This study is the first of its kind to evaluate the impact of AI technologies in healthcare on medical education and clinical training by analysing perceptions of trainee doctors across a range of hospital and community-based specialities in London, UK. Overall, doctors perceive that clinical AI will have a positive impact on their training (58% agree). Domain-based analysis reveals more mixed perceptions. The overwhelming majority (82%) report insufficient training in AI topics, with strong support for formal training in these topics.
Domains of perceived positive impact
Doctors were most optimistic that clinical AI would improve training in research, audit and quality improvement. These are key education domains and can be challenging to fulfill without significant time and effort commitment outside of work. AI systems can improve the efficiency of research and audit by rapidly and accurately analysing large volumes of data. This may explain trainees’ positive perceptions in this regard [23]. An indirect but desirable effect is that clinical AI could free up doctors’ time to spend doing other educational activities, which was a common positive theme throughout.
Developing skills in evidence-based medicine is a key training requirement, but keeping up with rapidly changing clinical guidelines can be burdensome for doctors. AI-based decision support systems are automatically updated with the latest literature. Trainees perceived this to be a positive impact, enabling them greater exposure to the latest evidence to improve their knowledge and the quality of care they provide.
Domains with skeptical perceptions
Trainees perceived that clinical AI could reduce their training in practical skills, clinical judgement and decision-making. Developing these skills requires iterative practice, formation of heuristics, personal reflection, varied clinical experience and time. Participants reported that decision support systems, robotics and automatic image analysis could reduce training opportunities in these domains leading to deskilling. Another skeptical perception was increased administrative workload leading to information overload. Clinical AI developers should ensure that these technologies do not impede workflow to enable their adoption in clinical practice.
Medical educators should note that there are some areas in which training may be harmed by clinical AI. Involving clinicians in the development of these algorithms (such as ground-truth labelling and procedural training) will help trainees continue to develop these skills, because the AI will depend on their training to mimic behavior. This will also increase trust in the AI technologies and improve explainability to patients.
Impact on interpersonal and ethical development
The impact on interpersonal and ethical skills training featured in both positive and negative perception themes. Doctors envisage that AI will automate tasks, freeing up more time to develop communication skills. Trainees are optimistic that this will enhance their ability to provide patient-centred care. Conversely, doctors express concern that accountability is unclear when AI is part of clinical decisions, which could cause human deskilling in ownership and probity. Both points of view are valid; AI in healthcare has already created a new ethical landscape [24]. Governing bodies and medical educators should work collaboratively to produce ethical and legal frameworks that will protect and enable clinicians to develop these skills effectively in the age of clinical AI.
Perceptual variations by clinical specialism
Medical, surgical and community-based (General Practice, Psychiatry) specialities had a higher proportion of ‘AI optimistic’ trainees Fig. 3). This may be due to a lower clinical acuity in these specialities and higher administrative workload. Trainees’ in these specialities may perceive that AI will improve their workflow to free up time for training and educational activities such as communication skills development.
Acute specialities (Emergency Medicine, Acute Medicine, Anesthesia, Intensive Care) and child and maternal health (Paediatrics, Obstetrics and Gynaecology) had more ‘AI Negative’ and ‘Indifferent’ responses. This might be due to the higher clinical acuity in these specialities, including emergency procedures and diagnostic ambiguity, which may rely on experience or ‘gut-feeling’. These skills are notoriously hard to model for AI development [25]. This mirrors trainees’ perceptions that training in practical procedures and clinical judgement might be reduced by clinical AI.
Clinical Radiology trainees, although a small proportion of respondents, were highly optimistic. This echoes positive attitudes reported previously [26]. Radiology has experienced the most AI advances in clinical practice already [27, 28] so Radiology trainees are most likely to already first-hand experience of AI’s impact on their training. This may explain their positive perceptions compared to other specialities.
Overall, specialities with higher capacity for automation were more optimistic; the implication is that rather than a panacea, delivery of medical education in the AI age will need to be tailored to these subtle variations between specialities.
