Supervised Contrastive Learning-Based Domain Generalization Network for Cross-Subject Motor Decoding.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Decoding electroencephalogram (EEG) signals for motor imagery and execution (MI/ME) is crucial for brain-computer interfaces.
• Cross-subject decoding faces challenges due to domain shift, where data distributions vary between individuals.
• Existing domain adaptation methods are impractical when target subject data is unavailable.

Purpose of the Study:

• To develop a calibration-free, highly accurate cross-subject EEG decoding system for MI/ME.
• To address the domain shift problem in EEG-based brain-computer interfaces.
• To propose a novel supervised contrastive learning-based domain generalization network (SCLDGN).

Main Methods:

• Designed a feature encoder for discriminative EEG representation learning.
• Employed deep correlation alignment for domain-invariant feature extraction.
• Utilized supervised contrastive learning with domain-agnostic mixup for class-level alignment.

Main Results:

• The proposed SCLDGN effectively learns domain-invariant and class-relevant discriminative representations.
• Achieved superior performance in cross-subject MI/ME decoding across six datasets compared to state-of-the-art methods.
• Ablation studies and visualizations confirmed the method's generalization mechanism and revealed neurophysiologically meaningful patterns.

Conclusions:

• SCLDGN offers a robust solution for calibration-free, cross-subject EEG decoding.
• The approach enhances the practical applicability of brain-computer interfaces.
• The method demonstrates significant advancements in learning domain-agnostic and class-specific EEG features.