XGBoost–SHAP-based interpretable predictive website for postoperative in-hospital mortality in patients with acute aortic dissection

Abstract

Developing a convenient and accurate prognostic model for patients with acute aortic dissection (AAD) using routine demographic and clinical data is challenging. This study aimed to develop an interpretable web-based prediction model using extreme gradient boosting with Shapley additive explanations (XGBoost–SHAP) to forecast postoperative in-hospital mortality (PIHM) in patients with AAD using readily available clinical data. In this retrospective study, we analyzed the admission records of 782 patients with AAD from the Fourth Hospital of Hebei Medical University, who were partitioned into a training set ( n = 470) and a test set ( n = 312). In parallel, 318 patients with AAD from the Second Hospital of Hebei Medical University were enrolled as an independent external validation cohort to assess the generalizability of the developed model. Key risk factors associated with PIHM were identified using the least absolute shrinkage and selection operator (LASSO) and logistic regression analysis. The SHAP method was integrated with XGBoost to create an interpretable predictive model, which was evaluated using receiver operating characteristic (ROC) curve, calibration, discrimination, and clinical utility across the three datasets, culminating in the development of a predictive website for PIHM in patients with AAD. LASSO regression and logistic regression highlighted several critical features associated with PIHM, including acute kidney injury, α-hydroxybutyrate dehydrogenase, Stanford A classification, carbon dioxide, blood chloride, monocyte count, hematocrit, upper gastrointestinal bleeding, and heart failure. The XGBoost-based predictive model demonstrated acceptable performance across all three datasets (sensitivities of 97.9%, 96.8%, and 95.3%; specificities of 71.3%, 67.8%, and 70.9%; accuracies of 84.5%, 82.4%, and 83.3%; AUC values of 0.845, 0.737, and 0.788; and F1 scores of 0.911, 0.899, and 0.903). Ultimately, an XGBoost–SHAP-based website was developed to potentially support clinical decision-making, risk stratification, and patient management. This study develops a model to predict the risk of PIHM in patients with AAD and presents an interpretable XGBoost-SHAP-based prediction website with potential utility for clinical decision-making, risk stratification, and patient management. Not applicable.