An end-to-end atrial fibrillation detection by a novel residual-based temporal attention convolutional neural network with exponential nonlinearity loss

Abstract

Atrial fibrillation (AF) is one of the most common abnormal heart rhythms, which is caused by the fast contraction of the two upper atria. Despite of the fact that convolutional neural network (CNN) has been applied to electrocardiogram analysis for AF rhythm, it cannot achieve the expected performance due to the lack of consideration for temporal features and the imbalance problem. In order to make the network concentrate on the learning of AF temporal features, we propose a residual-based temporal attention block (RTA-block). The RTA-block utilizes residual learning to generate temporal attention weights, which enhance informative features related to AF. Powered by the RTA-block, a residual-based temporal attention convolutional neural network (RTA-CNN) is further proposed for AF detection. The network can automatically focus on the parts with more sematic information to achieve better performance. In addition, we propose a novel exponential nonlinearity loss (EN-Loss), which addresses the imbalance problem by changing the nonlinearity of the loss function. We evaluated our framework on the single lead ECG classification dataset of The PhysioNet Computing in Cardiology Challenge 2017. The experimental results show that the proposed RTA-CNN with EN-Loss can obtain competitive results over the state-of-the-arts classification networks, which proves the method’s effectiveness.