基于电子病历的预训练模型在急性呼吸道感染识别中的应用

Application of pretrained model based on electronic medical record in recognition of acute respiratory infection

目的评价基于电子病历的预训练模型对急性呼吸道感染(ARI)的识别效果。方法收集重庆大学附属三峡医院2021年12月1至31日就诊于呼吸与危重症科、发热门诊、急诊门诊、儿科、感染科的所有患者的病历资料。使用基于Transformer的双向编码器表征(BERT)预训练模型对病历进行ARI病例识别与判断,以医学专业人员根据ARI病例定义判断的结果作为“金标准”,计算模型识别ARI的灵敏度、特异度、与“金标准”的一致性及受试者工作特征曲线下面积(AUC),评价模型对ARI病例识别的准确性。结果含有3817条电子医疗记录的测试集中,共有1200条经人工判定的ARI病历。模型判定ARI共计1205例,灵敏度为92.67%(1112/1200),特异度96.45%(2524/2617),与“金标准”的一致性Kappa值为0.89,AUC为0.95。模型在男性和女性中识别ARI的准确性相近(AUC分别为0.95和0.94),且对未成年ARI病例识别较18~59岁及≥60岁(AUC分别为0.94,0.89和0.89)更准确。相较于住院患者,目前模型对门诊患者的ARI识别更好,AUC分别为0.74和0.95。结论使用基于电子病历的预训练模型对ARI病例判定具有良好的效果,特别是针对门诊患者及未成年患者。该模型在利用电子病历进行医疗机构ARI病例监测方面表现出良好的潜力。

Objective To evaluate the recognition of acute respiratory infection(ARI)by a pretrained model based on electronic medical records(EMRs).Methods 38581 EMRs were obtained from Chongqing University Three Gorges Hospital in December 2021.Bidirectional encoder representation from transformers(BERT)pretrained model was used to identify ARI in EMRs.The results of medical professionals were considered as the gold standard to calculate the sensitivity,specificity,Kappa value,and area under the curve of the receiver operating characteristic(AUC).Results There were 3817 EMRs in the test set,with 1200 ARIs.A total of 1205 cases were determined as ARI by the model,with a sensitivity of 92.67%(1112/1200)and a specificity of 96.45%(2524/2617).The model identified ARI with similar accuracy in males and females(AUCs 0.95 and 0.94,respectively),and was more accurate in identifying ARI cases in those aged less than 18 than in adults 18-59 and adults 60 and older(AUCs 0.94,0.89 and 0.94,respectively).The current model had a better identification of ARIs in outpatient patients than that in hospitalized patients,with AUCs of 0.74 and 0.95,respectively.Conclusion The use of the BERT pretrained model based on EMRs has a good performance in the recognition of ARI cases,especially for the outpatients and juveniles.It shows a great potential to be applied to the monitoring of ARI cases in medical institutions.