaw JE, Magli- 38. Motaghedi R, Bae JJ, Memtsoudis SG, Kim DH, Beathe JC, Paroli ano DJ, Wluka AE, Cicuttini FM (2016) Associations between L, YaDeau JT, Gordon MA, Maalouf DB, Lin Y, Ma Y, Cunning- television viewing and physical activity and low back pain in ham-Rundles S, Liu SS (2014) Association of obesity with inflam- community-based adults: a cohort study. Medicine (Baltimore) mation and pain after total hip arthroplasty. Clin Orthop Relat Res 95:e3963. https ://doi.org/10.1097/md.00000 00000 00396 3 472:1442–1448. https ://doi.org/10.1007/s1199 9-013-3282-2 21. Janssen SA (2002) Negative affect and sensitization to pain. Scand 39. Murphy KP (2012) Machine Learning: A Probabilistic Perspec- J Psychol 43:131–137 tive. The MIT Press, Cambridge 22. Julius D, Basbaum AI (2001) Molecular mechanisms of nocicep- 40. Okifuji A, Hare BD (2015) The association between chronic pain tion. Nature 413:203–210. https ://doi.org/10.1038/35093 019 and obesity. J Pain Res 8:399–408. https ://doi.org/10.2147/jpr. 23. Juran JM (1975) The non-Pareto principle; Mea culpa. Qual Prog s5559 8 8:8–9 41. Paice JA, Portenoy R, Lacchetti C, Campbell T, Cheville A, Citron 24. Kaunisto MA, Jokela R, Tallgren M, Kambur O, Tikkanen E, M, Constine LS, Cooper A, Glare P, Keefe F, Koyyalagunta L, Tasmuth T, Sipilä R, Palotie A, Estlander A-M, Leidenius M, Levy M, Miaskowski C, Otis-Green S, Sloan P, Bruera E (2016) Ripatti S, Kalso EA (2013) Pain in 1,000 women treated for breast Management of chronic pain in survivors of adult cancers: ameri- cancer: a prospective study of pain sensitivity and postoperative can society of clinical oncology clinical practice guideline. J Clin pain. Anesthesiology 119:1410–1421. https ://doi.org/10.1097/ Oncol 34:3325–3345. https ://doi.org/10.1200/jco.2016.68.5206 aln.00000 00000 00001 2 42. Pareto V (1909) Manuale di economia politica, Milan: Società 25. Kehlet H, Jensen TS, Woolf CJ (2006) Persistent postsurgical editrice libraria, revised and translated into French as Manuel pain: risk factors and prevention. Lancet 367:1618–1625. https:// d’économie politique. In:Giard et Briére, Paris doi.org/10.1016/s0140 -6736(06)68700 -x 43. Pergolizzi JV, Gharibo C, Ho K-Y (2015) Treatment considera- 26. Kendall MG (1938) A new measure of rank correlation. Biom- tions for cancer pain: a global perspective. Pain Pract 15:778–792. etrika 30:81–93. https ://doi.org/10.1093/biome t/30.1-2.81https ://doi.org/10.1111/papr.12253 27. Knaster P, Estlander AM, Karlsson H, Kaprio J, Kalso E (2016) 44. Portenoy RK (1992) Cancer pain: pathophysiology and syn- Diagnosing depression in chronic pain patients: DSM-IV major dromes. Lancet 339:1026–1031 depressive disorder vs. beck depression inventory (BDI). PLoS 45. R Development Core Team (2008) R: a language and environment ONE 11:e0151982. https://doi.or g/10.1371/journal.pone.01519 82 for statistical computing 28. Kullback S, Leibler RA (1951) On Information and Sufficiency. 46. Ray L, Lipton RB, Zimmerman ME, Katz MJ, Derby CA pp 79–86. https ://doi.org/10.1214/aoms/11777 29694 (2011) Mechanisms of association between obesity and chronic 29. Lasselin J, Kemani MK, Kanstrup M, Olsson GL, Axelsson J, pain in the elderly. Pain 152:53–59. https ://doi.org/10.1016/j. pain.2010.08.043 Andreasson A, Lekander M, Wicksell RK (2016) Low-grade inflammation may moderate the effect of behavioral treatment 1 3 Breast Cancer Research and Treatment (2018) 171:399–411 411 47. Schaible HG (2007) Peripheral and central mechanisms of pain 54. Trimble EL, Ungerleider RS, Abrams JA, Kaplan RS, Feigal EG, generation. Handb Exp Pharmacol:3–28 Smith