AI-Supported Adaptive Simulation for Diagnostic Disclosure in Medical Students: A Randomized Controlled Trial

Diagnostic disclosure is a complex communication task that requires learners to integrate interpersonal attunement, structured information delivery, and condition-specific reasoning in real time. We conducted a randomized controlled trial comparing conventional diagnostic communication training with the same training supplemented by an AI-supported adaptive virtual patient simulation designed to provide additional deliberate practice and individualized, just-in-time feedback. Eighty undergraduate medical students were randomized 1:1 and completed standardized-patient encounters involving disclosure of a new diagnosis of type 2 diabetes mellitus before and after training. Performance was assessed by blinded physician raters using an adapted Kalamazoo rubric. Among students with complete pre–post data (conventional training, n = 25; AI-supported training, n = 26), both groups showed substantial improvement. Mean gains were numerically larger in the AI-supported group, with small-to-moderate standardized effects across selected communication domains; however, baseline-adjusted group-by-time interactions did not reach conventional thresholds for statistical significance, indicating that any added mean effects beyond conventional training remain uncertain. Exploratory person-level analyses suggested greater heterogeneity of improvement in the AI-supported condition, including a higher density of large gains in higher-order communication components. These findings should therefore be interpreted as exploratory rather than confirmatory. AI-supported adaptive simulation appears feasible as an adjunct to communication training, but adequately powered studies are needed to clarify effect magnitude, mechanisms, and generalizability across training contexts.