心脏超声报告文本结构化研究  

Research on text structuring of cardiac ultrasound report

作　　者：曹明源 贾真[3] 路强 张瀚博 田川 李运明 杨孝光 CAO Mingyuan;JIA Zhen;LU Qiang;ZHANG Hanbo;TIAN Chuan;LI Yunming;YANG Xiaoguang(Office of Medical Information and Data,Medical Support Center,the General Hospital of Western Theater Command of the PLA,Chengdu 610083,Sichuan Province,China;Department of Information,Medical Support Center,the General Hospital of Western Theater Command of the PLA,Chengdu 610083,Sichuan Province,China;School of Computer and Artificial Intelligence,Southwest Jiaotong University)

机构地区：[1]中国人民解放军西部战区总医院医疗保障中心医学信息数据室,成都610083 [2]中国人民解放军西部战区总医院医疗保障中心信息科,成都610083 [3]西南交通大学计算机与人工智能学院

出　　处：《中国数字医学》2025年第3期53-60,共8页China Digital Medicine

基　　金：全军保健专项科研课题(21BJZ39);西部战区总医院院管课题(2021-XZYG-A14)。

摘　　要：目的:探讨自然语言处理方法用于心脏超声报告文本提取的可行性,并构建心脏超声文本标准化数据库。方法:根据心脏超声报告内容,将提取数据元分为描述类和数值类数据元,利用TPLinker模型抽取描述类数据元,并利用基于规则的模式匹配方法抽取数值类数据元。结果:抽取976例心脏超声报告,模式匹配方法抽取数值类数据元的准确率和召回率均高于95%。TPLinker模型的F1值为88.62%,并横向对比CasRel、SpERT、CopyRE关系抽取模型,发现对于提取描述类数据元,TPLinker模型在准确率和召回率评价上优于其他关系抽取模型。结论:TPLinker模型结合基于规则的模式匹配抽取方法可用于心脏超声报告文本信息的结构化提取。Objective To explore the feasibility of natural language processing methods for text extraction of cardiac ultrasound reports,and to build a standardized database of cardiac ultrasound texts.Methods According to the content of cardiac ultrasound reports,the extracted data elements were divided into descriptive data elements and numerical data elements.The descriptive data elements were extracted by using the TPLinker model,and the numerical data elements were extracted by using the rule-based pattern matching method.Results 976 cases of cardiac ultrasound reports were extracted,and the accuracy and recall rate of numerical data elements extracted by pattern matching method were higher than 95%.The F1-score of the TPLinker model was 88.62%,and the horizontal comparison of the CasRel,SpERT,and CopyRE relational extraction models revealed that,the TPLinker model was superior to the other relational extraction models in the evaluation of the accuracy and recall rate for extracting descriptive data elements.Conclusion TPLinker model applied with rule-based pattern matching extraction method can be used for structured extraction of text information in cardiac ultrasound reports.

关 键 词：心脏超声检查报告 模式匹配 TPLinker模型 自然语言处理 

分 类 号：R540.45[医药卫生—心血管疾病] R319[医药卫生—内科学]