A deep learning framework for ECG denoising and classification

Cardiovascular disease (CVD) is a major cause of mortality worldwide. To facilitate early prevention and timely diagnosis of CVD, daily electrocardiogram (ECG) monitoring has become an important tool. However, various noises introduced into ECG signals will affect the diagnostic utility of these recordings. In addition, it is hard for non-specialists to interpret the ECG signals. To address this problem, we propose CS-TRANS, a novel deep learning framework for ECG denoising and classification. This framework mainly incorporates: (i) convolutional neural networks (CNN) regulated by stationary wavelet transform (SWT), called CNN-SWT; and (ii) Transformer encoder. The convolution kernel constraints and architecture of SWT are introduced into CNN to learn both the linear and non-linear time-frequency features more efficiently, while the transformer encoder further boosts global feature extraction. We implemented CS-TRANS and compared it with the state of the arts using the MIT-BIH arrhythmia database. The evaluation results show that compared to other algorithms, CS-TRANS improves the output signal-to-noise ratio (SNR) by 10% in denoising, and achieves 0.5% higher accuracy in classification, while saving half to even more than 90% of the parameters of the model.