Defining the limits of pulmonary vein isolation: electroanatomic substrate predicts recurrence in first-time paroxysmal atrial fibrillation ablation

Background Pulmonary vein isolation (PVI) remains the treatment of choice for paroxysmal atrial fibrillation (PAF). However, the clinical distinction between paroxysmal and persistent AF may not reflect the underlying atrial substrate. Automated extraction of electroanatomic mapping (EAM) features now enables rapid substrate characterization at baseline. We hypothesized that these EAM-derived metrics obtained via automated machine learning platform within CARTONET, can identify PAF patients in whom PVI alone is insufficient, outperforming traditional comorbidities and AF classification in predicting recurrence. Methods We analyzed 263 consecutive patients with clinically defined PAF undergoing first-time, PVI-only ablation. Baseline anatomic and voltage metrics were extracted automatically from CARTONET ML models within minutes of case upload (Figure). Recurrence of atrial tachyarrhythmia (≥30 seconds, beyond a 60-day blanking period) was assessed at 12 months. Demographic and clinical variables were compared using chi-square testing; EAM-derived geometric and low-voltage zone (LVZ) features (<0.5 mV and <0.2 mV) across posterior, inferior, lateral, anterior, roof, mitral, and non-PV/non-LAA regions were compared using independent-sample t-tests. Results At one year, 61 of 263 patients (23%) experienced recurrence of atrial arrhythmias. Demographic variables (age, sex, BMI, atrial volume) and clinical comorbidities except hypertension (p = 0.002) were not predictive in this PAF cohort. In contrast, automatically extracted regional LVZ metrics from CARTONET strongly correlated with recurrence, particularly within the posterior (p = 0.012), inferior (p = 0.003), right lateral (p = 0.002), roof regions (p = 0.008) as well as globally in the left atrium excluding the left atrial appendage and the pulmonary veins (p = 0.008). Larger LVZ burden corresponded to modest effect sizes (Cohen’s d 0.35–0.45, Table). Conclusions In patients with clinically defined PAF undergoing first-time PVI, rapidly derived ML-based EAM features identify substrate-driven recurrence risk far better than clinical classification or comorbidities. Automated extraction of atrial voltage and geometry enables near-real-time substrate assessment, offering a scalable path toward personalized, substrate-guided ablation strategies beyond the traditional "paroxysmal vs persistent" framework.