A deep learning-based framework For ECG signal denoising based on stacked cardiac cycle tensor

• Published in: Biomedical signal processing and control 2022-01, Vol.71, p.103275, Article 103275
• Main Authors: Rasti-Meymandi, Arash, Ghaffari, Aboozar
• Format: Article
• Language: English
• Subjects: Convolutional neural network; Deep model; Denoising; ECG; Non-local cycles
• ISSN: 1746-8094; 1746-8108
• DOI: 10.1016/j.bspc.2021.103275

•We have considered the pseudo-periodicity of the ECG signal by detaching each cardiac cycle and creating a stacked cardiac cycle (SCC) tensor to form a 2-dimensional signal.•We proposed a deep neural network modified with novel local/non-local cycle observation modules (LNC modules) to account for the non-local cycles of ECG signal.•The proposed method shows a significant improvement over the CNN model that only considers the 1-dimensional time series of ECG signals. The Electrocardiogram (ECG) signal is one of the frequently used non-invasive physiological measurement techniques for heart diagnosis. However, ECG signal is often contaminated with various noise and artifacts which make the diagnosis a challenging task. Recent deep learning models have had promising results in dealing with the noises, however, they only considered the 1D time series of the ECG signal. This paper presents a novel deep learning-based Electrocardiogram (ECG) denoising approach based on the periodicity of the ECG signals. In this work, ECG cardiac cycles are stacked together to form a 2D signal which will be fed to a convolutional neural network (CNN) model. Accordingly, the correlation between cardiac cycles can be exploited, resulting in an efficient and robust ECG denoising. The proposed CNN model is equipped with a novel local/non-local cycle observation (LNC) module to account for the correlation between the cycles. The proposed framework is applied to the publicly available MIT-BIH Arrhythmia database. Various experiments on different noise conditions have been conducted to evaluate the effectiveness of the design. The results have shown the superiority of our framework over the existing state-of-the-art approaches in terms of the Root-Mean-Square Error (RMSE) and improvement in Signal-To-Noise Ratio (SNRimp).