Multi-domain feature extraction and Sand Cat Swarm Optimized Broad Learning System for EEG-based Motor Imagery...

Brain-Computer Interfaces (BCIs) enable the translation of brain activity into executable commands, with Motor Imagery (MI)- based systems gaining prominence for their intuitive and non-invasive control. Electroencephalography is widely used due to its portability and time resolution, though its non-stationary and subject-specific nature poses major challenges for reliable classification. This research proposes a lightweight and efficient classification architecture that first selects discriminative filter bands based on Event-Related Desynchronization (ERD) scores. It then integrates Empirical Mode Decomposition (EMD), the Hilbert-Huang Transform (HHT), Riemannian Geometry (RG), and Common Spatial Pattern (CSP)-based feature extraction with a Broad Learning System (BLS) classifier. The BLS parameters are optimized using the Sand Cat Swarm Optimization (SCSO) algorithm to enhance convergence speed, avoid local minima, and improve generalization. EMD separates the EEG signal into a set of Intrinsic Mode Functions, while HHT extracts instantaneous amplitude and frequency features, effectively modeling the nonlinear and dynamic properties of EEG signals. Performance assessment was done on two datasets: the BCI IV 2a dataset and a clinical stroke EEG dataset. It achieved classification accuracies of 90.78% on BCI-IV 2a and 96.41% on the stroke dataset. The proposed approach also showed competitive generalization performance in All-subjects and Leave-One-Subject-Out (LOSO) validation settings. Analysis reveals that the proposed pipeline effectively extracts discriminative features and handles inter-subject variability, illustrating its applicability to real-world BCI systems.