Integrating traditional statistics and explainable machine learning for risk stratification of atrial fibrillation recurrence after pulsed field ablation: a hypothesis-generating single-center study

Predicting recurrence after pulsed field ablation (PFA) for paroxysmal atrial fibrillation (AF) remains challenging, particularly in early-stage patients with minimal structural remodeling. Traditional risk scores from the radiofrequency ablation era may not be applicable to this new technology and patient profile. This single-center retrospective study included 92 symptomatic paroxysmal AF patients undergoing first-time PFA. The endpoint was AF recurrence after a 3-month blanking period. We integrated Cox regression with an pilot explainable machine learning (XAI) pipeline using a Random Forest model. Internal validation employed stratified 10-fold cross-validation and bootstrap resampling. Model interpretability was assessed using permutation importance and SHAP (SHapley Additive exPlanations) analysis. Performance was evaluated via calibration, discrimination (AUC), and decision curve analysis (DCA). Over a median 28-month follow-up, 13 patients (14.1%) recurred. Multivariable analysis preliminarily identified preoperative log(NT-proBNP) as an independent predictor (HR = 2.15, P = 0.008). While XAI suggested log(NT-proBNP) as the most influential feature, secondary predictors exhibited high instability with importance ranges largely overlapping with zero. The model demonstrated modest calibration (Brier score 0.143) and modest discrimination with high uncertainty (AUC 0.697, 95% CI 0.254–1.000). Decision curve analysis preliminarily suggested a trend toward clinical net benefit between 10 and 15% risk thresholds. An exploratory three-tier risk stratification (NT-proBNP > 300 pg/mL, fluoroscopy time > 22.5 min, LAD > 30 mm) differentiated recurrence risk (3.0%, 11.4%, 33.3% across low-, intermediate-, high-risk groups; Log-rank P < 0.001). This hypothesis-generating pilot study proposes preoperative NT-proBNP as a key candidate predictor for PFA recurrence and demonstrates a transparent XAI framework for small-sample analysis. All findings are preliminary and necessitate validation in larger, prospective cohorts, underscoring the need for PFA-specific prediction tools.