DeeBayes: An interpretable deep Bayesian network for ECG signal restoration

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Electrocardiogram (ECG) signals are crucial for diagnosing arrhythmias but are often corrupted by noise.
• Noise in ECG signals can lead to misdiagnoses, highlighting the need for effective denoising techniques.

Purpose of the Study:

• To develop a novel variational inference method for ECG signal denoising.
• To create a unified Bayesian framework combining noise estimation and signal restoration.
• To introduce the Deep Bayesian ECG Signal Restoration Network (DeeBayes) for accurate ECG denoising.

Main Methods:

• Developed the Deep Bayesian ECG Signal Restoration Network (DeeBayes) using deep learning and variational inference.
• Integrated data-driven deep learning with model-driven generative approaches for interpretability and generalization.
• Employed a unified Bayesian framework for noise estimation and signal denoising, handling non-independent and identically distributed (non-iid) noise.

Main Results:

• DeeBayes demonstrated superior performance in denoising ECG signals compared to state-of-the-art methods.
• The method effectively restored ECG signals across various signal-to-noise ratio (SNR) levels.
• Experimental results validated the model's ability to handle complex, non-iid noise patterns.

Conclusions:

• The proposed DeeBayes method offers an effective and adaptive solution for ECG signal denoising.
• This approach enhances diagnostic accuracy by providing cleaner ECG signals.
• DeeBayes represents a significant advancement in applying AI for medical signal processing.