Spatio-Spectral Feature Representation for Motor Imagery Classification Using Convolutional Neural Networks.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Electroencephalogram (EEG)-based brain-computer interfaces (BCIs) decode user intentions noninvasively.
• Convolutional neural networks (CNNs) show promise for EEG analysis but face challenges due to high dimensionality and inter-subject variability.
• Existing methods struggle to preserve the multivariate structure and dependencies within EEG feature spaces.

Purpose of the Study:

• To develop a spatiospectral feature representation method for EEG data that preserves multivariate information.
• To enhance the decoding accuracy of CNN models in brain-computer interfaces.
• To address the variability issues in EEG signals across sessions, subjects, and trials.

Main Methods:

• Constructed 3-D feature maps by combining subject-optimized and subject-independent spectral filters.
• Stacked filtered EEG data into tensors to create spatiospectral representations.
• Implemented a layer-wise decomposition model within a 3-D-CNN framework for single-trial classification.

Main Results:

• Achieved average accuracies of 87.15% (±7.31) on BCI competition data IV_2a, 75.85% (±12.80) on IV_2b, and 70.37% (±17.09) on OpenBMI data.
• Demonstrated superior performance compared to state-of-the-art techniques.
• The decomposition model identified neurophysiologically plausible electrode channels and frequency domains, validating the approach.

Conclusions:

• The proposed spatiospectral feature representation effectively preserves multivariate EEG information.
• The 3-D-CNN framework with layer-wise decomposition provides reliable single-trial classification for BCIs.
• The method offers a robust solution for EEG-based BCIs, overcoming limitations of previous approaches.