Disentangling high-dimensional functional brain network structure in major depression via cortically inspired sparse encoding and contrastive learning

Functional brain networks are high-dimensional complex systems whose large-scale organization is disrupted in major depressive disorder (MDD), yet characterizing these disruptions remains challenging due to the extreme dimensionality of neuroimaging data and the substantial entanglement of connectivity patterns between patients and healthy individuals. Here we introduce an integrated computational framework that combines cortically inspired sparse encoding with contrastive manifold learning to disentangle the high-dimensional structure of functional brain networks in MDD from resting-state functional magnetic resonance imaging (rs-fMRI) data. The Spatial Pooler from Hierarchical Temporal Memory (HTM-SP) — a neocortex-derived algorithm that implements nonlinear competitive inhibition and Hebbian self-organization — serves as an unsupervised sparse encoder that reduces dimensionality while preserving the topological relationships within the network. Contrastive principal component analysis (cPCA) subsequently separates disorder-specific variance from shared neural architecture by identifying directions of maximal between-group divergence in the high-dimensional connectivity space. Evaluated on a cohort of 70 MDD patients and 70 matched healthy controls, the HTM-SP ＋ cPCA pipeline achieved a classification accuracy of 86% and an F1-score of 87%, substantially outperforming conventional approaches. Critically, among the top 20 functional connections contributing most to the discriminative cPCA components, only 5 (25%) showed statistically significant group differences in traditional univariate testing, demonstrating that contrastive learning captures coordinated multivariate network alterations invisible to standard statistical approaches. The framework identified 129 discriminative functional connections, of which 34 exhibited significantly reduced connectivity in MDD, implicating distributed cross-network alterations spanning sensorimotor, dorsal attention, visual and frontal executive systems. The spatial distribution of these alterations — predominantly bridging separate functional modules rather than confined within individual subnetworks — suggests impaired integrative network organization consistent with systems-level models of depression as a disorder of large-scale network dysregulation. These results demonstrate that combining biologically grounded sparse encoding with contrastive manifold separation provides both enhanced diagnostic sensitivity and interpretable insights into the emergent network-level reorganization underlying MDD.