<![CDATA[Mu rhythm event-related desynchronization (ERD) is a key indicator of motor imagery activity based on EEG signals. However, accurate classification of ERD remains challenging due to the nonlinear nature of EEG signals and inter-session variability. This study proposes a motor imagery classification approach using a Spatio-Temporal Graph Neural Network (ST-GNN) model that leverages nonlinear time-frequency features extracted via Variational Mode Decomposition (VMD) and Synchrosqueezing Transform (SST). The dataset was collected from a single healthy subject across five separate sessions, each consisting of two conditions: relaxation and motor imagery. After preprocessing and segmentation, features were extracted and represented as spatio-temporal graphs to be processed by the ST-GNN. The model was evaluated using metrics such as accuracy, F1-score, AUC-ROC, and the Session Stability Index (SSI). The results show that the ST-GNN achieved an accuracy of 94.2%, F1-score of 94.1%, and AUC-ROC of 96.1%, along with high prediction stability across sessions. This performance outperformed baseline models including CNN, CSP+SVM, and STFT+MLP.These findings support the hypothesis that ERD is a distributed brain network phenomenon and demonstrate that the ST-GNN approach with VMD/SST-derived features is a promising strategy for developing adaptive and accurate BCI systems.]]>
Copyrights © 2026
