| 摘 要: 本研究针对心肌病诊断“多源信息过载”“知识更新滞后”及传统大模型“幻觉频发”等痛点，提出基于检索增强生成(RAG)的辅助诊断系统。以阜外医院2024年642份含5类心肌病分型的临床报告为数据集，整合国际指南、权威文献构建知识库，通过多阶段混合检索与精准提示工程关联患者信息与权威知识。实验显示，该系统在心脏超声(93.5%, 0.93)、核磁共振报告(87.2%, 0.87)诊断准确率和F1分数上显著优于主流模型，证实RAG架构有效性，为智能辅助诊断系统发展提供支撑。 |
| 关键词: 大语言模型 检索增强生成 心肌病 辅助诊断 临床决策支持 |
|
中图分类号: TP391
文献标识码:
|
| 基金项目: 福建省自然科学基金项目（2023J01138）、宁波大学One health交叉学科研究项目（HY202409） |
|
| Research on Clinical Auxiliary Diagnosis of Cardiomyopathy Using Retrieval-Augmented Large Language Models |
|
Yang ChangJi1, Zhang XuanDe1, He HuiXin2, Liu Jie3, Duan YanFeng4, Huang Xin5
|
1.School of Electronic Information and Artificial Intelligence, ShaanxiSUniversitySofSScienceS STechnology;2.School of Computer Science and Technology, Huaqiao University;3.FUWAI Hospital, CAMS PUMC;4.Zhongguancun Xirui Institute for Precision Study of Heart, Brain and Tumor;5.School of Information Science and Engineering, Ningbo University
|
| Abstract: This study addresses the pain points of cardiomyopathy diagnosis, such as "multisource information overload," "lagging knowledge updates," and the frequent "hallucinations" of traditional large language models, by proposing a retrieval-augmented generation (RAG)-based auxiliary diagnosis system. Using a dataset of 642 clinical reports from Fuwai Hospital in 2024, which includes five types of cardiomyopathy classifications, the study integrates international guidelines and authoritative literature to build a knowledge base. It then links patient information with authoritative knowledge through multi-stage hybrid retrieval and precise prompt engineering. Experimental results show that the system significantly outperforms mainstream models in diagnostic accuracy and F1 score on echocardiography reports (93.5%, 0.93) and cardiac magnetic resonance (MRI) reports (87.2%, 0.87). This confirms the effectiveness of the RAG architecture and provides support for the development of intelligent auxiliary diagnosis systems. |
| Keywords: Large Language Models, Retrieval-Augmented Generation, Cardiomyopathy, Auxiliary Diagnosis, Clinical Decision Support |
