Motor Imagery EEG Decoding Method Based on a Discriminative Feature Learning Strategy.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Deep learning methods for motor imagery electroencephalography (EEG) decoding are advancing rapidly.
• Existing methods often lack feature discrimination, limiting decoding accuracy.
• Overfitting is a significant challenge in deep learning-based EEG decoding.

Purpose of the Study:

• To propose a discriminative feature learning strategy to enhance EEG decoding accuracy.
• To address the limitations of classification loss in existing deep learning models.
• To mitigate overfitting in deep learning-based EEG decoding.

Main Methods:

• Introduced a discriminative feature learning strategy incorporating central distance loss (CD-loss), central vector shift, and central vector update.
• Developed a data augmentation method using a circular translation strategy to expand datasets.
• Validated the approach on two public motor imagery EEG datasets (BCI competition IV 2a and 2b).

Main Results:

• The proposed method significantly improved feature discrimination.
• The central distance loss and vector shift strategies enhanced inter-class separation.
• The circular translation data augmentation effectively expanded datasets without information loss.
• The method achieved the highest average accuracy and demonstrated good stability compared to state-of-the-art approaches.

Conclusions:

• The proposed discriminative feature learning strategy effectively improves EEG decoding accuracy.
• The combination of CD-loss, central vector shift, and data augmentation offers a robust solution for motor imagery decoding.
• The method shows promise for advancing brain-computer interface technologies.