摘要
目的构建住院病案首页的人工智能初筛模型并应用于分析,旨在为改善住院病案首页质量提供参考。方法随机抽取某院2020年1月1日-2021年8月31日期间出院病案5000份作为模型构建对象,按照7:3的比例将所有住院病案首页分为训练集(n=3500)和测试集(n=1500)。基于双向丰富语义的预训练语言模型(BERT)机制构建的BERT-迭代膨胀卷积神经网络(IDCNN)-多头注意力机制(MHA)-随机条件场(CRF)构建住院病案首页辅助分析模型,并对两个集合的住院病案首页进行分析。同时由病案室质控医师对训练集及测试集的住院病案首页进行初级质控,而后由5年工作经验以上的主治医师以上职称对住院病案首页信息进行复核质控,记录人工质控结果。最后将模型与人工分析结果进行一致性检验,验证BIMC模型的分析效果。结果训练集及验证集中BIMC模型与人工评估结果的符合率分别为93.00%(3255/3500)和90.73%(1361/1500);两个集合中R-CNN模型与人工评估结果具有较高的一致性[Kappa=0.921(训练)/0.915(验证),P均<0.001]。结论构建的BIMC模型对于住院病案首页分析效果与人工分析一致性较高,运用于住院病案首页分析的可行性较好,但仍需要进一步纳入更多住院病案首页提升模型对错误信息的识别准确性。
Objectives This study aims to construct an artificial intelligence preliminary screening model for the front page of inpatient medical records and apply it to analysis,so as to provide a reference for improving the quality of the front page of inpatient medical records.Methods A total of 5000 discharged medical records from a hospital from January 1,2020 to August 31,2021 were randomly selected as the object of model construction.The front pages of all inpatient medical records were divided into a training set(n=3500)and a test set(n=1500)according to a ratio of 7:3.The BERt-iterative expansive Convolutional neural network(IDCNN)-multi-head attention mechanism(MHA)-stochastic conditional field(CRF)was constructed based on the bidirectional rich semantic pre-trained language model(BERT)mechanism to construct the auxiliary analysis model of the front page of inpatient medical records and analyzed the front page of two sets of inpatient medical records.At the same time,the quality control doctor in the medical records room conducted primary quality control on the front page of the inpatient medical records in both the training set and the test set.Subsequently,the quality control information on the front page of the inpatient medical records was reviewed by the attending physician with more than 5 years of work experience,and the manual quality control results were recorded.Finally,the consistency test between the model and manual audit results was carried out to verify the audit effect of BIMC model.Results The coincidence rates of BIMC model and manual evaluation results in training set and validation set were 93.00%(3255/3500)and 90.73%(1361/1500).The R-CNN model in the two sets had a high consistency with the human evaluation results[Kappa=0.921(training)/0.915(validation),both P<0.001].Conclusions The constructed BIMC model has a high consistency with manual analysis in the analysis of the front page of inpatient medical records,and it is feasible to apply it to the analysis of the front page of inpatient medic
作者
买丽克·伊明
吴芳
Mailike Yiming;Wu Fang(Department of Medical,The Seventh Affiliated Hospital of Xinjiang Medical University,Urumqi 830028,Xinjiang Uygur Autonomous Region,China;不详)
出处
《中国病案》
2024年第7期18-20,共3页
Chinese Medical Record
基金
新疆维吾尔自治区自然科学基金项目(2021D01C057)。
关键词
人工智能模型
双向丰富语义的预训练语言模型
住院病案首页
疾病诊断相关分组
Artificial intelligence model
Bidirectional encoder representations from transformers
Front page of medical records
Diagnosis related groups
