Challenging Deep Learning Methods for EEG Signal Denoising under Data Corruption.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Electroencephalogram (EEG) signal acquisition is often hindered by noise, such as from human movement.
• Severe noise can lead to corrupted EEG channels, rendering them devoid of useful information.
• Deep learning methods show promise for denoising EEG signals.

Purpose of the Study:

• To conduct the first benchmark study on the performance of deep learning-based EEG signal denoising methods when faced with corrupted channels.
• To evaluate a wide variety of datasets, models, and evaluation tasks for EEG denoising.
• To provide a comprehensive assessment of current EEG denoising techniques.

Main Methods:

• Development of a benchmark study design to systematically evaluate EEG denoising algorithms.
• Inclusion of diverse EEG datasets representing various noise conditions and channel corruption scenarios.
• Application and assessment of multiple deep learning models for EEG signal denoising and corrupted channel imputation.

Main Results:

• Performance of EEG deep learning denoising models is highly dependent on the dataset used for evaluation.
• Significant variability exists in how different models handle corrupted EEG channels.
• The benchmark highlights specific challenges and limitations of current deep learning approaches in real-world noisy EEG data.

Conclusions:

• A standardized benchmark is crucial for reliably assessing EEG denoising methods.
• Future development of EEG deep learning models should prioritize robustness to corrupted channels.
• Performance evaluation across a broad suite of datasets is essential for advancing EEG signal processing.