EEG CHANNEL SELECTION USING EVOLUTIONARY ALGORITHM FOR BCI APPLICATION

Abstract

Brain-Computer Interfaces (BCIs) offer a revolutionary means of communication between the human brain and external systems, enabling individuals to control devices without the need for any muscular activity. This technology holds immense potential in various domains, particularly in neurorehabilitation, where it assists patients with motor impairments, and in assistive technologies, providing control over prosthetic limbs, wheelchairs, and communication devices. Among the various BCI paradigms, Motor Imagery (MI) has emerged as one of the most promising approaches. In MI- based BCIs, users are trained to imagine specific movements—such as moving the left or right hand—without performing any actual physical motion. To monitor brain activity, Electroencephalography (EEG) is the most widely used modality in MI-based BCIs. EEG captures electrical activity from the scalp with high temporal resolution, making it particularly effective for tracking the fast neural dynamics associated with motor imagery. Additionally, EEG is non-invasive, portable, and cost-effective, offering significant advantages over other neuroimaging techniques such as fMRI and MEG. However, EEG signals are inherently high-dimensional and noise-prone, with artifacts stemming from muscle movement, eye blinks, and external interference. These challenges necessitate robust feature extraction and channel selection techniques to ensure accurate and efficient classification of MI tasks. Identifying the most informative channels and transforming the raw EEG into meaningful features are critical steps in reducing redundancy, improving signal quality, and enabling reliable real-time performance. In this work, we propose an enhanced EEG-based framework for binary motor imagery classification, focusing specifically on distinguishing left- vs. right-hand imagery. To extract discriminative features, we employ Advanced Graph Signal Processing (AGSP), a novel approach that treats EEG signals as data on a graph, where nodes represent EEG channels and edges capture functional connectivity between brain regions. This graph-based representation enables the extraction of features that incorporate both spatial and structural information, offering deeper insights into brain dynamics compared to traditional time-series analysis. AGSP leverages graph spectral transforms to highlight connectivity-driven neural patterns relevant to MI classification. To further improve system performance, we implement the Set-based Integer-coded Fuzzy Granular Evolutionary (SIFE) algorithm for intelligent channel selection. SIFE utilizes swarm intelligence to explore the search space and select the most informative subset of EEG channels, effectively reducing dimensionality while preserving key discriminative features. This not only boosts classification accuracy but also reduces computational cost and enhances the system’s real-time capabilities. For the final decision-making, we adopt an ensemble classification strategy by integrating multiple classifiers such as Random Forest, XG-Boost, and Ada-Boost. Ensemble learning enhances robustness and generalization by combining the strengths of individual models. The integration of AGSP for structural feature extraction, SIFE for optimized channel selection, and ensemble learning for robust classification results in a highly efficient and accurate framework for decoding binary motor imagery tasks from EEG signals.