Explainable machine learning for risk prediction of acute cardiac tamponade during atrial fibrillation ablation

Cardiac tamponade is a rare yet catastrophic complication during atrial fibrillation (AF) catheter ablation. Influenced by multiple procedural and patient-related factors, its prediction remains highly challenging. This study aimed to develop and interpret a machine learning-based predictive model for cardiac tamponade during AF catheter ablation. Data were retrospectively collected from 1481 patients who underwent AF catheter ablation at a tertiary hospital in Nanjing, China, between October 2014 and December 2024. After identifying key predictors of intraoperative cardiac tamponade via least absolute shrinkage and selection operator (LASSO) regression, eight machine learning algorithms were trained using Python libraries. Model performance was evaluated through cross-validation, and SHapley Additive exPlanations (SHAP) analysis was performed to interpret the best-performing model. The XGBoost model exhibited the optimal overall performance, with an area under the curve (AUC) of 0.972 in the training set and 0.908 in internal validation, demonstrating excellent calibration and the highest clinical net benefit. SHAP analysis identified five major predictors: operator experience, D-dimer level, total heparin dose, AF type, and left atrial diameter. These predictors represent multidimensional determinants associated with procedural technique, coagulation status, and cardiac anatomy. The XGBoost-based predictive model showed strong discriminative ability and interpretability for predicting cardiac tamponade during AF catheter ablation, which supports accurate preoperative risk stratification and guides intraoperative management to enhance procedural safety and precision. External validation across multiple centers is required to confirm the generalizability of the model.