Implications for clinical curricula and recommendations for trainers
As clinical practice changes so must clinical education. This study brings the need for formal AI training in clinical curricula into sharp relief, and confirms a willing appetite for this from trainee doctors. The impact of AI on medical education can take different routes. Direct routes leverage AI technology to improve the delivery of training itself. Indirect routes benefit education by streamlining workflow to free up more time for education and training. Although the majority (72%) of respondents in our survey were yet to regularly encounter AI systems in their training and education, it is an area of active research ranging from assisted radiology teaching [29] to virtual reality for surgical skills development [30] and automated assessment of procedural performance [31]. Medical curricula should be reviewed to leverage these technologies to directly improve the delivery of clinical education.
We propose that medical curriculum makers consider a new set of AI-specific skills. These include data input and management, mathematics and statistics, communicating AI outcomes to patients and AI-specific ethics. Medical training curricula are already saturated with limited room for new topics so practical training in ‘Applied AI’ would be most feasible. Alongside an overview of common ML architectures, this should include balanced training in clinical AI interpretation including data bias, overfitting and the potential for harm.
Navigating the current ML research landscape is challenging. Common pitfalls include over-optimistic conclusions from ‘human versus clinician’ studies that are usually retrospective and prone to bias [32], lack of standardized benchmarks and no universally accepted AI evaluation metrics [33]. Training in ‘Applied AI’ must additionally equip clinicians with skills in critical literature analysis.
Training in ‘Applied AI’ will need to be supported by e-learning, didactic teaching, assessment in examinations, supervised clinical learning events and personal reflection (Fig. 5). Although this has been considered as a priority for the future health workforce [2, 3], ‘Applied AI’ topics remain widely absent from clinical training curricula [9, 19]. Ultimately this will negatively affect the quality of patient care by missing out on the myriad benefits of clinical AI systems.
Our survey was not prefaced by any educational material on AI, for two reasons. First, it avoided prolonging the time participants would need to commit, which would have reduced response rates. Second, it would have biased the responses towards taking the point of view implicitly expressed in the educational material. Even if we had striven for neutrality, it would be difficult to achieve it and even more difficult to document that we had achieved it.
The optimistic perceptions of AI reported in this study may lessen as doctors gain more first-hand experience of AI in their clinical practice and training (including its weaknesses such as data bias and the black-box effect) [20]. Conversely, the opposite may occur; indeed in this study the participants working in the speciality most likely to have already experienced AI technologies (Radiology) were actually the most ‘AI optimistic’. Curriculum development must provide a balanced view, recognizing AI limitations, weaknesses and potential for harm.
Limitations
This study provides a snapshot of the opinions of trainee doctors working in the UK NHS. The survey was hosted primarily by postgraduate training centers in London (UK), where research team members were based to maximise response rate (72%). Although these centers administrate postgraduate training for hospital and community based trainees (enabling representation of trainees in General Practice, Psychiatry etc.), the results of our survey could be biased in favor of trainees working in London, who may have specific experiences due to local uptake of AI technologies in urban compared to rural areas. Based on our results, we recommend a survey of all UK NHS postgraduate centers to gain a cross-sectional representation of trainee experience, and to elucidate any geographical variations in experience and opinion. We also recommend the inclusion of medical undergraduate students, since they are the clinical workforce of the future, and most likely to be directly affected by the impact of AI technologies on medical training.
Participants’ level of AI knowledge was not collected in our survey. Participants’ role at the time of responding was collected, with only 2 respondents being currently involved with clinical AI research. The majority of respondents neither encountered AI technologies regularly in their clinical practice (68%) or training (72%). Future work would benefit from participants self-rating their level of AI knowledge to contextualize findings.
The impact on communication and interpersonal skills was not assessed in the Likert-type part of our survey, Thematic analysis of free-test responses revealed important trainee perceptions in these areas. Future evaluation of trainees’ attitudes should further probe the perceived impact on domains such as communication, professionalism, leadership and probity, which are key elements of all clinical training curricula.