Distinguishable spatial-spectral feature learning neural network framework for motor imagery-based brain-computer...

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalogram (EEG) spatial and spectral features are crucial for motor imagery (MI) classification.
• Common Spatial Pattern (CSP) and Filter Bank CSP (FBCSP) are established methods for extracting these features from MI-related EEG.
• Improving the separability of CSP features is key to advancing MI-based Brain-Computer Interfaces (BCIs).

Purpose of the Study:

• To propose a novel Distinguishable Spatial-Spectral Feature Learning Neural Network (DSSFLNN) framework.
• To enhance the separability of spatial-spectral features for improved MI classification.
• To evaluate the DSSFLNN framework's performance against state-of-the-art methods in MI-based BCIs.

Main Methods:

• The DSSFLNN framework first extracts FBCSP features from raw EEG signals.
• It employs two squeeze-and-excitation modules for band-wise and class-wise re-calibration of CSP features.
• A parallel convolutional neural network module learns distinguishable spatial-spectral features, followed by a fully connected layer for classification.

Main Results:

• The DSSFLNN framework achieved a mean Cohen's kappa of 0.7 on BCI competition IV datasets 2a and 2b, outperforming state-of-the-art methods.
• Additional experiments confirmed that combining band-wise and class-wise feature learning yields significantly better performance than using either individually.
• The proposed framework demonstrates superior decoding performance for MI-based BCIs.

Conclusions:

• The DSSFLNN framework effectively enhances the discriminative power of spatial-spectral features for MI tasks.
• This approach offers a significant advancement in the performance of MI-based BCIs.
• The study highlights the benefit of integrated band-wise and class-wise feature learning strategies.