Atrial Fibrillation Driver Localization From Body Surface Potentials Using Deep Learning

Abstract

Atrial fibrillation (AF) is characterized by complex and irregular propagation patterns. Multipoint intracardiac mapping systems present a limited spatial resolution, which makes it difficult to identify AF drivers and ablation targets. These AF onset locations and drivers responsible for AF perpetuation are main targets for ablation procedures. Although noninvasive electrocardiographic imaging (ECGI) and inverse problem-based methods have been tested during AF conditions, they need an accurate mathematical modeling of atria and torso to get good results. In this work, we propose to model the location of AF drivers from body surface potentials (BPS) as a supervised classification problem. We used deep learning techniques to address the problem. We were able to correctly locate the 92% and 96% of drivers in the test and training sets, respectively (accuracy of 0.92 and 0.96), while the Cohen's Kappa was 0.89 for both sets. Therefore, proposed method can help to identify target regions for ablation using a noninvasive procedure as BSP mapping.