Artificial Intelligence Beat-Tracking Improves Regional Rate

Abstract Background Pulmonary vein isolation (PVI) is the most effective approach to treat atrial fibrillation (AF), but 30-40% of patients require additional therapy. CAPLA and historical trials such as RADAR posit that ablating rapid sites on the posterior left atrium (or elsewhere) may improve success, yet have mixed results. Importantly, AF rate estimates may be inaccurate since AF electrograms are often complex, with spurious deflections that may be counted to overestimate rate (fig. A, red). Hypothesis We hypothesized that artificial intelligence (AI) systems trained on millions of electrograms may indicate AF rate better than conventional tools, by ‘filtering’ spurious activations. Methods We studied N=229 AF patients at ablation (68.2±8.7 Y, 13.7% females, 71.6% non-paroxysmal AF). We developed AI tools to track individual AF cycles from >10M unipolar electrograms in various left and right atrial regions in a training cohort (N=174), filtered 0.05-500 Hz. We applied this AI AF-Beat tracking model to identify activations in a matched-validation cohort (N=55), that we compared to conventional dV/dt marking and to experts informed by monophasic action potential (MAP) recordings. Results Fig. B shows AF in a 66 year old man, where dV/dt often marked non-physiological activations within repolarization inferred from high quality unipolar electrograms. Conversely, the AI system excluded spurious deflections and better correlated with experts (black). Fig. C shows, overall, that AI provided a higher accuracy than dV/dt (Observed value [F1 score; 95% CI]: 0.858 [0.836 - 0.878] vs. 0.747 [0.723 - 0.770]) and a lower percentage of spurious activations (0.234 [0.212 - 0.250] vs. 0.455 [0.432 - 0.478]]) when compared against expert markings. Fig. D shows that AI beat tracking better identified AF cycle length (rate; blue) than standard approaches (green) that substantially underestimate cycle length (p<0.01). Conclusions A novel AI-based beat tracking approach for AF, trained in a large registry, more accurately detected AF rates in clinical electrograms than standard approaches that overestimate rate. AI-based beat tracking in AF could have clinical utility in guiding ablation and better characterizing regional AF activity.