Abstract
The electrocardiogram (ECG), as an essential basis for the diagnosis of cardiovascular diseases, is usually disturbed by various noise. To obtain accurate human physiological information from ECG, the denoising and reconstruction of ECG are critical. In this paper, we proposed an ECG denoising method referred to as LSTM-DCGAN which is based on an improved generative adversarial network (GAN). The overall network structure is composed of multiple layers of convolutional networks. Furthermore, the convolutional features can be connected to their time series order dependence by adding LSTM layers after each convolutional layer. To verify the effectiveness and the denoising performance of the improved network structure, we test the proposed algorithm on the famous MIT-BIH Arrhythmia Database with different levels of noise from the MIT-BIH Noise Stress Test Database. Experimental results show that our method can remove the single noise and the mixed noise while retaining the complete ECG information. For the mixed noise removal, the average SNR(imp), RMSE, and PRD are 19.254 dB, 0.028, and 10.350, respectively. Compared with the state-of-the-art methods, DCGAN, and the LSTM-GAN methods, our method obtains the higher SNR(imp) and the lower RMSE and PRD scores. These results suggest that the proposed LSTM-DCGAN approach has a significant advantage for ECG processing and application in complex scenes.