A Dimensionality Reduction Approach for Motor Imagery BCI using Functional Clustering, Graph Signal Processing, and Differential Evolution

This paper aims to address the dimension reduction and classification of electroencephalogram (EEG) signals within the context of motor imagery brain-computer interface (MI-BCI). By leveraging modern brain signal processing tools, specifically Graph Signal Processing (GSP) and meta-heuristic techniques, we introduce the K-GLR-DE approach. This methodology encompasses functional clustering, Kron reduction, regularized common spatial patterns with generic learning (GLRCSP), and differential evolution (DE). Our approach is underpinned by a comprehensive structural-functional framework that carefully shapes the architecture of the brain graph. Edge weights are assigned based on geometric distance and correlation, imbuing the model with physiologically meaningful connectivity patterns. Graph reduction involves strategically employing physiological regions of interest (ROIs) and Kron reduction to select informative subgraphs while preserving vital information from all graph vertices. Feature extraction integrates total variation calculation and the GLRCSP method, followed by dimension reduction using the DE algorithm. The extracted features are then evaluated using well-established machine-learning classifiers. The validation process is carried out using Dataset IVa from BCI Competition III, providing a tangible benchmark for the performance of the K-GLR-DE approach. Significantly, the SVM-RBF classifier stands out as the top performer, achieving a remarkable average accuracy of 96.46{+/-}0.83. Noteworthy is our approachs capacity to notably augment MI-BCI classification performance across diverse training trial scenarios, encompassing limited, small, and conventional settings.