A multimodal explainable ai model integrating electrocardiographic signatures and left atrial strain to predict atrial fibrillation recurrence after catheter ablation

Background Atrial fibrillation (AF) recurrence after catheter ablation remains common and inadequately predicted by conventional clinical scores. Electrical remodeling captured by ECG and structural remodeling reflected by left atrial strain may provide complementary risk information. We aimed to develop an explainable multimodal model integrating ECG and atrial mechanics to improve recurrence prediction. Purpose To develop and internally validate an explainable multimodal model integrating 12-lead ECG features and left atrial strain to improve prediction of 1-year atrial fibrillation recurrence after catheter ablation. Methods We retrospectively analyzed 1,240 patients undergoing AF ablation. Baseline 12-lead ECG, left atrial volume index, peak left atrial reservoir strain, and clinical variables were integrated using a gradient boosting model with internal cross-validation. Model performance was compared with AI-ECG alone, echocardiographic parameters, and CHA2DS2-VASc using AUROC, NRI, decision curve analysis, and SHAP interpretation. Results A total of 1,240 patients were analysed, including 806 men (65%) and 434 women (35%); 347 (28%) experienced atrial fibrillation recurrence at 1 year. A gradient boosting multimodal model integrating baseline 12-lead ECG features, peak left atrial (LA) reservoir strain, LA volume index, and clinical variables was internally validated. The multimodal model achieved an AUC of 0.84 (95% CI 0.81–0.87), sensitivity 80%, specificity 78%, precision 67%, recall 80%, and F1-score 0.73. In comparison, AI-ECG alone yielded an AUC of 0.76 (95% CI 0.72–0.79), LA strain plus diameter 0.72 (95% CI 0.69–0.75), and CHA2DS2-VASc 0.61 (95% CI 0.58–0.64). Net reclassification improvement versus CHA2DS2-VASc was 31.2%, with an integrated discrimination improvement of 0.24. SHAP analysis identified peak LA strain as the strongest inverse predictor, followed by P-wave morphology features and LA volume index, demonstrating complementary structural and electrical remodeling contributions to recurrence risk. Conclusion An explainable multimodal model integrating ECG-derived electrical markers and left atrial strain significantly improves prediction of 1-year atrial fibrillation recurrence after ablation compared with conventional risk scores. By capturing complementary electrical and structural remodeling, the model enables more precise risk stratification and supports personalized procedural and post-ablation management strategies.