Multi-Period Attention for Automatic ECG Classification

Abstract:
Electrocardiogram (ECG) is of great importance to the diagnoses of heart diseases, and has been increasingly studied in the field of deep learning. Many sophisticated deep learning methods have been transferred to the classification of ECG due to the success in image classification. However, these methods have not perfectly adapted to the periodic characteristic of ECG which is an essential aspect. In this paper, we propose to address this issue with an efficient attention block called Multi-Period Attention, in which periodic masks have been generated based on periodic function. Besides, the masks have been mutually independent and adaptively adjusted according to different channels to avoid overfitting. We have applied the blocks into ResNet18 and conducted experiments on clinical 12-lead ECG dataset PTB-XL. The results have showed that our method has comparative ability over the other models with much more parameters.