Self-supervised contrastive learning for EEG-based cross-subject motor imagery recognition.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Electroencephalography (EEG) is widely used in brain-computer interfaces (BCIs) due to its non-invasive nature and high-resolution data.
• EEG datasets often suffer from scarcity and require extensive labeling, limiting model generalization due to inter-individual variability.

Purpose of the Study:

• To develop a novel self-supervised contrastive learning framework for decoding motor imagery (MI) signals in cross-subject scenarios.
• To address data scarcity and improve the generalization performance of EEG-based BCIs.

Main Methods:

• A convolutional neural network and attention mechanism encoder was designed for self-supervised contrastive learning.
• The framework uses a pretext task of data augmentation to train the network, minimizing distances between similar data transformations and maximizing distances between dissimilar ones.
• This method extracts deep features from EEG signals without requiring labeled data.

Main Results:

• The proposed framework achieved cross-subject classification accuracies of 67.32% (BCI IV IIa), 82.34% (BCI IV IIb), and 81.13% (HGD).
• Performance surpassed existing methods on three public motor imagery datasets.
• Demonstrated superior efficacy in decoding motor imagery signals.

Conclusions:

• The self-supervised contrastive learning method shows significant promise for enhancing cross-subject transfer learning in motor imagery-based BCIs.
• This approach effectively mitigates challenges associated with EEG data scarcity and inter-individual variability.
• Offers a pathway to more robust and generalizable BCI systems.