journal article
LitStream Collection
Home
doi: 10.1093/bjs/znac462pmid: 36680383
Facing the challenge of every second patient developing complications arising from surgery1, it is intriguing that the next major advancements in perioperative care will be technological2. These technological advancements can be as simple as digitalization of information, facilitating standardization, and aiding decision support3. A recent example of this was the demonstration of a reduction in postoperative morbidity and mortality in patients undergoing surgery for pancreatic cancer4. The authors selected evidence-based recommendations for individualizing postoperative clinical observations as well as the use and timing of standard imaging techniques. This standardized and individualized algorithm was the real innovation; however, the platform accelerates and facilitates its implementation. With the development of digitalization in the healthcare sector and new additional data sources, including data from electronic health records, wearables, and patient-reported outcome measures, we are on the verge of a revolution in healthcare. This can only be a reality, however, if these data can be transformed and understood by clinicians. Machine learning (ML) algorithms offer a new avenue for integrating multidomain data into improvements in perioperative care at the time of diagnosis, during surgery, or at discharge from hospital (Fig. 1). A prerequisite for this improvement is to solve challenges regarding understanding the potential and limitations of ML, while at the same time building on previous methodologies that have already proven to be valuable in a clinical setting, but whose full potential has not been realized owing to lack of implementation5. Although many prediction models have been developed over the years6, they are not widely implemented in the diagnostic setting, for prognostication, or in selecting patients for different perioperative treatment trajectories. An important consideration is the lack of clear clinical benefit and external validation, limiting the benefits outside the hospital setting, region or country in which the model has been developed7. The prognostication tools are also for the major part rules-based, and therefore not designed to compute complex, non-linear relationships leading to inaccurate models that cannot capture physiological interactions8. Fig. 1 Open in new tabDownload slide Data to be captured at each phase of the surgical journey with scope for introducing machine learning-based prediction models ML, machine learning; MDT, multidisciplinary team. Artificial intelligence is a branch of computer science that describes the computer automation of intelligent tasks that are usually performed by humans. ML is a subset of artificial intelligence which aids human learning through algorithms to perform tasks without being programmed explicitly to do so. The advanced data modelling techniques are, therefore, separate from traditional statistical models and tests that many clinicians are familiar with. The aim for the clinician is to provide the best possible perioperative care. The volume and complexity of medical information and the availability of data may be overwhelming, hence the need for methodologies that can integrate data inputs and produce actionable insight into the health status of patients. Although ML provides a robust set of data analysis tools, the associated complexities are also a barrier to clinical utility. A collaboration between clinicians and data scientists is a prerequisite, as both clinical knowledge and technical expertise are needed to create these digital solutions to improve patient care. Without sufficient collaboration there is a possibility of trivial findings, incorrect use, and false interpretation, leading at best to a waste of resources and, in the worst case, patient harm. A good place to start is to examine at which stages during the patient journey there is a clinical need for additional insight, and the technical possibility to implement a ML model9. The obvious possibilities are at the time of diagnosis, during and immediately after surgery, and at the time of discharge (Fig. 1). The relevant data to select for development of a ML algorithm are important to identify. A common application is prediction of an unobserved variable based on a