基于人工神经网络模型住院病案首页数据缺陷评估分类模型的构建

Construction of a Classification Model for Data Quality Defect Assessment of Medical Record Homepage based on Artificial Neural Network Model

目的 探讨通过构建人工神经网络模型分析问题住院病案首页的可行性。方法 从某院2018年1月1日-2020年12月31日归档病案中采用随机抽样法抽取3000份,其中2018年、2019年及2020年每年各提取1000份。统计分组后抽取住院病案首页存在问题的情况,通过运用自然语言处理技术从中文电子化病历中提取与住院病案首页相关的关键词语、短句,并应用循环-卷积神经网络(RCNN)模型对提取信息进行训练,构建质量缺陷分类模型。模型构建完成后对其使用情况进行评估,最后分析模型应用的效果。结果 纳入研究的3000份病案中,共计897份住院病案首页存在缺陷,缺陷率为29.9%,人工复检后发现问题主要包括诊断错误等8类;神经网络模型的准确率、敏感度、特异度、阳性及阴性预测值均显著高于人工复审,差异具有统计学意义(P<0.05);2020年采用人工网络模型分析后住院病案首页问题病案数量显著下降(P<0.05)。结论 RCNN模型运用于住院病案首页审核后能有效减少问题发生率,但仍需要扩大样本量来增加模型学习度。

Objectives This study aims to explore the feasibility of constructing an artificial neural network model to analyze the problematic medical record front pages. Methods From January 1, 2018 to December 31,2020, 3,000 medical records were selected by random sampling, of which 1,000 were extracted each year in 2018,2019, and 2020.Problems on the front pages of medical records were extracted after statistical grouping. Then, by using natural language processing technology to extract keywords and short sentences related to the medical record front pages from Chinese electronic medical records and using the recurrent-convolutional neural network(RCNN)model to train the extracted information, a quality defect classification model is constructed.After the model was built, its usage was evaluated, and the effect of the model application was analyzed.Results Among the 3000medical records included in the study, a total of 897 medical records had defects on the front page, with a defect rate of 29.9%. The problems found after manual re-examination mainly included 8 types of diagnostic errors.The accuracy, sensitivity, specificity, and positive and negative predictive values of the neural network model were significantly higher than those of the manual review, and the difference was statistically significant(P<0.05).In 2020, the number of medical records with problems on the front page dropped significantly after the analysis of the artificial network model(P<0.05).Conclusions The RCNN model can effectively reduce the problem rate of the front page of medical records after the front page of medical records is reviewed, but it is still necessary to expand the sample size to increase the learning degree of the model.