Enhanced EEG classification via Riemannian normalizing flows and deep neural networks

Electroencephalography (EEG) signals present unique challenges for classification tasks due to their non-stationary and high-dimensional nature. In this paper, we propose a novel method that combines Riemannian geometry with deep learning to classify multi-class EEG data. Specifically, we compute covariance matrices of EEG signals and map them onto the tangent space of the Symmetric Positive Definite (SPD) manifold. A deep neural network architecture, termed NeuroSPDNet, is designed to effectively learn from these tangent space features. The method incorporates cross-validation for robust performance evaluation and utilizes Receiver Operating Characteristic (ROC) and Precision-Recall (PR) curves to assess classification effectiveness across four classes. Experimental results demonstrate that the proposed approach achieves an overall accuracy of 96.23% with high precision and recall, significantly outperforming traditional approaches in EEG signal classification.