Masked Autoencoder for ECG Representation Learning

Abstract

In recent years, self-supervised methods have been widely used in representation learning for electrocardiogram (ECG), but most of the existing methods are based on contrastive learning. Contrastive learning methods usually rely on a large number of negative sample pairs and data augmentation. In this paper, we propose a masked autoencoder-based ECG representation learning model. Our approach is to mask the original ECG signal with a high ratio and then use the autoencoder to reconstruct the original ECG signal. To obtain better ECG features, our model not only extracts local features of ECG using multi-scale convolution, but also global features of ECG using transformer. Our model first pre-trains on the ECG datasets and then fine-tunes on each ECG classification task. Experimental results show that our model outperforms the extant SOTA models for self-supervised learning.