Study on the risk prediction model of pneumonia after resection of colorectal cancer
Objectives To establish a risk prediction model for postoperative pneumonia(POP)in patients undergoing resection of colorectal cancer.Methods A total of 317 patients who underwent resection of colorectal cancer under general anesthesia in Beijing Shijitan Hospital,Capital Medical University from September 1st,2018 to September 1st,2021 were selected,and the general data of the patients was collected.The basic characteristic risk variables of POP were screened by Boruta method,and the risk prediction model of POP were established by repeated cross validation,hyperparameter optimization and synthetic minority over-sampling technique(smote)which included four models such as logistic regression(LR),k-nearest neighbor(KNN),the classification and regression tree(CART),and random forest(RF),the confusion matrix parameters of the four prediction models were calculated,and the AUC of ROC curve,the precision recall curve(PRC)and the decision curve analysis(DCA)were used to evaluate the distinguishing ability,calibration ability and net benefit of the four prediction models,respectively.Results Among the 317 patients,there were 112 males and 205 females,aged from 31 to 91 years,with an average age of(64.8±10.8)years,and there were 28 cases(8.83%)of POP.The basic characteristic variables included in Boruta screening were preoperative Hb,preoperative ALB,BMI,preoperative venous thromboembolism(VTE)score,prognostic nutritional index(PNI),duration of operation,duration of anesthesia,and amount of intraoperative crystalloid dosage.The performance of RF prediction model was the best,and the AUC of ROC curve was 0.995(95%CI:0.991-0.999,P<0.05),and the maximum approximate exponent was 0.909,the corresponding risk cut-off was 0.910.The AUC of PRC was 0.996,and the predicted probability of POP was consistent with the actual observation probability.DCA suggested that when the risk threshold was 10%,the net benefit of intervention based on prediction model was higher than that of all intervention or no intervention,and the net benefit of intervention based on the prediction model was 0.975.Conclusions The risk prediction model for POP in patients undergoing resection of colorectal cancer based on maching learning is effectvie and has application value in screening high risk population of POP.
resection of colorectal cancerpostoperative pneumonia(POP)random forest(RF)confusion matrix(CM)partial dependence graph
万方数据
维普
