K-Medoids-based ICA and Wavelet-Enhanced Multi-resolution Convolutional Network for Motor Imagery EEG

In BCI systems, addressing the limitations where raw MI-EEG signals exhibit insufficient representation of task-related features in spatial-temporal and time-frequency domains with heavy reliance on manual intervention, and existing convolutional neural networks struggle to simultaneously capture local EEG features and global long-term neural oscillation patterns, we propose an end-to-end compact deep convolutional neural network. Our model integrates multi-branch convolutional networks with temporal convolutional networks to extract multi-resolution features. By combining KMedoids ICA and discrete wavelet transform, the original EEG signals are converted into a feature matrix composed of multi-layer wavelet coefficients. Experimental results under subject-independent validation demonstrate that our method achieves classification accuracies of 70.08% and 70.16% on BCI Competition IV-2a and Physionet four-class motor imagery datasets respectively, outperforming other classical deep learning models. Feature visualization reveals clearer inter-class separation and higher intra-class consistency compared with alternative approaches.