Abstract
BACKGROUND: Non-invasive estimation of intracardiac electrograms (EGMs) from body surface potential measurements (BSPMs) could reduce reliance on invasive mapping and enable safer, patient-specific characterization of atrial arrhythmias. Conventional inverse problem formulations, such as Tikhonov regularization, are limited by ill-posedness, sensitivity to anatomical inaccuracies, and low spatial resolution. OBJECTIVE: In this work, we propose a dual-branch deep learning (DL) architecture based on a variational autoencoder (VAE) to directly reconstruct atrial EGMs from BSPMs. METHODS: A dataset of 680 BSPM-EGM pairs was generated using biatrial computational models simulating a wide spectrum of rhythms, including sinus rhythm, atrial fibrillation (AF), ectopic activity, and fibrotic substrates. The network learns a shared latent representation of BSPMs, simultaneously optimized for BSPM self-reconstruction and EGM prediction. Performance was assessed across two phases: a baseline dataset with well-represented rhythms (sinus and multirotor AF), and an extended dataset with rarer arrhythmic classes. Evaluation employed multiple temporal and spectral metrics, as well as spatial voltage and phase mapping. RESULTS: Results show that stratified training yielded the most balanced performance, particularly in AF, with improved correlation, peak detection precision, and spectral coherence compared to baseline and regularized models. Against the zero-order Tikhonov method, the proposed DL model preserved waveform morphology and spectral content more faithfully across rhythm types. CONCLUSIONS: These findings demonstrate that non-invasive, data-driven EGM reconstruction is feasible and can capture physiologically relevant temporal and spatial dynamics. By providing more coherent functional information from BSPMs, DL-based approaches may support individualized diagnosis and guide ablation strategies in atrial arrhythmia care.