A Data-Driven Gaussian Process Filter for Electrocardiogram Denoising.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Gaussian Process (GP)-based filters are effective for electrocardiogram (ECG) filtering but can be computationally intensive with arbitrary hyperparameters.
• Existing methods often lack data-driven optimization, leading to suboptimal performance in clinical applications.

Approach:

• Developed a data-driven GP filter utilizing the ECG phase domain, assuming Gaussian distribution for time-warped ECG beats.
• Simplified computations of mean and covariance matrices for efficient, hyperparameter-free GP filter implementation.
• Evaluated performance against a wavelet-based filter using the PhysioNet QT Database, measuring Signal-to-Noise Ratio (SNR) improvement and QT-interval estimation accuracy.

Key Points:

• The proposed GP filter significantly outperforms the benchmark wavelet filter across all tested noise levels (-5 to 30dB).
• Achieved superior results in reducing QT-interval estimation error bias and variance compared to the state-of-the-art filter.
• Demonstrated enhanced signal-to-noise ratio (SNR) improvement for the novel GP filter.

Conclusions:

• The data-driven GP filter is a versatile and efficient tool for ECG preprocessing in clinical and research settings.
• Applicable to ECG signals of varying lengths and sampling frequencies, providing reliable performance.
• Offers confidence intervals for performance, enhancing its utility in medical diagnostics and research.