Atrial Fibrillation Detection Algorithm Based on Graph Convolution Network

Abstract

Atrial fibrillation (AF) is a common type of arrhythmia with a high incidence and risk, and it is difficult to monitor. Deep learning-based algorithms for AF detection have made preliminary progress, with Recurrent Convolution Neural Networks (RCNN) being widely used for AF detection. The RCNN algorithm combines electrocardiogram (ECG) morphology and rhythm information, where Convolution Neural Network (CNN) mainly processes morphological features and Recurrent Neural Network (RNN) mainly deals with rhythm information. However, a problem with this method is that direct splicing and fusion of ECG rhythm features and ECG morphological features cannot fully exploit the information of both features for AF detection because they have different representative spaces. Therefore, this paper proposes a novel AF detection algorithm based on graph convolution network, named AF-GCN. First, we transform heartbeats into a graph structure, called heartbeat diagrams. The node features of the heartbeat graph are the morphological information of each heartbeat, and the edge of the heartbeat is linked according to the time interval between heartbeats. Moreover, we calculate the edge weight according to the morphological similarity between heartbeats. Therefore, the morphological information and rhythm information of the ECG signal can be unifiedly expressed in one heartbeat. Then, using the Graph Convolution Network (GCN), the morphological and rhythmic features of the ECG signal can be effectively extracted, modeled, and fused in one network modal. We evaluated our AF-GCN method on two popular and widely recognized AF detection benchmarks, MIT-BIH AFDB and Physionet Challenge 2017 AFDB datasets. The results show that AF-GCN outperforms other existing methods.

Keywords

• Atrial fibrillation
• ECG-based
• graph convolution neural network