FL-SDGIN: A federated graph learning approach for schizophrenia diagnosis integrating static and dynamic brain functional networks.

Recent magnetic resonance imaging (MRI) studies have revealed connectivity abnormalities in brain networks of schizophrenia (SZ). Graph Neural Networks (GNN) through their powerful graph embedding ability provide novel approaches for brain network analysis in SZ. However, current functional MRI (fMRI) based SZ diagnostic models exhibit limitations including insufficient utilization of static and dynamic functional connectivity (FC/dFC), inadequate modeling of dynamic features while neglecting temporal variability, and lack of consideration for multi-site data privacy and heterogeneity. To address these issues, we propose a federated learning-based static-dynamic graph isomorphism network (FL-SDGIN) for SZ diagnosis. The framework first employs temporal convolutional networks to extract temporal variability of dFC, constructing temporal variability guided attention adjacency matrices. Dynamic graph isomorphism networks (DyGIN) then capture spatiotemporal topological patterns, while graph isomorphism networks (GIN) extract static topological features, enabling multidimensional characterization of brain networks. Simultaneously, federated averaging (FedAvg) is employed for cross-site model training while avoiding direct sharing of raw imaging data. Experimental results under random cross-validation show that FL-SDGIN achieves a classification accuracy of 0.815 and outperforms the evaluated baseline models. Additional leave-one-site-out analysis indicates that cross-site generalization remains challenging under site and cohort heterogeneity. The interpretability analysis suggests that candidate SZ-related regions include the thalamus, posterior cingulate gyrus, postcentral gyrus, middle temporal gyrus, and superior temporal gyrus.