Comparing artificial-intelligence based tracking of atrial fibrillation waves with clinical phenotypes in patients undergoing ablation

Abstract Background Atrial fibrillation (AF) is a major cause of morbidity and mortality, which is reduced by ablation. However, predicting response is difficult and mostly done using static chart data, since it is difficult to interpret AF electrogram features or link them with clinical presentation. There is a need for a metric that can accurately predict responders to AF ablation. Hypothesis We hypothesized that mapping AF waves for repeating patterns (fig. A) near the pulmonary veins (PV), using novel artificial-intelligence (AI) tools developed from multipolar data in AF, may reflect response to ablation and clinical phenotypes. Methods We developed an AI-system to identify potentially repeating AF waves on multipolar catheters (fig A), trained on >10 million electrograms in N=229 richly characterized AF patients (68.2±8.7 years old, 71.6 % persistent AF). In Fig B., AI was used to track AF waves near the PVs for 1 minute in a validation cohort population (26,640 electrograms, N=37). We compared peri-PV AF vectors to ablation outcome, classification (persistent vs non-paroxysmal AF), stroke risk factor (CHADS2-VASc), and left ventricular ejection fraction (LVEF). Results In Fig. B, wave vectors varied over 1 min yet predominantly exited PVs in a 60 Y man with PeAF and successful ablation, and did not predominantly exit PVs in a 70 Y woman with failed ablation at 1 year. In Fig. C, PV-exiting AF waves were more likely in successful vs failed ablation (<0.05), but did not reflect AF type (p=NS), CHADS2-VASc <2 vs >2 (p=NS) nor LVEF <50% vs >50% (p=NS). Conclusions A novel AI-based system to track AF waves trained using millions of electrograms identified ablation responders vs non-responders by AF waves exiting pulmonary veins. Future studies should examine if AF wave tracking, a novel metric of organization, is independent of other clinical AF classifications.Figure 1