Dual-Manifold Contrastive Learning for Robust and Real-Time EEG Motor Decoding

Area of Science:

• Neuroscience and Biomedical Engineering
• Machine Learning for Signal Processing

Background:

• Brain-computer interfaces (BCIs) offer potential for consumer electronics but face challenges in accuracy, stability, and real-time processing.
• Current electroencephalography (EEG) decoding methods struggle with electrode shifts and require efficient feature extraction.

Purpose of the Study:

• To develop a hybrid decoding framework addressing limitations in current EEG-based BCIs.
• To enhance decoding accuracy, stability, and real-time performance for human-computer interaction.

Main Methods:

• A hybrid framework combining manifold learning (non-negative matrix factorization) and contrastive learning.
• A dual-manifold model for feature extraction and a joint training strategy for improved stability.
• Optimization for real-time interaction with a system latency of 100 ms.

Main Results:

• Achieved superior decoding performance on a constructed motor EEG dataset (F1-scores of 0.7382 for motor imagery, 0.8361 for motor execution).
• Demonstrated robustness against reduced electrode counts and altered spatial distributions.
• Significantly reduced system latency for real-time BCI applications.

Conclusions:

• The proposed hybrid decoding framework offers a promising solution for reliable and portable BCI systems.
• The method enhances EEG signal decoding, paving the way for more effective human-computer interaction.
• The framework's stability and accuracy show potential for widespread adoption in consumer electronics.