基于BERT多层网络的医疗实体抽取模型研究

Research on Medical Entity Extraction Model Based on the BERT Multi-layer Network

电子病历是由医生根据病人描述和检查结果进行推断总结出来的,是以非结构化文本形式进行存储和管理,是医疗信息化的核心数据资产,其基本信息单元是医疗实体。传统的实体识别方法是基于规则、词典机器学习的方法,这些在性能、效率和准确度上难以满足医疗信息化的发展需求。本文提出基于BERT的多层网络模型,简称为BBC,并将其应用于克拉玛依市中心医院电子病历信息抽取中,提取腹部超声检查结果中的症状实体。实验结果表明,本文提出的模型显著优于现有的方法,实体预测的F1值提升了2.3%。

Electronic medical records are deduced and summarized by doctors based on patient description and examination results. They are stored and managed in the form of unstructured text. It is the core data asset of medical informatization, and its basic information unit is medical entity. The traditional entity recognition method is based on rules and dictionary machine learning, which is difficult to meet the development needs of medical informatization in terms of performance, efficiency and accuracy. This paper proposes a multi-layer network model based on BERT, referred to as BBC, and applies it to the electronic medical records information extraction of Karamay Central Hospital to extract symptom entities in the results of abdominal ultrasound examination. The experimental results show that the model proposed in this paper is significantly better than the existing methods, and the F1 value of entity prediction is increased by 2.3%.