Development of a multimodel predictive framework for gangrenous cholecystitis with SHAP-based interpretability

BACKGROUND: Gangrenous Cholecystitis (GC) is a critical pathological type of acute cholecystitis that tends to induce gallbladder wall necrosis or perforation, thereby leading to localized or diffuse peritonitis and is associated with high morbidity and mortality. Due to its nonspecific clinical symptoms and strong heterogeneous manifestations, timely and accurate identification of GC preoperatively remains a clinical challenge. To address this unmet clinical need, we developed an interpretable machine learning (ML) model using routinely available preoperative variables to facilitate the early prediction of GC and support clinical decision-making. METHODS: This was a retrospective cohort study, including 487 patients with acute cholecystitis who underwent laparoscopic cholecystectomy in our hospital between February 2020 and July 2023. After initial univariate analysis and logistic regression, nine ML models were constructed, including logistic regression (LR), random forest (RF), XGBoost, LightGBM, support vector machine (SVM), artificial neural network (ANN), k-nearest neighbor (KNN), linear discriminant analysis (LDA), and a Stacking ensemble model. LASSO regression was used to select key preoperative predictors. Model performance was evaluated using area under the receiver operating characteristic curve (AUC), accuracy, F1-score, sensitivity, specificity, decision curve analysis (DCA), and calibration curves. SHAP (SHapley Additive exPlanations) analysis was performed to enhance model interpretability. RESULTS: LASSO regression identified five key preoperative predictors: preoperative C-reactive protein (pre-CRP), gallbladder diameter, pericholecystic exudation, albumin (ALB), and basophil percentage (BA). Among all constructed models, XGBoost and RF exhibited superior predictive performance, with AUC values of 0.90 and 0.89, respectively. Both models demonstrated robust discrimination, good calibration, and significant net clinical benefit across the entire range of decision thresholds. SHAP analysis revealed that pre-CRP and gallbladder diameter were the most influential predictors of GC, while BA and ALB showed inverse or sample-dependent associations with GC occurrence. CONCLUSIONS: We developed an interpretable, non-invasive, and high-performance ML framework for preoperative GC prediction using routine clinical variables. This framework provides a valuable clinical tool to assist surgeons in decision-making and early risk stratification of patients with suspected GC, potentially improving patient outcomes by enabling timely intervention.