基于BERT特征的双向LSTM神经网络在中文电子病历输入推荐中的应用

目的:针对当前电子病历录入中的便捷性与规范性的双重需求,尝试利用神经网络算法来挖掘病历文本的语言习惯和承接关系,以便嵌入病历录入系统,提高医生输入的效率和质量。方法:设计了基于大数据的电子病历录入推荐工具,对骨创伤科病历文本进行清洗构建训练集后做特征编码,采用基于深度学习的BiLSTM网络算法,学习专科专病的语义信息。结果:分别基于BERT特征、独热编码、词向量三种文本表示,针对过往病历数据建立BiLSTM模型,预测下一句文本,结果表明使用BERT预训练模型特征的BiLSTM模型F1-score达到75.23%,且具有实际应用的价值。在专科专病文本推荐的场景下,BERT特征优于独热编码和Word2Vec词向量。

The Application of Bidirectional LSTM Neural Network Based on Features of BERT in the Entry Recommendation of Chinese Electronic Medical Records

Objective: To meet the demands of convenience and standardization of the current entry of electronic medical records, this paper tries to use the neural network algorithm to explore the language habits and associated relations of the medical record text, so as to embed into the medical record entry system and improve the efficiency and quality of entry by doctors. Methods: The recommendation tool for the entry of electronic medical records based on big data is designed. After cleaning of medical record text of Orthopedic Trauma Department and construction of the training set, the feature coding is completed. The BiLSTM network algorithm based on deep learning is used to learn the semantic information of specialized departments and diseases. Results: Based on BERT features, one-hot encoding and word vector, the BiLSTM model is established for the previous medical record data, and the next sentence of text is predicted. The results show that F1-score of BiLSTM model using features of BERT pre-training model is up to 75.23% and has the practical application value. In the scene of text recommendation for specialized departments and diseases, BERT features are better than one-hot encoding and Word2Vec word vector.