基于决策树法和Logistic回归预测神经外科老年住院患者医院感染风险的研究

Prediction of risk of nosocomial infection in elderly hospitalized patients of neurosurgery department based on decision tree model and logistic regression model

目的应用决策树卡方自动交互检测(CHAID)算法和二分类Logistic回归分析法分别构建神经外科老年住院患者医院感染风险预测模型,并对模型的预测结果做对比分析。方法回顾性分析2018年1月-2019年6月海南省人民医院神经外科≥60岁老年住院患者,应用CHAID算法和Logistic回归分析法分别建立风险预测模型,通过受试者工作特征曲线(ROC)的曲线下面积(AUC)对两种模型的预测效果进行对比评价。结果共收集患者1 111人,其中医院感染131人,感染发病率11.79%;CHAID法和Logistic回归分析法均显示住院天数≥31 d、使用呼吸机、泌尿道插管是医院感染发生的重要影响因素;决策树模型风险预测的正确率为88.2%,模型拟合效果较好,Logistic回归模型Hosmer-Lemeshow拟合优度检验显示模型拟合较好(χ^2=9.690,P>0.05);决策树模型AUC为0.881(95%CI:0.861~0.899),Logistic回归模型AUC为0.880(95%CI:0.860~0.899),两模型预测价值均为中等,其存在的差异无统计学意义(Z=0.188,P>0.05)。结论将两模型相结合可以从不同层面发现医院感染的影响因素,能更充分地了解各因素间的相互关系。医院感染风险模型的建立可以为加强院感防控措施提供参考依据,更有效地指导医院感染防控工作。

OBJECTIVE To construct the risk prediction models for nosocomial infection in elderly hospitalized patients of neurosurgery departments based on a decision tree Chi-Squared Automatic Interaction Detector CHAID algorithm and a binary Logistic regression analysis and observe the result of prediction of the models. METHODS The patients who were hospitalized the neurosurgery department of the People′s Hospital of Hainan Province from Jan 2018 to Jun 2019 and were aged no less than 60 years old were retrospectively analyzed, CHAID algorithm and Logistic regression analysis were employed to build the risk prediction models, and the prediction effects were evaluated and compared between the two types of models by means of area under curve(AUC) of receiver-operating-characteristic(ROC). RESULTS Of totally 1 111 patients who were enrolled in the study, 131 had nosocomial infection, with the incidence of infection 11.79%. Both CHAID and logistic regression analysis showed that the length of hospital stay no less than 31 days, use of ventilator and urinary tract intubation were the major influencing factors for the nosocomial infection. The accurate rate of risk prediction of the decision tree model was 88.2%, the model fit well;the test of goodness of fit of logistic regression model Hosmer-Lemeshow showed that the model fitted well too(χ^2=9.690,P>0.05). The AUC of the decision tree model was 0.881(95%CI:0.861~0.899), the AUC of the logistic regression model was 0.880(95%CI:0.860~0.899), both of the models had medium prediction value, and there was no significant difference(Z=0.188,P>0.05). CONCLUSION The combination of the two models may facilitate the discovery of influencing factors for nosocomial infection in different levels and understanding of the relationship among the factors. The risk prediction models for nosocomial infection may provide guidance for prevention and control of the nosocomial infection.