Self-supervised learning for electroencephalography (EEG) aims to acquire generalizable and transferable representations of brain activity. However, existing methods often yield structurally incomplete representations by failing to fully exploit the inherently multi-modal nature of neurophysiological signals across the spatial, temporal, and spectral domains. In this paper, we introduce DeFuseMod, a foundation model designed to learn holistic and structurally isomorphic representations of brain dynamics. DeFuseMod employs a Transformer architecture guided by a “disentangle-then-fuse" principle. It first decouples heterogeneous spatio-temporal features into parallel processing streams, which are subsequently reintegrated through a hierarchical fusion mechanism. This approach captures rich cross-domain interactions while preventing premature feature amalgamation. The architecture is fed by a neurophysiologically-grounded embedding layer that generates a dual-domain representation by concurrently capturing local temporal morphologies and spectral power distributions across classical frequency bands. For training, we enforce structural isomorphism through a masked reconstruction objective that penalizes not only point-wise reconstruction errors (Mean Squared Error, MSE) but also deviations in spectral composition and spatial covariance structure. Empirical evaluations across a comprehensive set of downstream Brain-Computer Interface (BCI) tasks demonstrate that DeFuseMod achieves state-of-the-art performance. This indicates that its architecture and learning paradigm produce representations that are both high-performing and faithful to the underlying neurophysiological dynamics.