DeeBayes: An interpretable deep Bayesian network for ECG signal restoration.

The electrocardiogram (ECG) is a valuable and non-invasive tool for detecting and preventing arrhythmias. However, in real-world situations, ECG signals are often contaminated by various types of noise, which can lead to clinical misdiagnoses. As a result, significant attention is given to developing methods to denoise ECG signals to ensure an accurate diagnosis and prognosis. This paper aims to develop a novel variational inference method that combines noise estimation and signal denoising within a unified Bayesian framework, specifically designed to effectively denoise ECG signals from any patient. Our method, the Deep Bayesian ECG Signal Restoration Network (DeeBayes), takes advantage of data-driven deep learning techniques, enabling efficient denoising through its explicit expression of posterior probabilities. Furthermore, DeeBayes incorporates the benefits of traditional model-driven approaches, particularly the strong generalization capabilities of generative models. This ensures that DeeBayes is both interpretable and adaptive for accurately estimating and removing complex non-independent and identically distributed (non-iid) noise patterns. Qualitative and quantitative experimental results conducted on noisy ECG signals with varying input signal-to-noise ratio (SNR) levels demonstrate that the proposed approach outperforms other state-of-the-art ECG signal restoration models, including those based on fully connected neural networks and convolutional neural networks. Source code is available at:https://github.com/marizvi/DeeBayes.