Development and validation of a LightGBM-based model with risk stratification for predicting early recurrence after catheter ablation for atrial fibrillation.

BACKGROUND: Atrial fibrillation (AF) is the most prevalent sustained arrhythmia, yet tools for predicting early recurrence (ER) after catheter ablation remain limited. This study aimed to develop a machine learning model to estimate ER risk following first-time AF ablation. METHODS: In this retrospective single-center study, 519 patients undergoing initial AF ablation were enrolled (ER rate: 9.2%). Eight predictors were selected via recursive feature elimination. A LightGBM model was constructed and internally validated against logistic regression and conventional risk scores. RESULTS: The LightGBM model achieved an AUC of 0.715 in training and 0.704 in testing, showing higher discrimination than logistic regression (AUC = 0.623) and traditional scores (APPLE AUC = 0.560). SHAP analysis identified mitral regurgitation severity, age, hemoglobin, and albumin as predominant predictors. Using a Youden-derived threshold (0.099), high- and low-risk groups exhibited significantly different recurrence rates in testing(11.4% vs. 2.9%; P < 0.05). CONCLUSION: We developed a LightGBM-based model integrating structural and metabolic features that modestly improves upon conventional approaches in predicting ER after AF ablation. This tool may facilitate personalized post-procedural management. Multicenter prospective validation is warranted.