Abstract 4136983: Estimating Atrial Fibrillation Rate in Patients Using Artificial Intelligence Referenced to Tissue Action Potentials

Background: Ablation by pulmonary vein isolation (PVI) is the most effective approach to treat atrial fibrillation (AF), but 30-40% of patients may not respond. Studies including the CAPLA trial suggest that ablating sites of rapid activation on the posterior left atrial wall or elsewhere may improve success. However, AF rate is rarely calibrated and, since electrograms often include noise or far field activity, may overestimate rate (underestimate cycle length) at many sites. Hypothesis: We hypothesized that artificial intelligence (AI) algorithms may more accurately indicate local AF rate, by excluding spurious activations referenced to simultaneous monophasic action potential (MAP) recordings, compared to conventional measurements. Methods: We studied N=303 AF patients at ablation (68.2±8.0 years, 17.8% females, 72.9% non-paroxysmal AF). Patients were matched N=229 into development and one quarter (N=74) into hold-out test cohorts. Patients underwent AF mapping using multipolar catheters, and a subset with MAP catheters (MedFact, GmbH). We used the development set to train an AI algorithm to reliably detect AF activations, which we compared in the test cohort to a standard dV/dt approach referenced to adjacent MAPs within 2 mm (N=20 patients) and manually annotated activations (N=64 patients). Results: Fig A shows AF in a 66 year old woman, where dV/dt often marked non-physiological activations within repolarization shown in the adjacent MAP. Conversely, the AI system excluded spurious deflections and better correlated with experts (black). In summary, Fig. B shows that AI provided a higher F1 score than dV/dt compared to MAPs (median[IQR]: 0.86 [0.5 - 1] vs. 0.67 [0.40 - 0.73], p<0.01) and experts (0.89 [0.73 - 1] vs. 0.72 [0.60 - 0.89]], p<0.01). Fig C shows that the AI approach better identified AF cycle length (rate; blue) than standard approaches (green) that substantially underestimate cycle length (p<0.05). Conclusions: In this large registry, a novel AI-based approach more accurately detected AF rates in clinical electrograms, referenced to experts and action potential recordings, than standard approaches which overestimated rate. AI may improve characterization of AF activity within the atria.