基于三种机器学习方法的慢性阻塞性肺疾病人群早筛模型的建立与验证  

Establishment and verification of early screening model of chronic obstructive pulmonary disease based on three machine learning methods

作　　者：母应姣 王子云[1] 苏旭 李凌[2] 周婕[2] 王艺颖 刘涛[1,2] MU Ying-jiao;WANG Zi-yun;SU Xu;LI Ling;ZHOU Jie;WANG Yi-ying;LIU Tao(Key Laboratory of Environmental Pollution and Disease Monitoring,Ministry of Education,School of Public Health and Health,Guizhou Medical University Guiyang,Guiyang 561113,China;不详)

机构地区：[1]贵州医科大学公共卫生与健康学院,环境污染与疾病监控教育部重点实验室,贵州贵阳561113 [2]贵州省疾病预防控制中心,贵州贵阳550004

出　　处：《现代预防医学》2024年第9期1677-1683,共7页Modern Preventive Medicine

基　　金：基于医防融合的健康管理中心大数据平台研究与示范(黔科合支撑[2021]一般447);贵州省2019年中央补助地方重大疾病防治项目;贵州省卫生健康委省级重点建设学科项目。

摘　　要：目的构建慢性阻塞性肺疾病(chronic obstructive pulmonary disease,COPD)患者筛检模型。方法采用多阶段分层随机抽样的方法,抽取贵州省≥40岁的常住居民4587名,对其进行问卷调查、体格检查及肺功能检查。经过单因素分析初步筛选模型纳入变量,经多因素logistic回归确定最终纳入变量。分别应用logistic回归(logistic regression,LR)、随机森林(random forest,RF)、支持向量机(support vector machine,SVM)构建COPD患者筛检模型,使用受试者工作曲线下面积(area under the curve,AUC)评价模型效果。使用delong法检验模型之间AUC的差异。结果根据多因素logistic回归分析结果,本研究将年龄、14岁前经常咳嗽、哮喘、每日吸烟量(支)、烹饪燃料与排风、有害气体暴露6种因素纳入LR、RF、SVM模型。三种模型训练集AUC分别为73.64%、87.14%、73.30%,测试集AUC分别为76.10%、70.96%、76.08%,均具有较好的筛检效果。Delong法结果显示,三种模型的筛检效果在训练集与测试集均存在一定差异。结论本研究通过年龄、哮喘等6个简单变量建立经济、快捷且有效的COPD患者筛检模型。Objective To establish a screening model for patients with chronic obstructive pulmonary disease(COPD).Methods By using the method of multi-stage stratified random sampling,4587 permanent residents≥40 years old in Guizhou Province were investigated by questionnaire,physical examination,and pulmonary function examination.Variables to be included into the model were screened by univariate analysis and then further screened by multivariate Logistic regression.Logistic regression(LR),random forest(RF)and support vector machine(SVM)were used to construct the screening model of COPD patients,and the area under the curve(AUC)was used to evaluate the effect of the model.Delong method was used to test the difference of AUC between models.Results According to the results of multivariate Logistic regression analysis,age,frequent cough before 14 years old,asthma,daily smoking,cooking fuel and exhaust,and harmful gas exposure were included in LR,RF and SVM models.The AUC of the three model training sets were 73.64%,87.14%,and 73.30%,respectively,and the AUC of the test set were 76.10%,70.96%,and 76.08%,respectively,all of which had good screening results.The results of Delong method showed that the screening effects of the three models were different between the training set and the test set.Conclusion This study established an economical,rapid,and effective screening model for COPD patients through six simple variables such as age and asthma.

关 键 词：LOGISTIC回归 随机森林 支持向量机 慢性阻塞性肺疾病 筛检 

分 类 号：R563.9[医药卫生—呼吸系统] R18[医药卫生—内科学]