基于大语言模型的医疗问答算法研究  

Research on Medical Question Answering Algorithm based on Large Language Model

作　　者：宋雪 李楠[1] 郑涛 郑兵 SONG Xue;LI Nan;ZHENG Tao;ZHENG Bing(Information Center of West China Hospital of Sichuan University,Chengdu 610041,Sichuan,China)

机构地区：[1]四川大学华西医院信息中心,四川省成都市610041

出　　处：《中国卫生信息管理杂志》2025年第2期269-274,290,共7页Chinese Journal of Health Informatics and Management

基　　金：四川省重大科技专项揭榜挂帅项目“医学诊疗大模型关键技术研发及应用示范”(2024ZDZX0017)。

摘　　要：目的研究基于大语言模型的医疗问答算法,提升医学智能问答效果。方法设计以E-Transformer为基础模块的医疗问答算法MedLLM,通过大量通识数据的预训练和医疗数据的监督微调,提升大语言模型在医疗领域的认知和问答能力。结果MedLLM算法在中文医疗问答数据集cMedQA2上进行测试,测试结果相较于现有医疗问答算法存在明显提升,同时在预测结果的相关性、正确性和完整性上实现了与医生相近的医疗诊断和建议。结论基于大语言模型的医疗问答算法可实现快速学习医疗领域知识并对问题展开回答,有效提升医学智能问答的准确性。Objective To study medical question answering algorithm based on large language model to improve the effectiveness of medical intelligent question answering.Methods A medical question answering algorithm MedLLM based on E-Transformer module was designed.Through the pre-training of a large number of general data and the supervised fine-tuning of medical data,the cognitive and question answering ability of large language model in the medical field can be improved.Results The MedLLM algorithm was tested on the Chinese medical Q&A dataset cMedQA2,and the test results were significantly improved compared with the existing medical Q&A algorithm.At the same time,the correlation,correctness and completeness of the predicted results achieved medical diagnosis and suggestions similar to those of doctors.Conclusion Medical question answering algorithm based on large language model can quickly learn medical knowledge and answer questions,which can effectively improve the accuracy of medical intelligent question answering.

关 键 词：大语言模型 TRANSFORMER 医疗问答 预训练 监督微调 算法研究 

分 类 号：R197.323[医药卫生—卫生事业管理] R319[医药卫生—公共卫生与预防医学]