Integrating Conventional Biostatistical Modeling and Explainable Machine Learning Algorithms for Advanced Risk Stratification of Atrial Fibrillation Recurrence Following Pulsed Field Ablation: A Hypothesis-Generating Single-Center Investigation Exploring Clinical, Electrophysiological, and Structural Cardiac Determinants of Post-Ablation Arrhythmia Relapse

Abstract

Background: Pulsed field ablation (PFA) represents a paradigm shift in catheter-based therapy for atrial fibrillation (AF), offering non-thermal, selective myocardial tissue destruction through irreversible electroporation. Despite high rates of acute procedural success, post-ablation AF recurrence remains a clinically significant challenge, with rates ranging from 20% to 45% at 12 months in contemporary series. Predictive models for recurrence risk have been limited by small cohort sizes, methodological heterogeneity, and reliance on traditional biostatistical approaches ill-suited to capture complex, non-linear biological interactions.

Methods: We conducted a single-center, retrospective observational study enrolling 248 consecutive patients undergoing first-time PFA-mediated pulmonary vein isolation (PVI) between January 2022 and September 2024. The primary endpoint was any documented AF recurrence beyond a 90-day blanking period, detected via 12-lead ECG, Holter monitoring, or implantable loop recorder. A comprehensive analytic pipeline was employed, integrating Cox proportional hazards modeling and multivariable logistic regression with a battery of explainable machine learning (XML) algorithms—including Random Forest, Extreme Gradient Boosting (XGBoost), LightGBM, Support Vector Machine, Multilayer Perceptron, and a stacking ensemble. Model interpretability was evaluated using SHapley Additive exPlanations (SHAP), permutation feature importance, and partial dependence plots. Internal validation was conducted via 5-fold cross-validation and bootstrapped calibration.

Results: AF recurrence occurred in 85 patients (34.3%) over a median follow-up of 14.2 months (IQR 9.8–18.6). On multivariable Cox analysis, independent predictors of recurrence included AF duration (HR 1.18 per 6-month increment, 95% CI 1.09–1.28), left atrial volume index (LAVI; HR 1.14 per 5 mL/m², 95% CI 1.06–1.23), left atrial diameter (HR 1.38 per 5 mm, 95% CI 1.21–1.57), and acute pulmonary vein reconnection (HR 2.61, 95% CI 1.29–5.27). The stacking ensemble demonstrated superior discriminative performance (AUC 0.847, 95% CI 0.794–0.900; Brier score 0.147) compared to conventional models (Cox: AUC 0.742). SHAP analysis identified LAVI, AF duration, LA diameter, and acute PV reconnection as the four highest-impact features across all ML architectures, with consistent non-linear interaction effects between structural and temporal predictors.

Conclusions: This hypothesis-generating investigation demonstrates that integrating conventional biostatistics with explainable ML substantially improves AF recurrence risk stratification following PFA. The convergence of results across methodologically orthogonal modeling approaches strengthens biological plausibility and identifies a core quartet of modifiable and stratifiable risk determinants. Prospective validation in multicenter cohorts and integration with wearable cardiac monitoring data are warranted to translate these findings into clinical decision support tools.

Keywords: Atrial fibrillation; Pulsed field ablation; Irreversible electroporation; Recurrence; Machine learning; XGBoost; SHAP; Risk stratification; Left atrial remodeling; Pulmonary vein isolation