Machine learning and artificial intelligence research for patient benefit: 20 critical questions on transparency, replicability, ethics, and effectiveness.
Vollmer S, Mateen BA, Bohner G, Király FJ, Ghani R, Jonsson P, Cumbers S, Jonas A, McAllister KSL, Myles P, Granger D, Birse M, Branson R, Moons KGM, Collins GS, Ioannidis JPA, Holmes C, Hemingway H. · 2020
Read Machine learning and artificial intelligence research for patient benefit: 20 critical questions on transparency, replicability, ethics, and effectiveness. by Vollmer S, Mateen BA, Bohner G, Király FJ, Ghani R, Jonsson P, Cumbers S, Jonas A, McAllister KSL, Myles P, Granger D, Birse M, Branson R, Moons KGM, Collins GS, Ioannidis JPA, Holmes C, Hemingway H. (2020) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc11515850.txt.
Introduction
Vollmer and a large team pose 20 critical questions for machine learning research aimed at patient benefit in this 2020 paper. Machine learning and AI open new ways to use growing health data. Yet the wider literature lacks openness, clear reporting for replication, attention to ethics, and clear proof of effectiveness. The authors trace part of the problem to missing best-practice guidance specific to these methods.
Their 20 questions form a framework called TREE, for transparency, reproducibility, ethics, and effectiveness. Research groups can use them to shape study design, conduct, and reporting. Editors and peer reviewers can use them to judge submissions. Funders and regulators are also addressed. The questions were built for interdisciplinary health projects. The authors offer the set as a preliminary solution and invite wider adoption.
Catalog record
| CREATOR | Vollmer S, Mateen BA, Bohner G, Király FJ, Ghani R, Jonsson P, Cumbers S, Jonas A, McAllister KSL, Myles P, Granger D, Birse M, Branson R, Moons KGM, Collins GS, Ioannidis JPA, Holmes C, Hemingway H. |
|---|---|
| YEAR | 2020 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 66 KB |
| LICENSE | CC BY 4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Machine learning and artificial intelligence research for patient benefit: 20 critical questions on transparency, replicability, ethics, and effectiveness By Sebastian Vollmer, Bilal A Mateen, Gergo Bohner, Franz J Király, Rayid Ghani, Pall Jonsson, Sarah Cumbers, Adrian Jonas, Katherine S L McAllister, Puja Myles, David Grainger, Mark Birse Published in: The BMJ DOI: 10.1136/bmj.l6927 ABSTRACT Machine learning, artificial intelligence, and other modern statistical methods are providing new opportunities to operationalise previously untapped and rapidly growing sources of data for patient benefit. Despite much promising research currently being undertaken, particularly in imaging, the literature as a whole lacks transparency, clear reporting to facilitate replicability, exploration for potential ethical concerns, and clear demonstrations of effectiveness. Among the many reasons why these problems exist, one of the most important (for which we provide a preliminary solution here) is the current lack of best practice guidance specific to machine learning and artificial intelligence. However, we believe that interdisciplinary groups pursuing research and impact projects involving machine learning and artificial intelligence for health would benefit from explicitly addressing a series of questions concerning transparency, reproducibility, ethics, and effectiveness (TREE). The 20 critical questions proposed here provide a framework for research groups to inform the design, conduct, and reporting; for editors and peer reviewers to evaluate contributions to the literature; and for patients, clinicians and policy makers to critically appraise where new findings may deliver patient benefit. Machine learning (ML), artificial intelligence (AI), and other modern statistical methods are providing new opportunities to operationalise previously untapped and rapidly growing sources of data for patient benefit. The potential uses include improving diagnostic accuracy,1 more reliably predicting prognosis,2 targeting treatments,3 and increasing the operational efficiency of health systems.4 Examples of potentially disruptive technology with early promise include image based diagnostic applications of ML/AI, which have shown the most early clinical promise (eg, deep learning based algorithms improving accuracy in diagnosing retinal pathology compared with that of specialist physicians5), or natural language processing used as a tool to extract information from structured and unstructured (that is, free) text embedded in electronic health records.2 Although we are only just beginning to understand the wealth of opportunities afforded by these methods, there is growing concern in the academic community that because the products of these methods are not perceived in the same way as other medical (eg, pharmacological) interventions,6 they do not have well defined guidelines for development and use and rarely undergo the same degree of scrutiny. NEED FOR GUIDANCE Several high profile publications have shown a lack of transparency,7 8 replicability,9 ethics,10 and effectiveness11 in the reporting and assessment of ML/AI based prediction models. This growing body of evidence suggests that while many best practice recommendations for design, conduct, analysis, reporting, impact