Designing artificial-intelligence based tracking of atrial fibrillation waves to identify ablation response

Ablation for atrial fibrillation (AF) is successful in many patients, yet it is difficult to identify responders. Physiologically, ‘electrical fingerprints’ should be important, yet clinical features (drivers, fractionation, others) are debated. We set out to identify responders using artificial intelligence (AI) by tracking AF wave direction, which may show repeating trends over time (fig A), rather than having to select a "site of interest."

We hypothesized that AF waves that predominantly exit pulmonary veins (PV), using an AI based wave tracking approach applied to multipolar catheters, may reflect active PVs and thus identify ablation response.

We developed an AI-system to track AF waves from >10 million electrograms in N=229 AF patients (68.2±8.7 years old, 71.6% non-paroxysmal AF) at ablation, recorded with multipolar catheters throughout the entire atria. Fig. A shows the design of the AI-system, that we applied to a validation cohort of N=37 patients to calculate AF vectors in spatial grids (covering ~3 cm2) for 1 minute. We compared wave maps near PVs for AF termination and 1 year outcome from ablation.

Ablation success was 71.4% at 1 year. AF waves shifted direction over 1 minute, but often with repeating vectors (Fig. A) and AI models converged . Fig. B shows AF waves predominantly exiting PVs in a 60 YO man with successful PV isolation at 1 year. Overall, Fig C. shows that AF waves were more likely to exit PVs (<0.05), in patients with successful vs failed ablation . Fig. D shows that wave vectors at PVs did not predict AF termination (p=NS).

A novel AI-based system trained on millions of AF electrograms showed that predicting long-term success of PVI likely includes mapping AF waves that predominantly exit the PVs. Future studies should correlate wave direction at other locations with clinical endpoints.Figure 1