On the Atrial Fibrillation Detection Performance of ECG-DualNet

Abstract

AbstractAtrial Fibrillation (AF) is a severe cardiac arrhythmia affecting a significant amount of the human population [1]. Quick diagnosis and treatment are critical in reducing the risk of severe sequelae such as stroke or heart failure. Rohret al. [2] recently proposed ECG-DualNet, a neural network for accurate AF detection in single-lead electrocardiogram (ECG) data. This short paper reports additional empirical results of ECG-DualNet to gain new insights on AF detection in single-lead ECG data with deep neural networks. We systematically analyze which ingredients of ECG-DualNet are crucial for achieving competitive AF detection results. We also scale the ECG-DualNet architecture to 130M parameters and perform large-scale supervised pre-training, providing additional empirical results. Finally, we provide recommendations for future research toward accurate and robust AF detection.