AI-based wave tracking reduces mapping burden in atrial fibrillation ablation

Background Even after PVI, efficient identification of additional targets remains challenging and often requires exhaustive mapping that poorly adapts to real-time AF dynamics. A navigation strategy that continuously infers activation and suggests directional moves could shorten procedures and improve outcomes. Hypothesis: A deep learning system that accurately identifies local activation from EGMs can provide practical intraprocedural catheter direction (up/down, left/right) toward sites linked to AF termination or long-term success, reducing mapping burden versus conventional approaches. Methods Using a registry of >20 million EGMs (N=236, ) (fig. 1), we trained an activation-timing model and embedded it into a retrospective guidance pipeline. On multi-electrode arrays (HD grid/basket), we computed predominant wave propagation in rolling time windows to generate directional guidance. In a validation cohort (69.0+/-8.3Y; 72.6 Paroxysmal) including ~5 million EGMs, we iteratively repositioned the array according to guidance and quantified (i) fraction of paths reaching successful ablation regions, (ii) number of moves to arrival, and (iii) concordance with AF termination or 1-year freedom from recurrence. Results Guidance based on AI-based wave tracking pointed to successful regions in the majority of cases and more often than control sites (85% vs 13%, fig. 2A). Applied successively, termination sites were reached within 4+1 movements from any arbitrary position, translating into 5 minutes, which compares favorably with typical 10–15-minute comprehensive mapping (fig. 2B). Performance was maintained across ranges of signal quality by leveraging AI beat-tracking to suppress spurious deflections and stabilize directional estimates (fig. 2C). Conclusions AI-based wave tracking can be operationalized into a practical tool during AF ablation to rapidly direct the catheter toward critical sites, potentially reducing mapping time and improving procedural efficiency. This approach complements PVI by revealing dynamic non-PV targets in a closed-loop workflow.Fig. 1.PopulationFig. 2.AI-based wave identification.