Ensemble Learning on Body surface Potential Mapping P waves: A Novel Approach to Predict Atrial fibrillation Recurrence Post-direct-current cardioversion

Abstract

Introduction Atrial fibrillation (AF) often recurs within 12 months after direct-current cardioversion (DCCV), yet current prediction approaches predominantly rely on single-lead or segment-level P-wave analysis their generalizability and spatial resolution. This work proposes a comprehensive method leveraging multi-lead P-wave features and ensemble learning to improve individualized recurrence prediction. Methodology 52 patients were recruited in this study (29 had AF recurrence and the rest had normal status). 131-lead body surface ECG recordings from the first 10 seconds post-DC were collected from each patient). P-waves were automatically detected from each lead, and global features were extracted across three domains: (1) Morphology and amplitude statistics from the time domain (2) Power spectrum from frequency domain (3) Time-frequency domain for wavelet-based components. After standardization and imputation, we selected to identify the top 50 predictors based on the different models. Nine shallow machine learning models were trained and compared, including, RF, CatBoost, and a Stacked Ensemble. Performance was evaluated via five-fold cross-validation using accuracy, precision, recall, F1-score, AUC, and other standard metrics. Results The Stacked Ensemble model achieved the highest accuracy (74%) and AUC (0.795), outperforming all base classifiers. XGBoost and LightGBM achieved moderate performance but failed to match the ensemble approach, as reflected in the ROC curves and precision-recall plots. ROC and PR curves confirmed its superior robustness. Notably, the most informative features included amplitude variability and frequency power ratios from leads 28, 40, 83, and 124, corresponding to the right central chest, lower left chest, upper left back, and lower left back—highlighting the predictive value of spatial P-wave dynamics respectively. Conclusion In conclusion, this global-feature-based, ensemble learning framework provides an efficient, interpretable, and scalable solution for post-DC recurrence prediction in AF patients. It offers significant implications for personalized treatment planning and early intervention strategies in clinical cardiology settings.