MA, Carter CL, Friedman MA (1993) Neoadjuvant therapy 48. Schou Bredal I, Smeby NA, Ottesen S, Warncke T, Schlichting in cancer treatment. Cancer 72:3515–3524 E (2014) Chronic pain in breast cancer survivors: comparison of 55. Udovičić M, Baždarić K, Bilić-Zulle L, Petrovečki M (2007) What psychosocial, surgical, and medical characteristics between survi- we need to know when calculating the coefficient of correlation? vors with and without pain. J Pain Symptom Manage 48:852–862. Biochem Med 17:10–15 https ://doi.org/10.1016/j.jpain symma n.2013.12.239 56. Ultsch A, Lötsch J (2015) Computed ABC Analysis for rational 49. Sipilä R, Estlander A-M, Tasmuth T, Kataja M, Kalso E (2012) selection of most informative variables in multivariate data. PLoS Development of a screening instrument for risk factors of persis- ONE 10:e0129767. https://doi.or g/10.1371/journal.pone.01297 67 tent pain after breast cancer surgery. Br J Cancer 107:1459–1466. 57. Vlaeyen JW, Crombez G, Linton SJ (2016) The fear-avoidance https ://doi.org/10.1038/bjc.2012.445 model of pain. Pain 157:1588–1589. https ://doi.org/10.1097/j. 50. Spielberger CD (1999) Staxi-2: state-trait anger expression inven-pain.00000 00000 00057 4 tory-2; professional manual. PAR, Psychological Assessment 58. Vlaeyen JW, Linton SJ (2012) Fear-avoidance model of chronic Resources musculoskeletal pain: 12 years on. Pain 153:1144–1147. https :// 51. Spielberger CD, Gorsuch RL, Lushene RE (1970) The State-Trait doi.org/10.1016/j.pain.2011.12.009 Anxiety Inventory (test manual). Consulting Psychologists Press, 59. Weiss SM, Indurkhya N (1995) Rule-based machine learning Palo Alto, California methods for functional prediction. J Artif Int Res 3:383–403 52. Tasmuth T, von Smitten K, Kalso E (1996) Pain and other symp- 60. Wright LJ, Schur E, Noonan C, Ahumada S, Buchwald D, Afari toms during the first year after radical and conservative surgery N (2010) Chronic pain, overweight, and obesity: findings from a for breast cancer. Br J Cancer 74:2024–2031 community-based twin registry. J Pain 11:628–635. https ://doi. 53. Torrance N, Smith BH, Bennett MI, Lee AJ (2006) The epide-org/10.1016/j.jpain .2009.10.004 miology of chronic pain of predominantly neuropathic origin. Results from a general population survey. J Pain 7:281–289. https ://doi.org/10.1016/j.jpain .2005.11.008 1 3
Breast Cancer Research and Treatment – Springer Journals
Published: Jun 6, 2018
It’s your single place to instantly
discover and read the research
that matters to you.
Enjoy affordable access to
over 18 million articles from more than
15,000 peer-reviewed journals.
All for just $49/month
Query the DeepDyve database, plus search all of PubMed and Google Scholar seamlessly
Save any article or search result from DeepDyve, PubMed, and Google Scholar... all in one place.
Get unlimited, online access to over 18 million full-text articles from more than 15,000 scientific journals.
Read from thousands of the leading scholarly journals from SpringerNature, Elsevier, Wiley-Blackwell, Oxford University Press and more.
All the latest content is available, no embargo periods.
“Hi guys, I cannot tell you how much I love this resource. Incredible. I really believe you've hit the nail on the head with this site in regards to solving the research-purchase issue.”Daniel C.
“Whoa! It’s like Spotify but for academic articles.”@Phil_Robichaud
“I must say, @deepdyve is a fabulous solution to the independent researcher's problem of #access to #information.”@deepthiw
“My last article couldn't be possible without the platform @deepdyve that makes journal papers cheaper.”@JoseServera