Combining AI interpretation of cholangiograms and clinical data for postoperative risk stratification after ERCP in cholodocholithiasis

Endoscopic retrograde cholangiopancreatography (ERCP) is the standard treatment for choledocholithiasis, yet postoperative risk stratification remains challenging. We aimed to develop an interpretable artificial intelligence (AI) framework that integrates cholangiographic imaging and clinical data to support postoperative risk stratification after ERCP-guided stone extraction. In this retrospective study, 960 patients who underwent ERCP for confirmed common bile duct (CBD) stones between October 2020 and December 2023 were analyzed. ERCP fluoroscopic images were segmented using a two-stage KM-UNet to identify bile ducts and stones. Quantitative geometric descriptors, including duct width, curvature, and stone size, were extracted from the segmentation results. These imaging-derived features were integrated with demographic, clinical, and laboratory variables and analyzed using a TabPFN (Tabular Prior-data Fitted Network) model. Model interpretability was evaluated using SHAP (Shapley Additive Explanations) and compared with L1-regularized logistic regression. During follow-up through December 2024, 124 patients experienced recurrent CBD stones (13.1%) and 836 remained recurrence-free (86.9%), while 129 had documented all-cause death status (13.4%). The multimodal AI model demonstrated strong performance for recurrence and all-cause mortality risk stratification. Imaging-derived morphological features—particularly bile duct tortuosity, maximal turning angle, and stone size—were among the strongest contributors to all-cause mortality prediction. Feature attribution analysis showed close agreement between SHAP-derived importance and regression-based feature selection, supporting the transparency and reliability of the model. Integrating quantitative cholangiographic morphology with clinical data may improve postoperative risk stratification in patients with choledocholithiasis after ERCP. This interpretable AI-based framework may serve as a decision-support tool for individualized recurrence and all-cause mortality risk assessment and follow-up planning.