Multi-Domain Feature Fusion with OVO-SVM for Multi-Class Motor Imagery EEG Classification and Motor Control Validation

The number of people suffering from severe motor disabilities due to stroke, multiple sclerosis, and spinal cord injuries is projected to reach 2–4% of the global population, highlighting the need for effective brain-computer interfaces (BCIs). This work proposes a robust multi-domain framework for four-class motor imagery (MI) EEG classification involving right hand, left hand, foot, and tongue imagery. The system fuses spatial, spectral, temporal, and nonlinear information using common spatial pattern (CSP), power spectral density (PSD), log-variance, wavelet transform, and bispectrum features. Classification is performed with a one-vs-one support vector machine (OVO-SVM) to enhance class separability and mitigate class imbalance. Beyond offline classification, an end-to-end EEG-driven control architecture is implemented that maps decoded MI decisions into coordinated control commands for two independent motors, enabling continuous translation of cognitive intentions into physical motion. Experiments on BCI Competition IV dataset 2a demonstrate strong performance, with the best subject achieving 95.83% accuracy and 0.94 kappa, and average accuracy and kappa of 74.3% and 0.65, respectively. System-level validation using an independent evaluation dataset confirms stable temporal prediction behavior, reliable motor command generation, and coordinated multi-motor control, indicating the framework’s suitability for real-world MI-based BCI applications.