Motor imagery task classification using spatial-time-frequency features of EEG signals: a deep learning approach for improved performance

Classification of electroencephalogram (EEG) signals according to the user-intended motor imagery (MI) task is crucial for effective brain-computer interfaces (BCIs). Current methods often encounter difficulties in attaining high classification accuracy. This study aims to improve accuracy by utilising spatial and time-frequency characteristics of multichannel EEG data using convolutional neural networks (CNN). EEG signals acquired from the sensory-motor region were subjected to time-frequency analysis, creating three-dimensional spatially informed time-frequency representations (SITFR). The CNN was trained and validated using SITFR matrices corresponding to four motor imagery tasks utilising the BCI Competition IV dataset IIa with a five-fold cross-validation technique. Gaussian noise data augmentation was applied to improve model robustness by increasing variability in EEG signals while preserving their structural integrity. Four time-frequency approaches, namely continuous wavelet transform (CWT), wavelet synchrosqueezed transform (WSST), Fourier synchrosqueezed transform (FSST) and synchroextracting transform (SET) were used for this experiment. The CNN model attained a mean test accuracy of 98.18% and kappa score of 0.98 for CWT-SITFR, outperforming other TFR methods. The accuracies obtained for FSST, WSST and SET were 97.47%, 94.38% and 91.82% with kappa scores of 0.97, 0.93 and 0.89 respectively. This approach enables the CNN to learn both time-frequency and spatial features, resulting in better performance compared with existing state-of-the-art techniques.