To the Editor, I was encouraged by the findings of a recent review by Duong et al1 discussing the applications of artificial intelligence (AI) to medical education. This topic of discussion is of particular interest because as a foundation doctor, I along with my colleagues have had recent exposure to radiology education as medical students and as foundation doctors in a district general hospital (DGH). From anecdotal discussion with other foundation doctors, it seems the amount of formal radiology teaching medical students in the UK receive varies depending on the university they attend and their course structure. This represents a potential cohort where AI precision education could be beneficial, particularly in those institutions with less formal radiology teaching. By using AI to assign students ‘must see’ cases as described, it could provide students more consistent exposure and experience to situations involving radiology they will encounter as a foundation doctor. An example of this would be interpreting chest x rays, a common scenario on call. Furthermore the solution mentioned in the article of addressing and correcting cognitive bias gives students more opportunity to correct and explain mistakes they may not otherwise have been able to address easily. This again applies to universities where less formal radiology teaching is provided. As supported by Taylor and Hamdy,2 medical education should incorporate combinations of different learning styles, learning phases, and feedback. In busy clinical environments addressing all of these for students can be understandably challenging due to time constraints and class sizes, for example. AI may therefore be useful in adapting to different learning styles and giving feedback in a class setting, so that those facilitating the sessions can focus more on specific questions from students. From experience as a foundation doctor in a DGH, and anecdotal discussion with foundation doctors at tertiary centres, the diversity of radiology cases at a DGH seems to be more limited. Therefore when rarer diagnoses present to foundation doctors at a DGH they may be more difficult to correctly identify. A framework such as ARIES could benefit smaller hospitals where junior doctors are less likely to see rarer cases by flagging up differentials they may not have considered, taking into account the clinical information provided as well. Moreover as a foundation doctor there is a lack of specific radiology training and curricula, therefore identifying one’s ‘unknown unknowns’ described in the Johari Window3 is more difficult. This is important because as already alluded to earlier interpretation of x rays, some CT scans and checking reports is a core part of on call as a foundation doctor. AI could have a useful role in developing junior doctors’ radiology knowledge if must-see cases and assessment were introduced as part of the core portfolio. AI could identify individual gaps in knowledge and learning needs doctors had been unaware of. Consequently in situations where formal radiology teaching is not provided or where clinicians have been unable to attend formal teaching due to on call commitments, AI integrated into the portfolio could allow foundation doctors to continue to develop their skills and focus on their weaker areas identified by AI. It must be acknowledged that the use of AI in education would need appropriate regulation and studies before introduction, especially given the potential impact on radiology as a specialty if AI systems provided inaccurate or substandard teaching to trainees. It would be paramount for future research to explore AI and precision education in radiology compared to current educational methods. If positive results were found it would be interesting for future studies to assess the effect of precision education on junior doctors in smaller hospitals compared to larger ones, and also on medical students with different course structures. 1Michael Tran Duong, 2Andreas M Rauschecker, 2Jeffrey D Rudie, 3Po-Hao Chen, 1Tessa S Cook, 4R. Nick Bryan and 1Suyash Mohan 1Department of Radiology, Perelman School of Medicine at the University of Pennsylvania, Philadelphia, Pennsylvania, USA 2Department of Radiology & Biomedical Imaging, University of California San Francisco, San Francisco, CA, USA 3Imaging Institute, Cleveland Clinic, Cleveland, OH, USA 4Department of Diagnostic Medicine, Dell Medical School, University of Texas at Austin, Austin, TX, USA We thank Dr Ellis for the proposed AI-augmented radiology education use cases for programs and institutions of the author’s home country. We are delighted by the shared excitement of the potential benefits that AI and integrated teaching files may bring to medical education. We agree that “precision” in AI-based precision education must encompass many levels, such that tools are tailored to individual regions, hospitals, programs, and trainees. Despite unequal access to certain cases due to frequency, geography or level of training, trainees at different stages and hospitals, from foundation doctors and registrars, to interns, residents and fellows, may one day benefit from AI through improved skills and greater knowledge. The goal of AI-augmented precision education is standardizing the level of knowledge while individualizing the delivery of that knowledge. To this end, AI will require extensive validation and regulation to ensure excellence in precision education. Similar to the Individualized Comparative Effectiveness of Models Optimizing Patient Safety and Resident Education (iCOMPARE) trials for standard and flexible duty hours for internal medicine programs,1,2 studies of AI in radiology education will require well-defined primary endpoints and standardized measures of success. Once AI programs have been developed “preclinically,” residency, foundation and registrar programs worldwide could conduct studies where trainees are randomized into multi year long subprograms with either standard or AI-augmented learning. Outcome measures to demonstrate superiority or non-inferiority of AI-augmented precision education could include patient safety outcomes (mortality, adverse effects from imaging, etc) and/or trainee performance on standardized tests of diagnostic accuracy. Medicine is context-dependent and ever-evolving. Similar to revising editions of a textbook, AI educational systems would need to be continually fine-tuned for modern technological advances, new medical knowledge, and changing patient populations. To enable physicians to understand the output of AI enough to learn from AI and then update AI, the AI must be relatively “explainable”.3 It may be necessary for trainees to consult multiple AI systems when perplexing cases arise. We also wish to emphasize that the intent of such AI programs is to benefit human students and allow human educators to focus on what they do best. Effective AI will empower rather than scrutinize trainees. Instead of trainees being told to “read more,” they can target their reading with cases. Effective AI will highlight certain diagnostic features and present relevant cases, and educators can reinforce certain strategies, pearls and pitfalls that are often difficult to learn through other methods. Thus, we envision AI systems as support, not surveillance, to trainees and as supplement, not replacement, to teachers as medicine moves towards precision education.
No takes yet. Share an insight, caveat, or question.
Liddy Ellis (2019) studied this question.
Synapse has enriched 5 related papers. Consider them for comparative context: