Answers are grounded in this paper and cite their sources.
BACKGROUND: Narrative feedback quality in surgical Entrustable Professional Activity (EPA) assessments is variable, limiting consistent evaluation and opportunities for improvement. We evaluated whether fine-tuned large language models (LLMs) could reproduce expert-assigned feedback quality scores and whether a simplified scoring rubric improved assessment consistency. STUDY DESIGN: This observational study evaluated 1,509 narrative comments from EPA micro-assessments in a general surgery residency program during the 2023-2025 academic years. Two surgical education experts scored feedback using the Quality of Assessment for Learning (QuAL) rubric, which assesses evidence, suggestion, and connection on a 0-5-point scale. A modified rubric (mQuAL) assigned 1 point per domain on a 3-point scale. Fine-tuned encoder and generative LLMs were evaluated using accuracy and F1 score against human-assigned scores. RESULTS: Overall interrater reliability was high (κ = 0.85), although raters disagreed on 27.8% of QuAL evidence scores. Binary evidence scoring with mQuAL reduced disagreement to 8.5%. Encoder and generative models achieved similar accuracy for suggestion (0.96 versus 0.94) and connection (0.95 vs 0.91), whereas the encoder outperformed the generative model for evidence (0.93 vs 0.77). Model performance was strongest with mQuAL. CONCLUSIONS: Fine-tuned LLMs reproduced expert assessments of surgical narrative feedback with high domain-specific accuracy, particularly using encoder models. Simplifying evidence scoring improved interrater agreement and model performance, supporting further evaluation of AI-assisted feedback quality assessment within surgical residency programs.
No takes yet. Share an insight, caveat, or question.
Kabbash et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: