Efficient multi-class motor imagery EEG classification with cosine convolution and channel-group spatial feature extraction

Motor imagery brain–computer interfaces (MI-BCIs) have extensive applications in assistive control and neurorehabilitation. However, existing MI electroencephalography (EEG) decoding methods exhibit limited interpretability, high computational complexity, and insufficient extraction of temporal, frequency, and especially spatial information from EEG signals. To address these limitations, we propose a novel deep learning model with Cosine Convolution and Channel Group Spatial Convolution (CosCGSC) for multi-class MI-EEG classification. First, the raw MI-EEG signals are processed by a temporal cosine convolution module, which utilizes fixed-frequency cosine filters to extract interpretable frequency features. These features are subsequently fed into the channel group spatial convolution module, where predefined channel groups are convolved along the channel dimension to capture multiscale spatial patterns. Finally, the multiscale features are concatenated and passed through an information fusion convolution block which integrates the spatial information and performs further feature extraction to enhance classification accuracy. The proposed CosCGSC model was evaluated on the publicly available BCI Competition IV dataset IIa and BCI Competition III dataset IIIa, achieving competitive classification accuracies of 77.12% and 80.47%, respectively. Furthermore, we visualize the learned representations across temporal, frequency, and spatial domains using Grad-CAM and weight analysis, offering insights into subject-specific features and enhancing the interpretability of the end-to-end decoding process. The proposed CosCGSC model thus provides an efficient and interpretable framework for robust and personalized MI-BCI applications.