Abstract 4363426: Artificial Intelligence-Based Optical Mapping Identifies Active Pulmonary Veins and Outcomes from Atrial Fibrillation Ablation

Background: A major hurdle in improving atrial fibrillation (AF) ablation is reliably identifying patient-specific targets. Mapping based on electrograms (EGMs) has had mixed results, in part because EGM are often challenging to interpret in AF. Optical mapping is an established gold standard for identifying AF waves and potential ablation targets in preclinical models, yet several factors including dye toxicity preclude its use in patients. Objective: To develop an artificial-intelligence (AI)-based system that emulates optical mapping of action potentials from clinical multipolar catheters (fig. A, top) in a large registry. In a hold-out test cohort, we tested if AI-based wave tracking could identify patients likely to respond to pulmonary vein isolation (PVI) by detecting waves exiting active PVs. Methods: coMAP is an AI-system trained using rigorous ground truth labels in >20 million EGMs in a registry of N=236 AF patients (69.0±8.3 years, 72.6% non-paroxysmal AF). It predicts AF activation timings based on unipolar and bipolar EGM features using a 6-layer recurrent neural network, trained on indices validated against human optical mapping of AF, paired with a random forest classifier (Fig. A, bottom). In Fig. B, AI-derived activation times across 3X3 or 4X4 multielectrode arrays were used to estimate predominant AF wave direction near the PVs over a 1-minute interval, and correlated to 1-year outcomes. Results: In the hold-out set in Fig. B, coMAP’s AI-based reconstruction of AF activation outperformed conventional dV/dt (82.4 vs 48.2%, p<0.01). Although AF wave patterns varied over time, consistent directions were identified in 59.7% of segments. Figs B and C show waves exiting the PVs across 6 AF cycles in a 60 Y man with persistent AF, whose ablation was successful at 1Y. Other patients showed variable or indeterminate wave trajectories. Overall, (fig. C) patients with consistent PV-exiting waves (i.e. active PVs) were more likely to have successful AF ablation at 1 year (p<0.05). AF wave direction did not reflect substrates CHADS2-VASc < 2 vs >2; p=NS; nor LVEF < 35% vs >35%; p=NS. Conclusions: An AI-based system to reconstruct optical mapping of AF from clinical multipolar catheters, trained on millions of EGMs, identified patients with PV-dependent AF and non-PV dependent AF in a large registry. Prospective studies are needed to validate AI-based AF wave tracking as a tool to guide ablation both at and beyond the PVs.