基于大语言模型的体检总检结论自动生成研究  

Automatic Generation of Healthcare Examination Summaries Using Large Language Models

作　　者：郑路程 李旭涛 徐敏[1] ZHENG Lucheng;LI Xutao;XU Min(Information Engineering College,Capital Normal University,Beijing 100048,China)

机构地区：[1]首都师范大学信息工程学院,北京100048

出　　处：《小型微型计算机系统》2024年第11期2569-2575,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(62177034)资助.

摘　　要：本文研究了基于大语言模型自动生成体检总检结论的方法.与常规文本摘要生成任务不同,体检总检结论的生成特别关注体检异常检查结果,要求生成结论不仅准确,还需遵循医学领域的专业知识和标准.为此,本文基于经医疗知识问答数据微调的大型预训练语言模型,提出了一个体检总检结论自动生成方法.该方法包括两个关键模块:1)异常信息抽取模块,利用少量标注数据增强模型在抽取科室小结中异常检查结果识别能力;2)结论项排序模块,使得生成内容符合体检总检结论的顺序规范.在真实体检数据集上的实验表明,这两个核心模块有效提升了总检结论生成质量.本文为医疗文档自动生成技术提供了新思路,展现了大语言模型在医疗人工智能应用中的前瞻性.This paper proposes a method for automatically generating comprehensive healthcare examination summaries based on large language models.Unlike conventional summary generation tasks,the generation of healthcare examination summaries focuses particularly on the abnormal results of the examination,demanding not only accuracy in the generated summaries but also adherence to professional knowledge and standards in the medical field.For this purpose,the paper presents a method for the automatic generation of comprehensive examination summaries,based on a large pre-trained language model fine-tuned with medical knowledge question-and-answer data.This method includes two key modules:1)an abnormal information extraction module,which enhances the model′s ability to identify abnormal information in the summaries of various examination departments using a small amount of annotated data;and 2)a summary item ordering module,ensuring that the generated content follows the standard sequence of comprehensive healthcare examination summaries.Experiments conducted on real healthcare examination datasets demonstrate that these two core modules effectively improve the quality of the generated comprehensive examination summaries.This study provides new insights into the technology for automatic medical document generation and showcases the potential and foresight of large models in the application of medical artificial intelligence.

关 键 词：体检科室小结 体检总检结论 大语言模型 异常文本抽取 结论项重排 医疗问答 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]