set of observed variables. Often a vast number of potential variables can be included as an input, but only a smaller subset is required. To solve this problem, a common step in developing prediction models is feature selection, whereby the features which best predict the outcomes are selected10. Selecting which co-variates improve performance allows researchers to retain more variables early in the process compared with traditional statistical methods11. ML provides the opportunity to include data that may not be expected to contribute directly to the model owing to an expected lack of biological importance. ML approaches, therefore, provide the possibility to integrate data on a much wider perspective, leading to better performance of the models9. Integrating and harmonizing institutional, regional, national, and international health registries in a common data model provides an opportunity to generate data infrastructures that may provide a unique platform to develop prediction models in the perioperative setting. For ML models to provide further value, existing data sources should be integrated across data domains, leading to more patients with greater granularity of information, including in treatment pathways12. Including rich descriptions of patient trajectories in predictive modelling might reduce heterogeneity in risk profiles and enhance model performance13. Examples of data sources that could be unified to give incremental potential for generating strong prediction models include registries focused on outcomes after surgery (for example a cancer registry), registries comprising data on socioeconomic factors, drug exposure, laboratory results, genomic information, pathology reports, and contacts with healthcare providers. These databases may each include instrumental data that can be used for generating prediction models focused on outcomes such as complications, readmissions or mortality after surgery9. In combination, the data will cover more physiological variation at the patient level, leading to models with better discrimination and calibration. As the number of patient records and the richness of the content increases, model performance can improve while accumulating further input variables or data that are not readily available. This can limit the feasibility of clinical implementation, unless solutions that automate the retrieval of patient data to populate the models are developed and deployed. Alternatively, parsimonious models can be created with fewer input variables, and, if performance is still sufficient for the clinical problem, the models can be deployed. The latter approach has the potential to improve the transparency and interpretability of the prediction models. Surgeons need to understand the dynamics behind ML-based prediction models in order to use them in a clinical setting. A prediction model may include variables that are not obviously relevant (for example a genotype where there are no data to confirm physiological relationship to the outcome predicted), or where the relationship between a variable and the outcome is paradoxical when all other variables are constant. This interesting concept needs to be communicated to the end user14. We are at the very beginning of a new era in medicine where technological advancements and use of ML-based prediction models can substantially improve perioperative treatments through better patient selection and at instrumental phases during the patient journey. This is not just selecting for surgical or anaesthetic approaches in theatre, but selecting for individualized prehabilitation, postoperative care, and rehabilitation. As stated by the surgical oncologist Blake Cady, ‘Biology is King, selection is Queen and technical manoeuvres by the surgeon are Princes and Princesses’15. For researchers who continue to undertake these studies in the future, it remains vital that they properly address the need, or the reasons, for applying ML to the problem. A clear understanding of best practice is needed16 and should be implemented beforehand. Based on this valid baseline, the potential added value of ML should be communicated clearly. Acknowledgements The authors thank E. M. Maravelia for assistance with figure generation, and K. B. Brauner and A. Rosen for critical review. Funding The authors have no funding to declare. Disclosure