Multiple ECG segments denoising autoencoder model

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Denoising autoencoders (DAEs) are effective for reducing noise in bio-signals like electrocardiograms (ECGs).
• Understanding the impact of data characteristics (correlated, uncorrelated, jittered) on DAE performance is crucial for optimizing bio-signal processing.
• Simultaneous Einthoven recordings (I, II, III) offer potential for leveraging signal correlations.

Purpose of the Study:

• To investigate the influence of correlated, uncorrelated, and jittered datasets on the performance of a denoising autoencoder (DAE) model for ECG signals.
• To determine the optimal number of hidden neurons for DAEs based on signal quality and computational efficiency using Akaike's information criterion.
• To evaluate DAE performance under various noise conditions, including mixed noise, motion artifacts, electrode movement, baseline wander, Gaussian white noise, and high-frequency noise.

Main Methods:

• ECG segments from simultaneous Einthoven recordings I, II, and III were concatenated to create correlated, uncorrelated, and jittered datasets.
• Akaike's information criterion was applied to find the optimal number of hidden neurons for DAEs.
• Datasets were corrupted with six types of noise to simulate real-world scenarios.
• Spectral analysis was employed to assess the impact of noise with overlapping power spectra on DAE performance.

Main Results:

• Denoising correlated ECG signals required significantly fewer hidden neurons compared to jittered signals.
• QRS-complex-based ECG alignment was found to be preferable for optimal denoising.
• Noises with slightly overlapping power spectra (e.g., baseline wander, high-frequency noise) were effectively removed with sufficient neurons.
• Gaussian white noise, with a completely overlapping spectrum, necessitated a low-dimensional, coarser signal reduction for effective recovery.

Conclusions:

• Leveraging correlations among simultaneous Einthoven I, II, and III ECG records improves DAE model performance.
• Enhanced signal-to-noise ratio and a reduced number of required hidden neurons are achievable by utilizing signal correlations.
• The study highlights the importance of data characteristics in optimizing DAE-based ECG denoising.