[2L1-OS-25] OS-25

[2L1-OS-25-04] Performance Comparison and Evaluation of Few-shot Example Selection for Generating Background, Causes, and Improvement Measures of Medical Incidents

〇Yuna Haseyama¹, Tomoki Ito², Hiroki Sakaji¹, Itsuki Noda¹ (1. Hokkaido University, 2. MITSUI & CO., LTD.)

Keywords:Benchmark Development, Large language model

In recent years, research has been conducted on methods for sample selection for few-shot learning, but there has been limited research related to medical incidents and patient safety. This study utilizes the Japanese Medical Incident Dataset (JMID), a dataset of medical incidents and near-miss cases collected and provided by the Japan Council for Quality Health Care and evaluates the generated results using BERTScore. Near-miss cases (referred to as ”hiyari-hatto” in Japanese) are incidents where accidents were narrowly avoided but could have potentially occurred. The JMID contains descriptions of various types of medical incidents and near-miss cases. In this study, these cases are classified into 18 categories based on similar content. In this paper, while presenting the results of zero-shot learning as a baseline, few-shot learning with similar cases and few-shot learning with randomly selected cases are compared and analyzed.