Motor Imagery Classification with Covariance Matrices and Non-Negative Matrix Factorization.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Motor imagery (MI) classification using electroencephalography (EEG) is crucial for brain-computer interfaces (BCIs).
• Selecting optimal EEG channels and features is vital for improving classification accuracy and reducing computational complexity.
• Existing methods often struggle with high dimensionality and overfitting, necessitating more efficient channel and feature selection techniques.

Purpose of the Study:

• To identify the minimal set of EEG channels for highly accurate motor imagery classification.
• To maintain an optimal Kappa score while reducing channel and feature redundancy.
• To enhance classification accuracy and reduce model overfitting through subject-specific channel selection.

Main Methods:

• Non-negative matrix factorization (NMF) for identifying important and discriminant EEG channels.
• Feature extraction utilizing Riemannian geometry on the manifold of covariance matrices.
• Neighborhood Component Feature Selection (NCFS) algorithm for selecting a concise subset of relevant features.

Main Results:

• Achieved 77.91% average classification accuracy on the BCI Competition IV,2a dataset.
• Obtained a mean Kappa score of 0.626.
• Demonstrated significant reduction in time complexity and overfitting by minimizing EEG channels and redundant features.

Conclusions:

• The proposed method effectively reduces computational load and overfitting in EEG-based motor imagery classification.
• Subject-specific EEG channel selection enhances model accuracy and robustness.
• This approach offers a promising strategy for optimizing BCI performance through efficient channel and feature selection.