Artificial Intelligence (AI), Clinical Decision-making, Clinical Reasoning, Medical Education
Conditions
Keywords
artificial intelligence, medical education, clinical reasoning, clinical decision-making, human-computer interaction
Brief summary
The purpose of the TEACH-AI study is to assess whether a brief, structured workshop on artificial intelligence can improve the performance of medicine doctors in training (i.e. residents) in their diagnostic and management reasoning. In this multi-site randomized controlled trial, internal medicine and family medicine residents are assigned either to receive an in-person workshop on safe, effective LLM use before a standardized AI-assisted assessment, or to complete the same assessment before receiving the workshop. Residents will review clinical cases that are fully synthetic, no protected health information is used, using a password protected LLM interface.
Detailed description
Large language models (LLMs) have rapidly entered routine use in medical education and clinical practice. Prior randomized trials have shown that, although LLMs can outperform individual clinicians on some reasoning benchmarks, providing physicians with access to an LLM does not necessarily improve their diagnostic or management performance, and LLMs alone may perform better than physician-plus-LLM teams. These findings suggest that human-AI collaboration is fraught, non-trivial and may require explicit training. TEACH-AI is a pragmatic, multi-site, two-arm randomized controlled trial embedded within protected residency didactic time. The primary objective is to determine whether a single in-person workshop on the basics of LLMs and best practices of prompting/verification strategies improves residents' performance on an AI-assisted assessment, compared with residents who complete the simulation before receiving the workshop. All participants will ultimately receive the same workshop and the same assessment (either workshop-first vs assessment-first). The simulation consists of multiple fully synthetic vignettes delivered via a password protected LLM interface. Residents interact freely with the LLM using natural-language prompts and then submit structured final responses regarding aspects such as leading diagnosis, differential, management plan, and/or justification. The platform will record prompts, model outputs, final answers, and timing. Vignette scoring combines correctness of the final diagnosis or management plan. Scoring is conducted by blinded faculty using standardized rubrics and then any discrepancies will be resolved through multiple rounds of discussions. The trial will enroll up to 200 residents across four ACGME-accredited programs (internal medicine at Beth Israel Deaconess Medical Center, Stanford University, and Cambridge Health Alliance; family medicine at AdventHealth Orlando).
Interventions
OTHERWorkshop-First
Participants will attend a workshop during their didactic time that will review various aspects of large language models including fundamentals and best practices of interacting and interpreting their outputs.
Sponsors
Beth Israel Deaconess Medical Center
Stanford University
Cambridge Health Alliance
AdventHealth
Study design
Allocation
RANDOMIZED
Intervention model
PARALLEL
Primary purpose
DIAGNOSTIC
Masking
SINGLE (Outcomes Assessor)
Masking description
The grading of responses will be performed by assessors blinded to participant identity and whether workshop-first or assessment-first assignment.
Intervention model description
The trial will be designed as a randomized, two-arm, single-blind parallel group study.
Eligibility
Sex/Gender
ALL
Healthy volunteers
Yes
Inclusion criteria
* Participants must be licensed physicians and have started at least post-graduate year 1 (PGY1) of medical training. * Training in Internal medicine or family medicine or emergency medicine. * Able to provide informed consent and complete assessments in English.
Exclusion criteria
* Not currently practicing clinically. * Resident directly involved in the design of the study, development of the workshop, or creation/piloting of the assessment vignettes. * Resident who declines or withdraws consent.
Design outcomes
Primary
| Measure | Time frame | Description |
|---|---|---|
| Total Score on Expert-Development Rubrics | Within 24 hours of assessment completion. | The primary outcome will be the number of correct responses for all cases using select questions from expert-developed scoring rubrics. The rubrics were developed using a Delphi consensus process by expert physicians as established by Goh et al (Nature Medicine, 2025; DOI: 10.1038/s41591-024-03456-y). The primary outcome will be analyzed at the case level, comparing performance between the randomized study groups, with a higher number of correct responses indicating a better outcome. |
Secondary
| Measure | Time frame | Description |
|---|---|---|
| Time Spent on Management | Within 24 hours of assessment completion. | Comparison of time spent per case between the two study groups. |
| Prompt Count | Within 24 hours of assessment completion. | Comparison of the number of resident-initiated prompts to the LLM per case between the two study groups. |
| Management Reasoning Using Expert-Derived Rubrics | Within 24 hours of assessment completion. | Comparison of management reasoning accuracy based on the number of correct responses per case between the two groups on expert-developed scoring rubrics. The rubrics were developed using a Delphi consensus process by expert physicians as established by Goh et al (Nature Medicine, 2025; DOI: 10.1038/s41591-024-03456-y). The primary outcome will be analyzed at the case level, comparing performance between the randomized study groups with more correct responses indicating a better outcome. |
| Diagnostic Reasoning | Within 24 hours of assessment completion. | Comparison of diagnostic reasoning accuracy based on the number of correct responses per case between the two groups. |
Countries
United States
Contacts
CONTACTJacob M Koshy, MD, MPH
Outcome results
None listed