Deep-learning-based electrode action potential mapping from annotation-free unipolar electrograms for substrate characterization in atrial fibrillation

Catheter ablation for non-paroxysmal atrial fibrillation (non-PAF) remains limited by difficulty in identifying abnormal conduction regions, and current mapping relies on excitation-timing annotation that is unreliable under complex conduction. We developed Deep-learning-based Electrode Action Potential Mapping (DEAP Mapping), which reconstructs membrane potential distributions from unipolar intracardiac electrograms (EGMs) without local activation-time annotation. A convolutional neural network was trained on 13,824 in silico samples from a fibrotic human atrial bidomain model. In ex vivo porcine hearts ( N = 6), DEAP Mapping reconstructed conduction delays and blocks with high similarity to optical mapping (phase-variance-map SSIM 0.87 ± 0.07), outperforming conventional activation maps (0.36 ± 0.14). In a retrospective clinical study ( N = 12 non-PAF patients), the conduction patterns reconstructed by DEAP Mapping partly differed from those of existing methods, and phase-variance analysis revealed spatially continuous conduction heterogeneity with interpatient variability. DEAP Mapping provides annotation-free, optical-comparable substrate visualization from clinical catheter signals, potentially supporting personalized ablation.