Deep Learning Models for Denoising ECG Signals

Abstract

Effective and powerful methods for denoising electrocardiogram (ECG) signals are important for wearable sensors and devices. Deep Learning (DL) models have been used extensively in image processing and other domains with great successes but only very recently they have been used in processing ECG signals. This paper presents two DL models, together with a standard wavelet-based technique for denoising ECG signals. First, a Convolutional Neural Network (CNN) is depicted and applied to noisy ECG signals. It includes six convolutional layers, with subsequent pooling and a fully connected layer for regression. The second DL model is a Long Short-Term Memory (LSTM) model, consisting of two LSTM layers. A wavelet technique based on an empirical Bayesian method with a Cauchy prior is also applied for comparison with the DL models, which are trained and tested on two synthetic datasets and a dataset containing real ECG signals. The results demonstrate that while both DL models were capable of dealing with heavy and drifting noise, the CNN model was markedly superior to the LSTM model in terms of the Root Mean Squared (RMS) error, and the wavelet technique was suitable only for rejecting random noise.