Modeling inter-subject heterogeneity improves multisite rs-fMRI identification of major depressive disorder: An interpretable graph neural network approach.
Authors
Affiliations (7)
Affiliations (7)
- College of Medicine and Biological Information Engineering, Northeastern University, Shenyang, China. Electronic address: [email protected].
- College of Medicine and Biological Information Engineering, Northeastern University, Shenyang, China. Electronic address: [email protected].
- College of Information Science and Engineering, Shenyang Ligong University, Shenyang, China. Electronic address: [email protected].
- College of Computer Science, Beijing University of Posts and Telecommunications, Beijing, China. Electronic address: [email protected].
- College of Medicine and Biological Information Engineering, Northeastern University, Shenyang, China. Electronic address: [email protected].
- College of Software, Northeastern University, Shenyang, China. Electronic address: [email protected].
- College of Medicine and Biological Information Engineering, Northeastern University, Shenyang, China. Electronic address: [email protected].
Abstract
Major depressive disorder (MDD) is a highly heterogeneous affective disorder, and the identification of reliable neuroimaging markers remains challenging. Resting-state functional magnetic resonance imaging offers a promising approach for characterizing altered brain functional organization in MDD. However, most existing studies model subjects independently and do not sufficiently consider inter-subject heterogeneity, particularly in multisite datasets. We developed an interpretable brain region-aware multimodal graph neural network (BRM-GNN) for rs-fMRI-based MDD identification using the multisite REST-meta-MDD dataset. After quality control, 1586 participants from 16 centers were included, comprising 821 MDD patients and 765 healthy controls. At the individual level, temporal features were learned from raw blood oxygen level-dependent signals and combined with functional connectivity information to characterize subject-specific brain networks. A multi-scale pooling fusion module was introduced to preserve local and global topological information while identifying disease-relevant brain regions. At the population level, a sex-aware heterogeneous population graph was constructed to model inter-subject relationships by integrating imaging and non-imaging information. Compared with existing graph-based methods, BRM-GNN achieved better identification performance while retaining neurobiological interpretability. The most informative regions were mainly located in prefrontal, cingulate, subcortical, and motor-related areas, suggesting that MDD may involve abnormalities in neural systems related to emotion regulation, reward processing, cognitive control, and psychomotor function. Our findings suggest that integrating subject-specific brain network features with sex-aware inter-subject relationships may improve multisite rs-fMRI-based MDD identification. This framework also highlights biologically plausible brain regions associated with MDD and may support future research on neuroimaging-informed biomarkers in affective disorders.