Machine learning based prediction models for outcomes following pulsed field ablation in atrial fibrillation

Abstract Background Catheter ablation is the most effective strategy for maintaining sinus rhythm in atrial fibrillation (AF); however, it is currently offered to only a minority of eligible patients. To optimize procedural benefit and cost-effectiveness, identifying candidates with the highest likelihood of long-term success is crucial. Purpose To develop and validate machine learning (ML)–based prediction models for arrhythmic recurrence following AF ablation using pulsed-field ablation (PFA), a novel and standardized ablation technology, in a large, multicenter cohort. Methods We included all consecutive patients undergoing PFA-based AF ablation across 19 centers. Demographic, clinical, and echocardiographic data routinely available in clinical practice were prospectively collected. Five supervised ML models were trained to predict 1-year arrhythmic recurrence. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC). To enhance clinical interpretability, SHapley Additive exPlanations (SHAP) were applied to quantify feature importance. Results A total cohort of 1688 patients with AF were enrolled (mean age of 62.7 ± 9.7 years, with 27.4% being female; about 31.0% presented with persistent AF, while 79.4% reported symptomatic AF). During a median 1-year follow-up atrial arrhythmic recurrence free rate was 81.4%. Performance of the tested ML models and most influential predictors of recurrence risk at SHAP analysis are highlighted in Figure 1 (panel A and B). The Random Forest (RF) model exhibited the strongest performance. Specifically, on the independent test set, the RF model achieved the highest AUC of 0.75 (95% CI 0.69-0.82), leading to its selection as the final classifier. SHAP analysis revealed Diagnosis-to-Ablation Time (mean SHAP values of 0.0513), BMI (0.0410), and indexed left atrial volume (0.0202) as the most influential predictors of recurrence risk. Conclusion This study presents the first ML based analysis specifically trained and validated on a multicenter cohort of AF patients treated with PFA. Present findings support the use of ML-driven risk stratification to enable personalized patient selection and improve clinical decision-making in AF ablation.