EEG Classification Using Contrastive Learning and Riemannian Tangent Space Representations

This paper introduces a novel advanced framework for the classification of Electroencephalography (EEG) signals through the integration of Riemannian geometry and a bespoke contrastive learning framework. Initially, the EEG signals are segmented and converted into covariance matrices, which are regularized to ensure positive definiteness, thus enabling their interpretation as Symmetric Positive Definite (SPD) matrices. These SPD matrices are then projected onto the Riemannian manifold, facilitating the extraction of discriminative features by utilizing Euclidean geometric operations within the tangent space. A specialized neural network architecture, TangentSpaceNet, is devised to map these features into a lower-dimensional space. In this space, a customized contrastive loss function, predicated on Euclidean distance, is implemented to enhance class separability. This approach substantially improves the robustness and accuracy of EEG signal classification. Empirical assessments highlight the effectiveness of the suggested approach, showcasing significant improvements in classifying performance 93% accuracy, 94%precision, 94% recall, and 94% F1-score. The novel fusion of Riemannian geometry and contrastive learning presents a sturdy and adaptable structure for analyzing EEG signals, presenting considerable prospects for utilization in brain-computer interfaces, cognitive neuroscience, and the identification of neurological conditions.