Denoising of ECG with single and multiple hidden layer autoencoders

Abstract

Denoising autoencoder (DAE) with single or multiple hidden layers is widely employed to denoise images and signals via dimension reduction. In machine learning systems, e.g., neural networks, non-linear transformations play an essential role in detecting relevant and suppressing irrelevant features present in the input data. Therefore, the single hidden layer and multiple hidden layer denoising autoencoders (SHL-DAE and MHL-DAE) with scaled activation function, e.g., hyperbolic tangent with various scale?, for denoising ECG signals are investigated in this work. SHL-DAE and MHL-DAE are trained and tested with ground truth, and Gaussian white noise (GWN) distorted ECG. The results show that MHL-DAE often yields better signal-tonoise-ratio (SNR) improvement for high noise level ECG signals than SHL-DAE. Furthermore, we found that selecting the activation function and the number of hidden neurons must be done carefully to achieve high denoising performance.