Towards Real-Time ECG Signal Denoising using Sparse and Shallow Running Denoising Autoencoder

Abstract

Background: Electrocardiogram (ECG) is a widely used diagnostic tool in medical practice, but noise often compromises the quality of the recorded ECG signals, especially during prolonged monitoring like Ambulatory Cardiac Monitoring (ACM). Denoising Au- toencoders (DAE) have shown promise in biomedical signal denoising due to their ability to learn complex features. However, the current state-of-the-art DAE models need long non-aligned input segment with multiple hidden layers to achieve acceptable denoising, leading to complex models. This complexity poses challenges for model’s interpretability and for practical implementation in real-time scenarios, particularly on portable devices like ACM. Methods: Various DAE models utilizing either dense or convolutional layers have been examined with respect to different input lengths, including QRS-aligned, non-aligned, and overlapping ECG segments. In this work, a novel Sparse Running Denoising Autoencoder (SRunDAE) model is proposed to denoise relatively short ECG segments without the need for R-peak detection algorithms for segment alignment. The proposed RunDAE model employs a sliding-window processing approach, which takes into account the correlation between consecutive, overlapping ECG segments. This work investigates the effects of two weight regularization techniques, L1- and L2-norm, with the aim of enhancing denoising performance, achieving more interpretable, meaningful, and sparse representations, and simplifying the architecture of the current RunDAE model. Results and Discussions: The results indicate that the running DAE models outper- form the traditional DAE models in denoising ECG signals. Moreover, introducing spar- sity to RunDAE model help to improve the generalization of the model in removing un- seen noise/artifacts and retaining ECG morphology, and the interpretability of the model by learning meaningful features. The simplicity and the efficiency of SRunDAE model facilitate the implementation of reliable real-time ECG denoising on limited-resources mircrocontroller, like Arduino. Conclusions: The concept of running DAEs has shown a significant advancement com- pared to the traditional DAEs. By imposing sparsity, the SRunDAE model has delivered remarkable denoising and offers practical benefits such as improved generalization, inter- pretability, simplicity, and suitability for real-time implementation. Besides, the sparse DAEs yield weights that effectively acting as basis functions suitable for sparse coding.