A class alignment network based on self-attention for cross-subject EEG classification.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalogram (EEG) signal variability necessitates subject-specific models.
• Existing domain adaptation methods for EEG focus on domain alignment, potentially neglecting crucial class boundaries.
• This can lead to weak feature-category correlations in classification tasks.

Purpose of the Study:

• To propose a novel adversarial learning model for bolstering subject-specific EEG classification.
• To leverage information from multiple subjects to improve individual EEG analysis.
• To address limitations in current domain adaptation strategies by focusing on both domain alignment and class separability.

Main Methods:

• Extracting shallow and attention-driven deep features from EEG signals.
• Employing a class discriminator with a novel discrimination loss function to align same-class features and diverge different-class features across domains.
• Utilizing two parallel, harmonized classifiers for joint decision-making.

Main Results:

• The proposed model effectively leverages multi-subject data for enhanced individual EEG classification.
• The adversarial strategy successfully aligns features across domains while maintaining class separability.
• Experimental validation on two public EEG datasets demonstrated the model's superior efficacy.

Conclusions:

• The novel adversarial learning approach significantly improves subject-specific EEG classification.
• The method effectively balances domain alignment and class discrimination for robust feature extraction.
• This work offers a promising direction for developing more accurate and reliable EEG analysis tools.