Wavelet-Based Denoising Diffusion Models for ECG Signal Enhancement

Diffusion models are emerging as a robust solution for achieving high-fidelity data denoising, often surpassing classical deep learning methods in quality under various conditions. Nonetheless, their slow training and inference speeds present a significant limitation, preventing their practical use in realtime applications. The objective of this paper is to improve the denoising performances of an existing diffusion model by introducing a new denoising diffusion scheme based on wavelets. In our scheme, we leverage the Discrete Wavelet Transform (DWT) to reduce the dimensionality of the input signals, before applying the denoising diffusion model. The proposed waveletdenoising diffusion model exhibits the capability to reconstruct clean ECG signals from corrupted ones. Also, the number of model parameters to be estimated has been reduced using the improved approach. The proposed improved model has been tested in a medical scenario using Electrocardiogram (ECG) data from the QT database, corrupted with real noises that are sourced from the MIT-BIH Noise Stress Test database. To assess the denoising performance, we employ metrics such as the Sum of the Square of Distances (SSD), the Absolute Maximum Distance (MAD), the Percentage Root-Mean-Square Difference (PRD), and the Cosine Similarity (Cosine Sim). The results of this work demonstrate the encouraging denoising capabilities of our proposed method, showcasing lower SSD, MAD, and PRD, along with higher Cosine Sim scores compared to the state-of-the-art methods.