Gaussian Mixture Connectivity with  \(\alpha \) -Renyi Regularization for EEG-Based MI Classification

Brain-computer interfaces (BCIs) have unveiled a transformative avenue for human computer interaction. So, motor imagery (MI) emerges as a compelling paradigm to mentally simulate movements without any overt physical execution. However, the inherent variability in brain signals, both inter and intra-subject, poses significant challenges to achieving robust classification. Here, we present the Gaussian Mixture Functional Connectivity Network with \(\alpha \) -Renyi Regularization (GMRRNet) for electroencephalography-based MI classification. GMRRNet employs a Gaussian mixture block made up of kernel-based functional connectivities with different bandwidths to extract discriminant spatial features. Also, it uses an \(\alpha \) -Renyi regularization to cut down on the mutual information between Gaussian-based channel matrices. Our method was evaluated on the GigaScience MI-EEG dataset, comprising recordings from 52 subjects performing left and right-hand MI tasks. The experimental results demonstrate that GMRRNet outperforms traditional models like EEGNet in terms of both accuracy and stability, particularly for subjects with medium to low MI skills. The model achieves an average accuracy of 71.7 \(\%\) , with a lower standard deviation, indicating improved consistency across trials. Finally, the spatial interpretability analysis highlights the GMRRNet capability to capture significant localized and global connections within the brain.