A Dimensionality Reduction Approach for Motor Imagery Brain-Computer Interface Using Functional Clustering and Graph Signal Processing.

Abstract

Background: This paper introduces an approach for dimensionality reduction and classification of electroencephalogram signals in motor imagery brain-computer interface (MI-BCI) systems.

Materials and methods: The proposed Kron-reduced generic learning regularization with differential evolution (K-GLR-DE) framework leverages graph signal processing (GSP) with a meta-heuristic optimizer, integrating functional clustering, Kron reduction, regularized common spatial patterns with generic learning (GLRCSP), and differential evolution (DE). Brain graphs are constructed within a structural-functional framework, where edge weights are defined based on geometric distances and correlations. Graph's dimensionality reduction is achieved by applying physiological regions of interest (ROIs) and Kron reduction to preserve essential topological-spectral features. Feature extraction is performed using graph total variation and GLRCSP, followed by DE-based feature selection.

Results: The approach was evaluated on BCI Competition III Dataset IVa and the PhysioNet eegmmidb dataset. The support vector machine with a radial basis function (SVM-RBF) classifier achieved superior performance, yielding a mean accuracy of 96.46% ± 0.81% on BCIC III-IVa.

Conclusions: The proposed K-GLR-DE method demonstrates significant performance in MI-BCI classification across various training conditions, including scenarios with small and limited training sets.