assessment, and clinical implementation can be borrowed from the traditional biostatistics and medical statistics literature,12 they are not sufficient to guide the use of ML/AI in research. Producing such guidance is a major undertaking due to the ever-growing battery of ML/AI algorithms and the multifaceted nature of assessing performance and clinical impact. Not taking action is unacceptable, and if we wait for a more definitive solution, we risk wasting valuable work,13 14 15 16 17 while allowing futile research to continue unchecked, or worse, translation of ineffective (or even harmful) algorithms from the computer bench to the bedside. Summary points Clinically relevant research using modern statistical methods (such as machine learning and artificial intelligence) is too often limited by one or more of TREE concerns (transparency, reproducibility, ethics, and effectiveness); addressing these concerns can facilitate appropriate translation from computer bench to patient benefitHere we propose 20 critical questions that offer a framework for users and generators of ML/AI researchFor research generators, the 20 questions can inform the way research groups design, conduct, and report their researchFor editors and peer reviewers, the checklist provides a starting point for evaluating the quality and clinical relevance of articles For the users of such research findings—including healthcare professionals, patients, and the public—the 20 questions highlight important issues for critical appraisal INITIAL FRAMEWORK We propose a series of 20 critical questions (box 1) to help identify common pitfalls that can undermine ML/AI based applications in health. The questions span issues of transparency, reproducibility, ethics, and effectiveness (TREE). Appendix 1 includes a brief description of how these questions were generated. The questions are not only relevant for those who use the findings (that is, patients and policy makers), but also for those who generate ML/AI health research. We envision this checklist of questions as providing a framework for journal editors, peer reviewers, and those who critically evaluate contributions to the literature; for researchers as a reference to inform the way that research groups design and conduct ML/AI research; for regulators judging algorithm approval; and for educators of clinicians and academic disciplines involved. Current practice in research publication is heterogeneous with relevant questions not clearly dealt with. Clearly further work is needed to build consensus on what constitutes acceptable practice and reporting, but we believe that adoption of this framework as a starting point, and of other related publications,18 will help to build trust in the underlying processes and results of health related ML/AI research. Box 1Critical questions for health related technology involving machine learning and artificial intelligenceInceptionWhat is the health question relating to patient benefit?What evidence is there that the development of the algorithm was informed by best practices in clinical research and epidemiological study design?StudyWhen and how should patients be involved in data collection, analysis, deployment, and use?Are the data suitable to answer the clinical question—that is, do they capture the relevant real world heterogeneity, and are they of sufficient detail and quality?Does the validation methodology reflect the real world constraints and operational procedures associated with data collection and storage?What computational and software resources are required for the task, and are the available resources sufficient to tackle this problem?Statistical methodsAre the reported performance metrics relevant for the clinical context in which the model will be used?Is the ML/AI algorithm compared to the current best technology, and against other appropriate baselines?Is the reported gain in statistical performance with the ML/AI algorithm justified in the context of any trade-offs?ReproducibilityOn what basis are data accessible to other researchers?Are the code, software, and all other relevant parts of the prediction modelling pipeline available to others to facilitate replicability?Is there organisational transparency about the flow of data and results?Impact evaluationAre the results generalisable to settings beyond where the system was developed (that is, results reproducibility/external validity)?Does the model create or exacerbate inequities in healthcare by age, sex, ethnicity, or other protected characteristics?What evidence is there that clinicians and patients find the model and its output (reasonably) interpretable?How will evidence of real world model effectiveness in the proposed clinical setting be generated, and how will unintended consequences be prevented?ImplementationHow is the model being regularly reassessed, and updated as data quality and clinical practice changes (that is, post-deployment monitoring)?Is the ML/AI model cost effective to build, implement, and maintain?How will the potential financial benefits be distributed if the ML/AI model is commercialised?How have the regulatory requirements for accreditation/approval been addressed? CRITICAL QUESTIONS Inception (questions 1-2) What is the health question relating to patient benefit? The vast majority of published clinical