Construction of a postoperative mortality risk model for patients with acute aortic dissection based on XGBoost-SHAP method
10.11855/j.issn.0577-7402.1728.2025.0805
- VernacularTitle:基于XGBoost-SHAP方法的急性主动脉夹层患者术后死亡风险模型的构建
- Author:
Xin ZHANG
1
;
Min FANG
;
Yi CAO
;
Ting-Ting LI
;
Xian-Kong LIU
;
Jia-Yi DANG
;
Xue-Sen ZHAO
;
Hong-Qin REN
;
Jia-Ze GENG
;
Kai-Wen WANG
;
Tie-Sheng HAN
;
Yong-Bo ZHAO
;
Dong MA
Author Information
1. 华北理工大学公共卫生学院,河北 唐山 063000
- Keywords:
machine learning;
prediction model;
acute aortic dissection;
postoperative mortality
- From:
Medical Journal of Chinese People's Liberation Army
2025;50(10):1226-1234
- CountryChina
- Language:Chinese
-
Abstract:
Objective To develop a predictive model for postoperative mortality risk in patients with acute aortic dissection(AAD)using the Extreme Gradient Boosting(XGBoost)algorithm combined with Shapley Additive Explanation(SHAP),and to establish a prediction website to serve as a diagnostic and therapeutic support platform for clinicians and patients.Methods A retrospective cohort study design was adopted.Data from 782 AAD patients who underwent surgical treatment at the Fourth Hospital of Hebei Medical University from January 2013 to December 2023 were collected,including basic information and initial serum biomarker test results.Patients were randomly divided into training and test sets at a 7:3 ratio.An external validation set consisting of 313 AAD patients admitted to the Second Hospital of Hebei Medical University from January 2020 to December 2023 was also established for further model validation.Variables were screened using LASSO regression,and an XGBoost machine learning model was constructed and interpreted using SHAP.The predictive performance of the model was evaluated using receiver operating characteristic(ROC)curve analysis.Using the Shiny package,the XGBoost model was deployed to shinyapps.io to create a prediction website for postoperative mortality risk in AAD patients.One patient was selected by simple random sampling from the test set and the external validation set respectively for the prediction example on the Shiny webpage.Results The XGBoost model demonstrated high predictive performance for postoperative mortality in AAD patients,with area under the ROC curve(AUC)values of 0.928(95%CI 0.901-0.956)in the training set,0.919(95%CI 0.891-0.949)in the test set,and 0.941(95%CI 0.915-0.967)in the external validation set.SHAP values indicated the following order of variable importance in the model(from highest to lowest):"lactate dehydrogenase""blood chlorine""multiple organ injury""carbon dioxide combining power""prothrombin time""α-hydroxybutyric acid""creatine kinase isoenzyme""Stanford classification""combined use of bedside blood purification""gender""acute kidney injury""gastrointestinal bleeding""brain injury"and"shock".A risk prediction website for adverse postoperative outcomes in AAD patients was developed using XGBoost-SHAP method(https://dun-dunxiaolu.shinyapps.io/document/)and validated with examples.One randomly selected patient from each of the test and external validation sets was applied:the predicted mortality risk value for patient 1(who died postoperatively)was 0.9539,and that for patient 2(who survived postoperatively)was 0.0206.Conclusions The XGBoost-SHAP model demonstrates high accuracy in predicting postoperative mortality risk for AAD patients.The online prediction tool established based on this model enhances the identification efficiency of high-risk postoperative mortality patients.
