基于RoBERTa的BiLSTM-CRF模型在中文病案实体识别中的应用研究

Research on application of RoBERTa-based BiLSTM-CRF model in entity recognition of Chinese medical record

目的:构建适用于中文电子病案审核的命名实体识别模型,提高医院统计部门相关病案审核的工作效率,为人工智能技术在医疗卫生行业的应用奠定基础。方法:基于1 700条真实中文电子病历,选取“症状体征”“疾病诊断”“治疗方式”“解剖部位”“影像检查”“手术”等作为主要实体,结合经人工审核的病案结果进行BIOES标注,基于RoBERTa的BiLSTM-CRF算法,构建中文病案实体识别模型。结果:所建实体识别模型在CCKS2017中准确率为94.80%,召回率为96.03%,F1值为0.95;在CCKS2019中准确率为81.91%,召回率为83.03%,F1值为0.83。与传统基于Word2Vec、BERT的识别模型相比,RoBERTa-BiLSTM-CRF模型具有更优的效果。结论:基于RoBERTa的BiLSTM-CRF模型在中文电子病案实体识别中效果良好,能够对中文电子病历进行有效的初步识别和筛查,提高相关统计人员的审核效率。

Objective To build a named entity recognition model applicable to the audit of Chinese electronic medical record, improve the efficiency of medical record audit in hospital statistics departments and to lay the foundation for the application of artificial intelligence technology in healthcare industry. Methods Based on 1,700 real Chinese electronic medical record, "symptoms and signs" "disease diagnosis" "treatment method" "anatomical site" "imaging examination" and "surgery" were selected as the main entities, and BIOES labeling was completed with the results of manually reviewed medical record. Based on BiLSTM-CRF algorithm of RoBERTa, the entity recognition model of Chinese medical record was built. Results In CCKS2017, the accuracy of the proposed entity recognition model was 94.80%, the recall rate was 96.03% and F1 value was 0.95;in CCKS2019, the accuracy was 81.91%, the recall rate was 83.03% and the F1 value was 0.83. Compared with the traditional recognition models based on Word2Vec and BERT,the RoBERTa-BiLSTM-CRF model has better results. Conclusion The RoBERTa-based BiLSTM-CRF model works well in entity recognition of Chinese electronic medical record, which can effectively perform preliminary recognition and screening of Chinese electronic medical record, and improve the audit efficiency of relevant statisticians.