The authors declare no conflict of interest. Data availability Data related to this Needle Point editorial can be accessed by contacting the corresponding author. References 1 GlobalSurg Collaborative and NIHR Global Health Research Unit on Global Surgery . Effects of hospital facilities on patient outcomes after cancer surgery: an international, prospective, observational study . Lancet Glob Health 2022 ; 10 : e1003 – e1011 Crossref Search ADS PubMed WorldCat 2 COVIDSurg Collaborative . Machine learning risk prediction of mortality for patients undergoing surgery with perioperative SARS-CoV-2: the COVIDSurg mortality score . Br J Surg 2021 ; 108 : 1274 – 1292 Crossref Search ADS PubMed WorldCat 3 van der Meij E , Anema JR, Leclercq WKG, Bongers MY, Consten ECJ, Schraffordt Koops SE et al. Personalised perioperative care by e-health after intermediate-grade abdominal surgery: a multicentre, single-blind, randomised, placebo-controlled trial . Lancet 2018 ; 392 : 51 – 59 Google Scholar Crossref Search ADS PubMed WorldCat 4 Smits FJ , Henry AC, Besselink MG, Busch OR, van Eijck CH, Arntz M et al. Algorithm-based care versus usual care for the early recognition and management of complications after pancreatic resection in The Netherlands: an open-label, nationwide, stepped-wedge cluster-randomised trial . Lancet 2022 ; 399 : 1867 – 1875 Google Scholar Crossref Search ADS PubMed WorldCat 5 Lammers DT , Eckert CM, Ahmad MA, Bingham JR, Eckert MJ. A surgeon’s guide to machine learning . Ann Surg Open 2021 ; 2 : e091 Google Scholar Crossref Search ADS WorldCat 6 Bilimoria KY , Liu Y, Paruch JL, Zhou L, Kmiecik TE, Ko CY et al. Development and evaluation of the universal ACS NSQIP surgical risk calculator: a decision aid and informed consent tool for patients and surgeons . J Am Coll Surg 2013 ; 217 : 833 – 842.e1–3 Google Scholar Crossref Search ADS PubMed WorldCat 7 Lin V , Gögenur S, Pachler F, Fransgaard T, Gögenur I. Risk prediction for complications in inflammatory bowel disease surgery: external validation of the American College of Surgeons’ National Surgical Quality Improvement Program Surgical Risk Calculator . J Crohns Colitis 2022 ; DOI: 10.1093/ecco-jcc/jjac114 [Epub ahead of print] Google Scholar OpenURL Placeholder Text WorldCat Crossref 8 van den Bosch T , Warps ALK, de Nerée Tot Babberich MPM, Stamm C, Geerts BF, Vermeulen L et al. Predictors of 30-day mortality among Dutch patients undergoing colorectal cancer surgery, 2011–2016 . JAMA Netw open 2021 ; 4 : e217737 Google Scholar OpenURL Placeholder Text WorldCat 9 Vogelsang RP , Bojesen RD, Hoelmich ER, Orhan A, Buzquurz F, Cai L et al. Prediction of 90-day mortality after surgery for colorectal cancer using standardized nationwide quality-assurance data . BJS Open 2021 ; 5 : zrab023 Google Scholar OpenURL Placeholder Text WorldCat 10 Remeseiro B , Bolon-Canedo V. A review of feature selection methods in medical applications . Comput Biol Med 2019 ; 112 : 103375 Google Scholar OpenURL Placeholder Text WorldCat 11 Bzdok D , Altman N, Krzywinski M. Statistics versus machine learning . Nat Methods 2018 ; 15 : 233 – 234 Google Scholar Crossref Search ADS PubMed WorldCat 12 Hripcsak G , Ryan PB, Duke JD, Shah NH, Park RW, Huser V et al. Characterizing treatment pathways at scale using the OHDSI network . Proc Natl Acad Sci U S A 2016 ; 113 : 7329 – 7336 Google Scholar Crossref Search ADS PubMed WorldCat 13 Jensen AB , Moseley PL, Oprea TI, Ellesøe SG, Eriksson R, Schmock H et al. Temporal disease trajectories condensed from population-wide registry data covering 6.2 million patients . Nat Commun 2014 ; 5 : 4022 Google Scholar Crossref Search ADS PubMed WorldCat 14 Prosperi M , Guo Y, Sperrin M, Koopman JS, Min JS, He X et al. Causal inference and counterfactual prediction in machine learning for actionable healthcare . Nat Mach Intell 2020 ; 2 : 369 – 375 Google Scholar Crossref Search ADS WorldCat 15 Cady B . Basic principles in surgical oncology . Arch Surg 1997 ; 132 : 338 – 346 Google Scholar Crossref Search ADS PubMed WorldCat 16 Kehlet H , Memtsoudis SG. Perioperative care guidelines: conflicts and controversies . Br J Surg 2020 ; 107 : 1243 – 1244 Google Scholar Crossref Search ADS PubMed WorldCat © The Author(s) 2023. Published by Oxford University Press on behalf of BJS Society Ltd. All rights reserved. For permissions, please e-mail: [email protected] This article is published and distributed under the terms of the Oxford University Press, Standard Journals Publication Model (https://academic.oup.com/journals/pages/open_access/funder_policies/chorus/standard_publication_model)
