A GAN Based Heart Sound Denoising Model

The current mainstream methods of heart sound denoising are susceptible to human subjectivity and leads to distortion of the heart sound signal after denoising. In this paper, a generative adversarial network (GAN) based model is proposed to denoise the heart sound signal. It learns the deep features of heart sound signals by building a deep neural network. Both clean heart sound signal and the noisy are fed into the generator in parallel during the training phase and the relevant loss function is introduced. This enables the network to better learn the feature information in the clean heart sound signal, which has been shown to increase the average signal-to-noise ratio (SNR) by about 1 dB. We use SNR, root mean square error (RMSE) and percent root mean square difference (PRD) as metrics. Then the proposed model was compared with denoising models based on wavelet transform (WT) and empirical modal decomposition (EMD). Experiments show that the method we proposed has a high overall SNR and low RMSE and PRD.