Generative AI-Assisted Simulation Training Is Associated with Higher Post-Intervention Diagnostic Communication Scores Across Type 2 Diabetes, Obesity, and Breast Cancer Scenarios.

Background

Diagnostic communication influences patient understanding, adherence, and shared decision-making in high-burden cardiometabolic disease and high-stakes oncologic care. However, scalable training models that allow standardized, repeatable practice and competency benchmarking remain limited. This study examined whether undergraduate medical students demonstrated higher diagnostic communication scores after completing a structured generative artificial intelligence (AI)-assisted simulation program across three clinically distinct diagnostic disclosure scenarios.

Methods

We conducted a prospective, single-arm, pre-post educational study in undergraduate medical students completing AI-assisted diagnostic communication training across T2DM, obesity, and breast cancer scenarios. Students underwent baseline in-person assessments with standardized human simulated patients, completed 10 asynchronous AI-assisted encounters per scenario using standardized scenario-specific prompts and automated feedback, and then completed post-intervention in-person assessments. Scenario order was randomized. Performance was scored live by two physician raters using an adapted 24-item, eight-domain rubric. Cross-scenario analyses included three-scenario completers (n = 56; scenario-specific paired samples up to n = 77). Without a control group, analyses were interpreted as within-student pre-post associations rather than causal effects.

Results

Students demonstrated higher post-test total rubric scores across all scenarios. Mean (SD) within-student changes were +24.26 (25.05) for T2DM, +26.17 (20.67) for obesity, and +36.31 (17.70) for breast cancer. Positive pre-post changes were observed across communication domains, with variation by clinical context. Exploratory analyses suggested limited cross-scenario gain-score associations and heterogeneous response patterns.

Conclusions

Generative AI-assisted simulation was associated with higher post-intervention diagnostic communication scores across three diagnostic disclosure scenarios. The single-arm design precludes causal attribution and does not exclude testing effects, rubric familiarization, maturation, or concurrent clinical learning. Controlled studies are needed to determine its comparative educational value.