Effectiveness of an AI-augmented teacher feedback model in improving medical students' clinical documentation skills: a retrospective cohort study.

BACKGROUND: Clinical documentation is central to evaluating medical students' clinical competence. However, traditional teacher feedback is often inefficient, inconsistent, or delayed. This study evaluates the effectiveness of an AI-augmented teacher feedback model in improving medical students' clinical documentation skills, in which AI-generated feedback was added to conventional teacher feedback. METHODS: Fifty medical interns participated in this retrospective cohort study (25 each in the AI-assisted enhancement group and a traditional group). After completing their medical records, the AI-assisted group received structured feedback reports generated by AI (DeepSeek V3), including dimension-specific scores and improvement suggestions, as well as teacher feedback. The traditional group received teacher feedback only. Thus, the intervention evaluated represents a combined AI-plus-teacher model, not an isolated test of AI. Both groups completed and were assessed on another medical record before the end of their rotation. All records were scored by an AI system according to the Admission Record Writing Quality Evaluation Standard (2024 Edition) in three phases: Phase 1-assessment of AI scoring reliability using the intraclass correlation coefficient (ICC) based on 20 records blindly evaluated by two experts, Phase 2-analysis of the correlation between post-rotation AI scores and Mini-Clinical Evaluation Exercise (Mini-CEX) scores, and Phase 3-ANCOVA for comparing post-rotation AI scores between groups, controlling for baseline AI scores and case complexity. RESULTS: AI scores demonstrated high consistency with expert ratings (ICC = 0.893) and a significant correlation with Mini-CEX scores (r = 0.579, p < 0.001). After controlling for baseline differences, the AI-assisted group achieved significantly higher post-rotation AI scores (11.15 points) than the traditional group (9.09 points), with a mean difference of 2.058 points (p = 0.041), representing a relative improvement of 22.6%. A subgroup analysis revealed a significant effect in low-complexity cases (p = 0.007) but not in medium-high complexity cases (p = 0.798). Sensitivity analyses showed that the main findings were not affected by covariate selection or sample exclusion. CONCLUSIONS: The generative AI scoring tool based on large language models was consistent with expert ratings. The combined AI-augmented teacher feedback model showed a preliminary association with higher post-rotation AI scores compared to traditional teacher feedback alone, particularly in low-complexity cases.