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

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Brain-Computer Interfaces (BCIs) offer non-invasive control via brain activity, with Motor Imagery (MI) systems being prominent.
• Electroencephalography (EEG) is favored for its portability and temporal resolution, but its non-stationary and subject-specific signals present classification challenges.
• Existing BCI systems struggle with reliable classification due to the inherent variability and complexity of EEG data.

Purpose of the Study:

• To develop a lightweight and efficient classification architecture for MI-based BCIs.
• To enhance the accuracy and generalization of EEG signal classification.
• To address the challenges of non-stationarity and inter-subject variability in EEG data.

Main Methods:

• A novel architecture combining Event-Related Desynchronization (ERD) band selection, Empirical Mode Decomposition (EMD), Hilbert-Huang Transform (HHT), Riemannian Geometry (RG), and Common Spatial Pattern (CSP) feature extraction.
• Integration of a Broad Learning System (BLS) classifier with parameters optimized by the Sand Cat Swarm Optimization (SCSO) algorithm.
• Utilized two datasets (BCI IV 2a and a clinical stroke EEG dataset) for performance evaluation, including All-subjects and Leave-One-Subject-Out (LOSO) validation.

Main Results:

• Achieved high classification accuracies: 90.78% on the BCI IV 2a dataset and 96.41% on the clinical stroke EEG dataset.
• Demonstrated competitive generalization performance across different validation settings, indicating robustness to inter-subject variability.
• The proposed pipeline effectively extracts discriminative features and handles individual differences in EEG signals.

Conclusions:

• The proposed lightweight and efficient BCI classification architecture significantly improves accuracy and generalization.
• The integration of EMD, HHT, RG, and CSP with an SCSO-optimized BLS classifier effectively models nonlinear EEG dynamics and addresses inter-subject variability.
• This approach shows strong potential for real-world BCI applications, particularly in motor imagery tasks and clinical settings like stroke rehabilitation.