重症监护病区医院感染风险预测模型构建

Construction of risk prediction model for nosocomial infection in intensive care unit

目的应用分类树(CHAID)算法和Logistic回归分析构建重症监护病区医院感染风险预测模型,并比较二者预测结果的优劣。方法回顾分析2016年1月-2017年6月入住重症监护病区(包括ICU、RICU、CCU/CNICU)超过48h及转出重症监护病区48h内的住院患者,应用CHAID算法和Logistic回归分别建立医院感染的风险预测模型,并对模型进行拟合优度检验评价模型,使用受试者工作特征曲线(ROC)曲线下面积(AUC)比较两种预测模型的优劣。结果共收集患者1 232例,其中院内感染182例,感染发病率14.8%;分类树模型和Logistic回归均显示住院天数≥10天、APACHEⅡ评分≥20、中心静脉插管日数≥7天是院内感染发生最重要的影响因素;分类树Risk统计量为0.286,模型拟合效果较好;分类树模型的灵敏度为83.5%,特异度59.3%,ROC AUC为0.788(95%CI 0.742～0.835);Logistic回归模型的灵敏度为80.2%,特异度81.3%,AUC为0.869(95%CI0.832～0.906);通过比较,分类树模型和Logistic回归模型两者结果差异有统计学意义(Z=4.656,P<0.001)。结论 Logistic回归模型的预测效果优于分类树模型,两个模型的分析结果相结合可以从不同层面发现医院感染的风险因素,为进一步预防与控制医院感染的发生提供参考依据。

OBJECTIVE To construct risk prediction models of nosocomial infection in intensive care unit by using Chi-Square Automatic Interaction Detection (CHAID) algorithm and the Logistic regression analysis, and compare the quality of the prediction results. METHODS Hospitalized patients admitted to Intensive Care Unit (ICU, RICU, CCU/CNICU) for more than 48 h and transferred out of Intensive Care Unit for less than 48 h from Jan. 2016 to Jun. 2017 were included in this retrospective study. Risk prediction models of nosocomial infection were constructed by using CHAID algorithm and the Logistic regression analysis. The risk models were evaluated by conducting goodness-of-fit tests. The area under ROC curve analysis was used to compare the two prediction models. RESULTS Among the 1232 patients collected, 182 cases of nosocomial infections occurred, and the infection rate was 14.8%. Classification tree model and Logistic regression showed that hospitalization time greater than 10 days, APACHE U score greater than or equal to 20 points, and more than 7 days of central venous intubationwere the most important influencing factors for nosocomial infections. The Risk statistics (0.286) revealed that CHAID model fitted the data well. The sensitivity, specificity and area under the ROC curve (AUC) of the CHAID model were 83.5%, 59.3% and 0.788 (95% CI 0.742-0.835), respectively. The sensitivity, specificity and AUC of the Logistic regression model were 80.2%, 81.3% and 0.869 (95% CI 0.832-0.906), respectively. There were significant differences in results of the CHAID model and the Logistic regression model (Z = 4.656, P <0.001). CONCLUSION The Logistic regression model is superior to the CHAID model in predicting nosocomial infections. The combination of the two models enable us to find out risk factors of nosocomial infection at different levels, which can provide a reference for further prevention and control measures of nosocomial infections.