Removal of Noise from ECG Signals using Residual Generative Adversarial Network

Removal of Noise from ECG has a great significance in diagnosis of cardiac diseases. Denoising is the foremost step in ECG signal pre-processing tasks. The existing denoising methods in the literature does not provide any linear relationship between signals and does not adaptively work for various types of noises. In this study, a Residual Generative Adversarial Network (R-GAN) structure is proposed for ECG noise filtering. R-GAN has a generator and discriminator unit. The generator network is designed using encoder, residual block, decoder and discriminator is designed using Convolutional layers. Signals are collected from MIT-BIH database for quantitative and qualitative analysis. Various denoising methods in the literature have been explored to make a fair comparison. Experimental results shows that our proposed methodology can effectively retain significant information carried by the given ECG signal compared to existing state of art methods.

I. Introduction

ECG signal when recorded by Holter is corrupted by various types of noised. ECG Morphology plays a vital role in diagnosis of cardiovascular risk factors[1],[2]. Early prediction of cardiovascular risk factors is important [3] By observing the shape, amplitudes and intervals the type of abnormality in ECG signal can be predicted. Based on these features various prediction models are proposed in the literature [4]-[5]. The first and foremost step in developing a predictive model is removal of noise. Hence, Denoising the ECG signal becomes important. Most common noises that affect ECG signal are baseline wander, powerline interference, artifacts, and motion of electrodes. Channel noises like additive white gaussian noises affect the entire frequency band of ECG signal [6].