Enhanced classification of motor imagery EEG signals using spatio-temporal representations

Deep learning has shown promising results in motor imagery brain-computer interfaces. However, most existing methods fail to account for the topological relationships between electrodes and the nonlinear features of electroencephalogram (EEG) signals. To address this, we propose a model combining Gramian Angular Fields (GAF) and Phase-Locking Value (PLV) with a parallel convolutional neural network (CNN). GAF captures time-domain nonlinear features, while PLV represents spatial features based on electrode topology. Comparative experiments between the end-to-end parallel CNN model and the model with spatiotemporal feature representation demonstrate that considering both time-domain correlations and electrode topology significantly enhances model performance. Furthermore, when separately evaluating the temporal and spatial features of EEG signals, the results confirm that jointly considering spatiotemporal features leads to a substantial improvement. On the Physionet dataset, our model achieves an accuracy of 99.73% in binary classification tasks and 83.37% in four-class classification tasks, showing improvement over the comparison algorithms used in the paper.