Local-Global Spherical Sliced Wasserstein Graph Kernel for Schizophrenia Identification

Accurate identification of schizophrenia (SZ) using functional connectivity (FC) is fundamentally challenged by the non-Euclidean geometry of neuroimaging data and information loss from arbitrary graph thresholding. To overcome these bottlenecks, this study proposes the local-global spherical sliced Wasserstein graph (LG-SSWG) kernel. By embedding FC profiles onto a hyperspherical manifold, the proposed framework utilizes geodesic optimal transport to construct a mathematically rigorous, positive-definite graph kernel. This threshold-free approach synergistically integrates macroscopic whole-brain topology and region-specific functional architectures. Evaluated on the public COBRE dataset, LG-SSWG achieves a superior 84% accuracy, significantly outperforming state-of-the-art baselines. Beyond predictive accuracy, the method exhibits strong clinical interpretability by identifying critical neuroanatomical biomarkers, particularly the left thalamus and right precuneus. Network analysis reveals pervasive functional hyperconnectivity within subcortical relays and cognitive networks in SZ patients. Notably, this hyperconnectivity correlates positively with PANSS positive symptom scores. Consequently, these altered thalamocortical and precuneus-anchored circuits indicate that SZ clinical symptoms may be closely associated with disrupted sensory gating and functional segregation. Ultimately, the LG-SSWG kernel provides a principled geometric learning tool that advances both the objective diagnosis and mechanistic understanding of schizophrenia.