基于自然语言处理和深度学习的急性呼吸道传染病早期识别模型的构建

Construction of early detection model for acute respiratory infectious diseases based on natural language processing and deep learning

目的通过深度学习算法构建急性呼吸道传染病早期识别模型,协助开展医疗机构呼吸道传染病早期识别工作。方法收集2012年1月一2023年3月北京某大型三甲医疗机构急性呼吸道感染性疾病6683例患者病历文本数据,使用基于自然语言处理(NLP)技术的双向编码器表征(BERT)训练词向量,结合卷积神经网络(CNN)和双向长短期记忆神经网络(BiLSTM)构建早期识别模型BERT_MCB,以受试者工作曲线、准确率、精确率、召回率和F1值等指标作为模型性能判断标准。结果BERT_MCB模型整体优于随机森林、BERT、BERT_CNN、BERT_RNN四组基线模型,其中准确率提高了1.20%~15.80%、精确率提高了1.66%~23.69%、召回率提高了0.25%~26.75%、F1值提高了0.66%~27.25%。结论本研究建立的急性呼吸道传染病早期识别模型可较为准确地识别出急性呼吸道传染病,表明深度学习算法在早期识别急性呼吸道传染病方面具有较好的应用前景。

OBJECTIVE To develop an early detection model for acute respiratory infectious diseases using a deep learning algorithm,and to assist in the early identification of respiratory infectious diseases in medical institutions.METHODS Medical records of 6683 patients with acute respiratory infections from a large tertiary medical institution in Beijing from Jan 2012 to Mar 2023 were collected.We used the bidirectional encoder representations from transformers(BERT)based on natural language processing technology to train word vectors.Combining convolutional neural networks(CNN)and bi-directional long short-term memory(BiLSTM),we created an early detection model called BERT_MCB.Its performance was evaluated based on the receiver operating curve,accuracy,recall,and F1.RESULTS The BERT_MCB model was overall better than the random forest,BERT,BERT_CNN,and BERT_RNN models.The accuracy rate of this model increased by 1.20%-15.80%,precision rate increased by 1.66%-23.69%,recall rate increased by 0.25%-26.75%,and F1 value increased by 0.66%-27.25%.CONCLUSION The early detection model for acute respiratory infectious diseases can accurately identify acute respiratory infectious diseases,which showed that deep learning algorithms have promising potential in the early identification of acute respiratoryinfections.