Presentation Information
[4Yin-B-10]An Attempt to Classify the Strength of Improvement Measures for Medical Incident Cases Based on RCA2
〇Yuna Haseyama1, Tomoki Ito2, Tomohiro Kimura3, Hiroki Sakaji4, Itsuki Noda4, Shusaku Tsumoto3 (1. Graduate School of Information Science and Technology, Hokkaido University, 2. National Institute of Information and Communications Technology, 3. Shimane University Faculty of Medicine, 4. Research Facluty of Information Science and Technology, Hokkaido University)
Keywords:
Large Language Models,Patient Safety
Medical accident reports are essential for preventing recurrence and improving patient safety. In Root Cause Analysis (RCA) and its extension RCA² (Root Cause Analysis and Action), corrective actions are classified as weak, intermediate, or strong based on how much they rely on human effort versus systemic safeguards. However, Japanese medical accident reports do not currently include such classifications, requiring manual evaluation.
This study constructs a dataset of reports annotated with RCA²-based strength labels and examines automatic classification using ModernBERT, conventional machine learning, and generative models. ModernBERT achieved the highest accuracy. Generative models tended to overestimate action strength, while traditional machine learning struggled to classify strong actions due to limited training data.
This study constructs a dataset of reports annotated with RCA²-based strength labels and examines automatic classification using ModernBERT, conventional machine learning, and generative models. ModernBERT achieved the highest accuracy. Generative models tended to overestimate action strength, while traditional machine learning struggled to classify strong actions due to limited training data.