计算机应用研究2025，Vol.42Issue(10)：2956-2963,8.DOI:10.19734/j.issn.1001-3695.2025.04.0089
基于提示引导多跳推理的医学诊断检索增强生成
Retrieval-augmented generation with prompt-guided multi-hop reasoning for medical diagnosis
摘要
Abstract
The complexity of medical diagnosis tasks is particularly prominent in the manifestations of symptoms and their as-sociations with diseases.Due to the complexities of"different symptoms for the same disease"and"same symptoms for diffe-rent diseases",medical diagnosis tasks impose higher requirements on the reasoning capabilities of models.Traditional RAG technologies,with their static retrieval and single-step reasoning,struggle to capture multi-level logical relationships(such as symptoms → departments → diseases → differential diagnosis).To effectively overcome this limitation,this study proposed a novel framework for the medical field:PGM-RAG.By integrating basic knowledge of the medical field,this framework provi-ded clear reasoning guidance for the model through the design of prompt information for each reasoning step.Meanwhile,this framework designed a quantitative rewriting mechanism around strictly control the accuracy of the content generated by large language models,thereby enhancing the reliability of the reasoning process and the precision of diagnostic results.Experiments on two public medical datasets,Huatuo-26M and WebMedQA,show that the proposed model outperforms the existing best methods by 12.6％and 8.9％in EM and F1 metrics,respectively.Ablation experiments demonstrate that the multi-hop rea-soning chain and the quantitative rewriting mechanism significantly improve the model's performance.关键词
动态检索增强生成/多跳思维链/医学提示学习/量化重写机制/医学诊断Key words
dynamic retrieval augmented generation/multi hop chain-of-thought/medical prompt learning/quantitative re-writing mechanism/medical diagnosis分类
信息技术与安全科学引用本文复制引用
秦乐,勾智楠,王培伍,张高飞,刘思雨,高凯..基于提示引导多跳推理的医学诊断检索增强生成[J].计算机应用研究,2025,42(10):2956-2963,8.基金项目
河北省自然科学基金资助项目(F2023207003) （F2023207003）
河北省高等教育教学改革研究与实践项目(2023GJJG187) （2023GJJG187）
河北经贸大学教学研究项目(2024JYQ09) （2024JYQ09）
