DepressionGraph: A Two-Channel Graph Neural Network for the Diagnosis of Major Depressive Disorders Using rs-fMRI

This study introduces DepressionGraph, a graph-based deep learning framework designed for accurate diagnosis of Major Depressive Disorder (MDD) using resting-state functional Magnetic Resonance Imaging (rs-fMRI). While prior studies have explored spatio-temporal Graph Neural Networks (GNNs), DepressionGraph distinguishes itself by implementing a two-channel architecture—Graph Convolutional Network (GCN) for spatial connectivity and Temporal GCN (T-GCN) for dynamic temporal analysis—tailored specifically to MDD pathology. Unlike earlier approaches, this framework explicitly integrates both static and dynamic patterns from functional connectivity graphs to improve diagnostic granularity. The model is trained and evaluated on publicly available rs-fMRI datasets from the REST-meta-MDD Consortium, employing a stratified 10-fold cross-validation and a separate hold-out test set to ensure generalizability. Graphs are constructed using Pearson correlation-based connectivity across 116 anatomical regions, and classification is performed via fused embeddings from the dual channels. The results demonstrate that DepressionGraph achieves 88.4% accuracy, 86.9% sensitivity, 89.6% specificity, and an AUC of 0.93, outperforming traditional models (SVM, Random Forest) and 2D CNN baselines, thus ensuring a fair comparison by aligning data dimensionality. Additionally, attention maps from the spatial channel provide interpretability by highlighting clinically relevant brain regions. DepressionGraph represents a robust, interpretable, and clinically viable advancement for non-invasive MDD diagnosis, with potential extensions to multimodal imaging and broader neuropsychiatric disorders.