Multiclass motor imagery classification with Riemannian geometry and temporal-spectral selection

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Brain-computer interfaces (BCIs) enable control of external devices using brain signals.
• Electroencephalography (EEG) is a common modality for BCI signal acquisition.
• Motor imagery (MI) decoding from EEG is crucial for intuitive BCI operation.

Purpose of the Study:

• To develop a robust feature extraction and selection framework for motor imagery (MI) based brain-computer interfaces (BCIs).
• To improve the accuracy and efficiency of decoding user intentions from EEG signals.
• To enhance the interpretability of features derived from MI EEG signals.

Main Methods:

• Application of Riemannian geometry to spatial-filtered covariance matrices for feature extraction.
• Development of a multiscale temporal-spectral segmentation scheme to enrich feature dimensionality.
• Utilization of a linear learning-based temporal window and spectral band (TWSB) selection method for optimal feature configuration.
• Classification of MI EEG signals using support vector machines (SVMs).

Main Results:

• Achieved average accuracies of 79.1% on BCI Competition IV dataset 2a and 83.1% on dataset 2b.
• Demonstrated an accuracy improvement of up to 6% by employing TWSB feature selection compared to using all features.
• Showcased significant reduction in computational burden through the TWSB selection method.
• Outperformed existing methods in decoding accuracy for MI-BCI.

Conclusions:

• The proposed framework provides interpretable feature information from motor imagery EEG signals.
• The method yields highly accurate and discriminative neural responses for MI-BCI.
• The approach facilitates improved performance in real-time motor imagery brain-computer interfaces.