Content area
Stroke affects millions worldwide, leading to severe motor and cognitive impairments. Effective rehabilitation is essential but labor-intensive. Robotic exoskeletons integrated with Brain-Computer Interfaces (BCIs) using Electroencephalography (EEG) enable user-driven rehabilitation, reducing therapist workload. However, real-time EEG classification remains challenging due to signal complexity.
This thesis develops EEG GLT-Net, a spectral Graph Neural Network (GNN) for real-time classification of EEG Motor Imagery (MI) signals at single time points ( 1/160 s). It introduces the EEG Graph Lottery Ticket (EEG GLT) method, which dynamically constructs adjacency matrices without prior knowledge of EEG channel relationships, improving accuracy and efficiency. Evaluation on PhysioNet shows superior performance over state-of-the-art (SOTA) methods.
Beyond stroke rehabilitation, EEG GLT is applied to economic forecasting, demonstrating its adaptability. Additionally, EEG Synergistic Gated Network (EEG SGNet), a CNN-GNN hybrid, enhances window-based EEG classification, validated on BCIC iv-2a and HGD datasets. Lastly, EEG RL-Net, a reinforcement learning model, optimises classification by selectively skipping uncertain time points, improving computational efficiency.
These contributions advance EEG-based rehabilitation, enabling intelligent, adaptive systems that enhance stroke recovery and broader neurorehabilitation applications.