Uppe Nanaji , Yam Krishna Poudel 2 1 ( January - March ) Although significant advances have been made in classification of electroencephalograms EEGs in biomedical signal analysis supervised approaches require high numbers of labeled data which are expensive and challenging to collect in medical environments Selfsupervised learning has become an exciting alternative as it can be used to obtain useful representation learning from recordings that are not labeled In this paper a selfsupervised geometric learning framework is proposed which combines the Symmetric Positive Definite SPD manifold representation and contrastive learning objective for robust EEG classification Multichannel EEGbased covariance structures are projected onto Riemannian manifolds where the nonlinear relationships between neural signals are preserved The objective which is a combination of the geometric and selfsupervised objectives significantly decreases the dependency on annotation by 65 and achieves a classification accuracy of 982 outperforming CNN Transformer and fully supervised baselines The proposed framework shows high generalizability and is applicable to motor imagery mental arithmetic and clinical braincomputer interface BCI tasks Copyright (c) World Journal of Pharmaceutical Seiences. All rights reserved 2026