Enhancing EEG Decoding with Selective Augmentation Integration.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Deep learning shows potential for electroencephalography (EEG) analysis.
• Challenges include scarce, noisy EEG data and limited data augmentation generality.
• Existing methods struggle with representational distortions and optimal augmentation selection.

Purpose of the Study:

• To develop an end-to-end EEG augmentation framework with adaptive mechanisms.
• To introduce NeuroBrain, a novel neural architecture for auditory EEG decoding.
• To enhance feature learning and mitigate representational distortions in EEG data.

Main Methods:

• Utilized contrastive learning to strengthen encoder feature learning and mitigate augmentation distortions.
• Incorporated a selective augmentation strategy for dynamic optimal augmentation combination determination.
• Introduced NeuroBrain, a neural architecture designed for capturing local and global dependencies in EEG signals.

Main Results:

• Demonstrated a 29.42% performance gain over HappyQuokka.
• Achieved a 5.45% accuracy improvement compared to EEGNet.
• Validated the framework's efficacy on SparrKULee and WithMe datasets.

Conclusions:

• The proposed framework and NeuroBrain architecture effectively address challenges in EEG analysis.
• The methods significantly improve performance in EEG decoding tasks.
• This work advances the state of the art in deep learning for EEG analysis.