特征选择技术在江西地区缺血性脑卒中合并肺部感染风险预测模型中的应用  

Application of feature selection technique in risk prediction model of ischemic stroke complicated with pulmonary infection in Jiangxi Province

作　　者：罗颢文 涂江龙[2] 刘建模 俞鹏飞 葛艳秋 易应萍[1] LUO Hao-wen;TU Jiang-long;LIU Jian-mo;YU Peng-fei;GE Yan-qiu;YI Ying-ping(Department of Information,Second Affiliated Hospital of Nanchang University,Nanchang,Jiangxi 330006,China;不详)

机构地区：[1]南昌大学第二附属医院信息处,江西南昌330006 [2]南昌大学第二附属医院神经内科,江西南昌330006 [3]南昌大学医学部公共卫生学院,江西南昌330006

出　　处：《现代预防医学》2020年第22期4038-4041,4052,共5页Modern Preventive Medicine

基　　金：国家重点研发计划(2018YFC1312902);国家自然科学基金(81960609);江西省重点研发计划(2018ACH80004)。

摘　　要：目的采用不同特征选择技术构建基于机器学习的预测模型,探讨江西地区缺血性脑卒中患者发生肺部感染的风险因素,为江西地区缺血性脑卒中合并肺部感染的控制和预防提供参考。方法分别采用互信息、Lasso回归、决策树对特征进行筛选,比较XGboost、SVM、随机森林、MLP、logistic回归在缺血性脑卒中合并肺部感染模型中的效果。结果Lasso回归优于其他两种特征选择方法,筛选的侵入性操作、NIHSS评分、中性粒细胞计数等15个特征纳入最终模型,与其他分类算法相比,MLP分类性能最好,AUC与约登指数分别是0.8740(95%CI:0.8694~0.8804)和0.6267。结论Lasso回归可以限制多重共线性带来的影响,并输出高风险因素,结合MLP分类算法,能够较好的预测缺血性脑卒中患者是否会发生肺部感染,为其精准防控提供借鉴,具有一定的临床实践意义。Objective The prediction model based on machine learning was constructed by using different feature selection techniques,the risk factors of pulmonary infection in patients with ischemic stroke in Jiangxi Province were discussed,it provides reference for the control and prevention of ischemic stroke complicated with pulmonary infection in Jiangxi province.Methods Features were screened by mutual Information,lasso regression and decision tree,and the effects of XGboost,SVM,random forest,MLP and logistic regression in the model of ischemic stroke with pulmonary infection were compared.Results Lasso regression was better than the other two feature selection methods,and 15 features including invasive operation,NIHSS score and neutrophilic granulocyte count were selected into the final model.Compared with other classification algorithms,MLP had the best classification performance,the AUC and Youden index were 0.8740(95%CI:0.8694-0.8804)and 0.6267,respectively.Conclusion Lasso regression can limit the impact of multicollinearity,output the risk factors of the model,combined with the MLP classification algorithm,it can better predict will happen pulmonary infection in patients with cerebral ischemic stroke,provide reference for prevention and control of its precision,it has certain clinical practical significance.

关 键 词：缺血性脑卒中 肺部感染 Lasso回归 MLP 

分 类 号：R181.2[医药卫生—流行病学]