Galema, Hidde A; van Ginhoven, Tessa M; Franssen, Gaston J H; Hofland, Johannes; Bouman, Claire G O T; Verhoef, Cornelis; Vahrmeijer, Alexander L; Hutteman, Merlijn; Hilling, Denise E; Keereweer, Stijn
doi: 10.1093/bjs/znad043pmid: 36861217
Singh, Tino; Lavikainen, Lauri I; Halme, Alex L E; Aaltonen, Riikka; Agarwal, Arnav; Blanker, Marco H; Bolsunovskyi, Kostiantyn; Cartwright, Rufus; García-Perdomo, Herney; Gutschon, Rachel; Lee, Yung; Pourjamal, Negar; Vernooij, Robin W M; Violette, Philippe D; Haukka, Jari; Guyatt, Gordon H; Tikkinen, Kari A O
Lim, Arthur J M; Mohamed, Abduraheem H; Hitchman, Louise H; Lathan, Ross; Ravindhran, Bharadhwaj; Sidapra, Misha M; Smith, George; Chetter, Ian C; Carradice, Daniel
doi: 10.1093/bjs/znad048pmid: 36894167
Showing 1 to 10 of 25 Articles
doi: 10.1093/bjs/znad035pmid: 36912116
BackgroundThe timing at which venous thromboembolism (VTE) occurs after major surgery has major implications for the optimal duration of thromboprophylaxis. The aim of this study was to perform a systematic review and meta-analysis of the timing of postoperative VTE up to 4 weeks after surgery.MethodsA systematic search of MEDLINE, Scopus, and CINAHL databases was performed between 1 January 2009 and 1 April 2022. Prospective studies that recruited patients who underwent a surgical procedure and reported at least 20 symptomatic, postoperative VTE events by time were included. Two reviewers independently selected studies according to the eligibility criteria, extracted data, and evaluated risk of bias. Data were analysed with a Poisson regression model, and the GRADE approach was used to rate the certainty of evidence.ResultsSome 6258 studies were evaluated, of which 22 (11 general, 5 urological, 4 mixed, and 2 orthopaedic postoperative surgical populations; total 1 864 875 patients and 24 927 VTE events) were eligible. Pooled evidence of moderate certainty showed that 47.1 per cent of the VTE events occurred during the first, 26.9 per cent during the second, 15.8 per cent during the third, and 10.1 per cent during the fourth week after surgery. The timing of VTE was consistent between individual studies.ConclusionAlthough nearly half of symptomatic VTE events in first 4 weeks occur during the first postoperative week, a substantial number of events occur several weeks after surgery. These data will inform clinicians and guideline developers about the duration of postoperative thromboprophylaxis.
BackgroundMechanochemical ablation (MOCA) is an alternative method to endovenous thermal ablation (EVTA) for the treatment of superficial venous incompetence that does not require tumescent anaesthesia. The aim of this study was to compare the outcomes from RCTs of MOCA versus EVTA.MethodsA search was conducted in MEDLINE, Embase, and the Cochrane Central Register of Controlled Trials (CENTRAL). Meta-analysis inclusion was restricted to RCTs comparing MOCA against EVTA. Outcomes included anatomical occlusion rate, disease-specific quality of life using the Aberdeen Varicose Vein Questionnaire, procedural and postprocedural pain, and rates of venous thromboembolism.ResultsFour RCTs were included in the meta-analysis comprising 654 patients. The anatomical occlusion rate at 1 year was lower after MOCA than EVTA (risk ratio 0.85, 95 per cent c.i. 0.78 to 0.91; P < 0.001). No significant differences were detected in procedural pain (mean difference −3.25, −14.25 to 7.74; P = 0.560) or postprocedural pain (mean difference −0.63, −2.15 to 0.89; P = 0.420). There were no significant differences in Aberdeen Varicose Vein Questionnaire score at 1 year (mean difference 0.06, −0.50 to 0.62; P = 0.830) or in incidence of venous thromboembolism (risk ratio 0.72, 95 per cent c.i. 0.14 to 3.61; P = 0.690).ConclusionThe rate of successful anatomical occlusion after MOCA is significantly lower than that after EVTA, but there is no difference in procedural and postprocedural pain between the two interventions. Long-term data are required to assess the impact of the reduced vein occlusion rate on clinical outcomes such as quality of life and reintervention.