Interpretable stacking ensemble combining CT radiomics and clinical features improves preoperative prediction of acute severe cholecystitis

In a two-center study of 492 patients, a stacking ensemble combining CT radiomics (TabNet) and clinical features (XGBoost) predicted acute severe cholecystitis with external test AUC 0.827, outperforming standalone models (0.753–0.792). SHAP showed key predictors: neutrophil %,…

• In five-fold cross-validation, the stacking model achieved mean AUC 0.850, surpassing TabNet (0.807) and XGBoost (0.799).
• SHAP analysis revealed that radiomic features contributed a higher overall weight to the ensemble than clinical features.
• The model was externally validated on 79 patients from a second center, yielding the lowest Brier score (0.156) and highest clinical net benefit.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.