Abstract 16400: Development of Deep Neural Network Models for Identification of Pulmonary Vein Potentials From Body Surface Electrocardiograms of Patients With Atrial Fibrillation

Abstract

[Background] Pulmonary vein (PV) potential is essential as the origin of atrial fibrillation (AF). However, no established method exists to evaluate PV potentials on a body surface ECG (BS-ECG). [Objective] To develop a method to identify PV potentials from BS-ECGs using state-of-the-art deep neural network (DNN) technology. [Methods and results] A cohort of 799 consecutive paroxysmal AF patients, who underwent Pulmonary Vein Isolation (PVI), were enrolled, and pre- and post-PVI 12-lead ECGs (12ECGs) were collected, amounting to a dataset of 1598 ECGs. Four DNN models were constructed, each employing distinct architectures: 1D-ResNet, 1D-DenseNet, 1D-EfficientNet-b3, and 1D-ViT (Vision Transformer). These models were rigorously trained and evaluated for the identification of PV potentials. The 1D-EfficientNet-b3-based model exhibited optimal performance, achieving an accuracy of 89.4% in the training set and 89.1% in the test set. The Gradient-weighted Class Activation Mapping (Grad-CAM) effectively demonstrated that the segments 200-300ms before the P-wave is crucial for identifying PV potentials. Furthermore, DNN models were developed for each single-lead ECG. While there were variations in the accuracy of PV potential identification among single-lead models, the V6-lead model demonstrated superior performance with an accuracy of 83.2%. [Conclusions] The findings of this research strongly suggest that our DNN-based model can accurately differentiate PV potentials from BS-ECGs. Additionally, Grad-CAM visualizations indicate that pertinent features for the identification of PV potentials are predominantly clustered in the P-wave frontal region. This signifies a promising advancement in the non-invasive detection and analysis of PV potentials, paving the way for improved management of paroxysmal AF.