Distinct Pulmonary Vein Antral Substrate-Activation Relationships Are Revealed Through Deep Learning of AF Electrograms

Abstract 2304117, presented at Western Atrial Fibrillation Symposium 2026

Current atrial fibrillation (AF) substrate definition includes both static (voltage) and dynamic (LAT, activation patterns) parameters. However, none of these parameters fully capture the nonlinear, time-evolving dynamics of AF. Although prior studies have characterized the structural and functional properties of the pulmonary vein (PV) antrum, its AF-state dynamical behavior has not been well defined and whether this enables patient tailored lesion sets. We hypothesized that high-density AF electrograms contain distinct nonlinear dynamical signs that predict the ablation targets around the PV antra.AF electrograms from a publicly available OpenEP dataset (4,207 vertices; 59,000 samples) were analyzed for substrate metrics (Shannon entropy, bipolar voltage, local activation time, rotor score, stable activation times) and nonlinear dynamics (lag-1 autocorrelation (AC1), recurrence quantification analysis - recurrence rate (RQA-RR), fluctuation analysis, sample entropy, Hurst exponent). A deep learning classifier was trained using both conventional substrate and dynamical features to differentiate PV-antral lesions (defined by wide area circumferential ablation (WACA)) from the remainder of the left atrium (LA). Model performance, feature importance, and ablation probability fields were evaluated.The model accurately identified PV-antral regions (AUC 0.93). Permutation feature analysis showed that nonlinear dynamical metrics, such as AC1 (temporal organization) and recurrence rate contributed most strongly to classification. Except unipolar Shannon entropy, activation-based features such as voltage, local activation time, and stable activation times had moderate effects. The learned probability map formed a sharply circumscribed ring around the PV antra, closely matching clinical WACA ablation targets. Lastly, a few extra-PV sites also exhibited strong dynamical signals in the probability map.AF electrograms contain unique nonlinear dynamics localized to the PV antrum. These features provided greater discriminatory value than traditional substrate markers and produced ablation probability fields that aligned with clinical WACA lesion sets. These findings suggest that AF nonlinear dynamics may carry additional information beyond conventional substrate mapping and help guide personalized ablation therapies.