Preoperative prediction of acute severe cholecystitis using an attentive interpretable tabular network (TabNet)-based radiomics model and a stacking ensemble: a two-center study

Objectives To develop an explainable and cost-effective predictive model for acute severe cholecystitis (ASC) utilizing stacking ensemble learning framework and SHapley Additive exPlanations (SHAP) algorithm.Methods This retrospective study was conducted on 492 patients with pathologically confirmed acute cholecystitis, collected from two tertiary hospitals between January 2020 and January 2023. The analysis was performed in December 2025. The patients divided into a training set (n = 413, Center 1) and an external test set (n = 79, Center 2). We developed a two-level stacking ensemble framework: level 1 employed attentive interpretable tabular network (TabNet) for CT radiomics and extreme gradient boosting (XGBoost) for clinical/imaging features as base learners; level 2 utilized logistic regression as a meta-learner to fuse predictions. The stacking model was compared with standalone TabNet and XGBoost models.Results In five-fold cross-validation, the stacking model achieved a mean area under the curve (AUC) of 0.850, surpassing standalone TabNet (0.807) and XGBoost (0.799). This superiority was maintained in the external test set (AUC: 0.827 vs. 0.753 and 0.792). In the external test set, the stacking model also yielded the lowest Brier score (0.156) and the highest clinical net benefit in decision curve analysis. SHAP analysis identified neutrophil percentage, gallbladder wall necrosis, and pericholecystic exudation as the most influential clinical predictors, with radiomic features providing a higher overall weight in the final ensemble.Conclusions The interpretable stacking model effectively integrates clinical and radiomic data to accurately predict ASC preoperatively.