Sparse dictionary learning neural networks for ECG signal denoising

Electrocardiography (ECG) signals are prone to different types of interference, which can impede evaluation and lead to false diagnosis. In this paper, we propose a novel hybrid approach for effectively removing baseline noise from ECG signals based on sparse dictionary learning (SDL). First, we use the K-singular value decomposition (K-SVD) algorithm to generate a dictionary of basis functions for typical types of ECG noise. Then, we employ a model-driven neural network based on the FISTA-Net architecture to approximate the noise in a given ECG signal as a linear combination of the dictionary atoms. Finally, the denoising is done by subtracting the reconstructed noise from the signal. Both steps of the proposed approach are learned from small datasets, while maintaining the interpretability of SDL. Our experiments using the Brno University of Technology ECG Quality Database demonstrate the superiority of this method compared to other relevant techniques, including quadratic programming and end-to-end deep learning, in both denoising performance and computational efficiency. Source code is available at https://github.com/galigergergo/ECGDenFISTA-Net.Clinical relevance-ECG is the basis for cardiovascular assessment but signal noise is a potential bias, especially in remote monitoring. False diagnosis might lead to adverse consequences for patients; therefore, noise removal is of utmost importance for patient safety.