Deep-learning based electroencephalogram denoising: a literature review

Area of Science:

• Neuroscience
• Signal Processing
• Artificial Intelligence

Background:

• Electroencephalography (EEG) is crucial for brain function research.
• EEG signals are prone to artifacts (e.g., electromyogram, electrocardiogram, electrical noise) that impede analysis.
• Existing deep learning methods show promise in EEG denoising but lack comprehensive review.

Purpose of the Study:

• To bridge the literature gap by reviewing deep learning-based EEG denoising strategies.
• To analyze the end-to-end denoising pipeline's influence on model performance and utility.
• To identify future research directions for improved EEG denoising systems.

Main Methods:

• Framework analysis of the end-to-end denoising pipeline.
• Examination of data/target construction, input representation, architecture, objective design, and evaluation.
• Discussion of selective/multi-task denoising, downstream validation, and model deployment.

Main Results:

• Deep learning approaches offer competitive reconstruction fidelity and artifact suppression.
• Pipeline components critically influence model assumptions, performance interpretation, and practical utility.
• Translating reconstruction performance to usable applications requires careful consideration of validation and deployment.

Conclusions:

• A comprehensive review of deep learning for EEG denoising is needed.
• Optimizing denoising pipelines is key to enhancing EEG data reliability.
• Future work should focus on developing interpretable and practical EEG denoising solutions.