Non-Invasive Electrogram Estimation from Body Surface Potential Mapping Using Artificial Intelligence

In this study, we propose a novel approach utilizing 3D Convolutional Neural Networks (CNNs) to estimate atrial electrical activity from Body Surface Potential Mapping (BSPM).The data used consists of a database comprising 16,800 simulations that encompass both sinus and ectopic beats, generated using a cellular automata model and diverse atrial geometries derived from a statistical shape model.Our method involves converting BSPM into 2D for each time instant as input for a CNN and outputting the corresponding 2D images of the electrical estimations in the heart.To evaluate the performance of our approach, we compared it with the traditional Tikhonov 0 order regularization method.Our results demonstrate that the proposed CNN-based method outperforms the traditional approach, surpassing it by 0.870.05vs. 0.730.05for average EGM correlation and 0.040.01 vs. 0.070.01for root mean squared error, achieving more accurate estimations of atrial electrical activity.These results highlight the effectiveness of deep learning methodologies, particularly 3D CNNs, in accurately estimating electrical activity from BSPM, without relying on detailed cardiac geometry or location information, and aiming to improve the current state-of-the-art electrocardiographic imaging solved by Tikhonov regularization.