Fusion of imaging and clinical features: an explainable machine learning model for predicting clinically relevant postoperative pancreatic fistula after distal pancreatectomy

Clinically relevant postoperative pancreatic fistula (CR-POPF) is a severe complication after distal pancreatectomy (DP). Current predictive models lack sufficient accuracy and generalizability. This multicenter retrospective study enrolled DP patients from three centers (2022–2024). Six core predictors were identified using support vector machine-recursive feature elimination (SVM-RFE), eXtreme Gradient Boosting (XGBoost), and Boruta algorithms. Ten machine learning models were constructed and validated internally and externally. Performance was assessed using the area under the receiver operating characteristic curve (AUC), calibration curves, decision curve analysis (DCA), and SHapley Additive exPlanations (SHAP) analysis. Six key predictors were selected: pancreatic computed tomography (CT) attenuation, transection site thickness, operation time, pancreas-to-psoas major muscle CT attenuation ratio, stump management, and main duct ligation. The Random Forest (RF) model achieved the highest AUC among the ten candidate models and yielded an AUC of 0.956 in internal validation and 0.828 in external validation. Calibration and clinical utility were excellent. SHAP analysis indicated that a lower pancreas-to-psoas major muscle CT attenuation ratio, greater parenchymal thickness, longer operation time, non-use of staplers, lower pancreatic attenuation, and non-ligation of the duct were primary risk factors. The RF model, utilizing accessible preoperative and early intraoperative data, accurately predicts CR-POPF with robust interpretability. It serves as a valuable tool for risk stratification, surgical planning, and personalized postoperative management.