An explainable multi-modal graph learning framework with attention-based GCN-GAT for EEG-based depression detection

Major Depressive Disorder (MDD) is phenotypic and neurophysiologically heterogeneous, and developing reliable computational models for it requires careful control of information leakage, dataset heterogeneity, and contributions from different information sources. In this study, a subject-level multimodal population-graph learning framework is presented, in which each individual is modeled as a node and inter-subject relationships are directly incorporated into the learning process. The proposed framework uses a hybrid Graph Convolutional Network (GCN)–Graph Attention Network (GAT) framework under strict leakage-controlled inductive evaluation; Harmonization, preprocessing, feature selection, and graph construction in each fold are performed solely on the training data. In addition to classification, the study design allows for controlled modality contribution analysis, explicit assessment of dataset heterogeneity and graph topology, and multi-method explainability in a single framework. In multimodal analysis, the full model achieved an accuracy of 94.8 ± 7.5% and a Matthews Correlation Coefficient (MCC) of 0.907 ± 0.134, while additional analyses showed substantial cohort and modality dependence of performance. Train-fitted harmonization markedly attenuated linear dataset-origin separability, although nonlinear dataset-specific structure remained. Explainability analyses also showed high agreement of Integrated Gradients and GradientSHAP, with complementary information from Permutation Feature Importance. Overall, the main innovation of this study lies in the integration of subject-level relational modeling, leakage-controlled evaluation, controlled multimodal contribution analysis, heterogeneity and topology auditing, and multi-method explainability into a single integrated and auditable framework, rather than in the claim of absolute architectural superiority or a single performance metric.