Statistics and behavior of clinically significant extra-pulmonary vein atrial fibrillation sources: machine-learning-enhanced electrographic flow mapping in persistent atrial fibrillation.

Introduction: Electrographic flow (EGF) mapping is an FDA 510(k)-cleared method for visualizing atrial activation wavefronts in atrial fibrillation (AF). Its clinical efficacy was demonstrated in the FLOW-AF randomized controlled trial, and its fundamental principles have been previously described. However, the underlying machine learning strategy used to develop and refine the EGF algorithm has not yet been detailed. Here, we present how our EGF Model-trained on procedural outcomes from 199 fully anonymized retrospective patient datasets-identifies clinically significant sources of AF and how this machine learning-driven hyperparameter optimization underlies its clinical effectiveness. We also examine the statistical characteristics of the identified sources and their impact on cycle length variability, offering insights into potential pathophysiological mechanisms. Methods and results: Unipolar electrograms were recorded from patients with persistent or long-standing persistent AF using 64-electrode basket catheters. The EGF Model processes these recordings to reconstruct divergent wavefront propagation patterns and quantify their temporal prevalence. We included 399 retrospective patients in total: 199 for training and optimizing 24 model hyperparameters, and 200 for subsequent analyses of source prevalence and characteristics. Our machine learning approach established an activity threshold, above which divergent wavefront patterns-termed "significant sources"-predicted AF recurrence. This threshold was validated in 85 prospective patients from the published FLOW-AF trial. Significant sources persisting post-procedure were associated with significantly higher recurrence rates than those successfully ablated. Notably, the majority of significant sources were not continuously active; however, when these sources switched "ON," the spatial variability of AF cycle lengths in the respective atrium decreased by more than 50%, suggesting an entraining effect. Conclusions: By systematically optimizing the EGF Model's hyperparameters based on clinical outcomes, we reliably detect and target key AF sources that, when ablated, improve procedural success. These findings, supported by the FLOW-AF trial, underscore the usefulness of clinical outcome-based machine learning to improve the efficacy of algorithm based medical diagnostics.