Decoding multiclass motor imagery EEG from the same upper limb by combining Riemannian geometry features and partial...

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Motor imagery (MI)-based brain-machine interface (BMI) systems face challenges in accurately classifying complex movements due to limitations in electroencephalogram (EEG) signal resolution and quality.
• Decoding multiclass MI EEG, especially from the same upper limb, presents significant obstacles for current classification methods.

Purpose of the Study:

• To develop a novel feature learning approach to enhance the classification accuracy of 6-class motor imagery tasks within the unilateral upper limb.
• To address the limitations of traditional methods like Common Spatial Pattern (CSP) and Filter-Bank CSP (FBCSP) in decoding complex MI EEG signals.

Main Methods:

• Utilized the Riemannian geometry (RG) framework to extract tangent space (TS) features from spatial covariance matrices of MI EEG trials.
• Applied partial least squares regression to reduce the dimensionality of TS features, creating more separable and compact representations.
• Validated the RG feature representations using linear discriminative analysis and support vector machine classifiers.

Main Results:

• Achieved an average classification accuracy of 80.50% with linear discriminative analysis and 79.70% with support vector machine classifiers.
• Demonstrated a significant increase in decoding accuracy for multiclass MI tasks from the same upper limb compared to CSP and FBCSP features.

Conclusions:

• The proposed Riemannian geometry-based feature learning approach offers a significant improvement over traditional methods for MI EEG classification.
• This method holds promise for advancing MI-based BMI applications, including robotic arm control and neural prosthetics for individuals with upper limb impairments.