A Gradient-Guided Spatio-Temporal Graph Convolutional Network for Population-Level Major Depressive Disorder Classification.

Abstract

Major depressive disorder (MDD) is a brain-related psychiatric disorder. Accurate diagnosis of MDD has long remained a challenge in clinical practice due to the high heterogeneity of its clinical manifestations. Graph neural network-based brain network analysis has shown great potential for MDD diagnosis. However, existing methods for constructing brain networks often overlook the continuous hierarchical organization of brain function or fail to effectively incorporate non-imaging information as a crucial complement for disease modeling. To address these, we propose a gradient-guided spatio-temporal graph convolutional network (GST-GCN) for population-level MDD identification. Specifically, GST-GCN first performs multi-scale modeling of the long-range dependencies and short-term dynamic fluctuations of blood oxygen level dependent signals along the temporal dimension, thereby capturing key temporal patterns associated with MDD. It then innovatively replaces conventional graph pooling with a gradient-guided hierarchical pooling to construct a spatial topology consistent with brain's intrinsic organizational principles. Within this well-structured spatial framework, the model is able to learn brain embeddings that more faithfully reflect individual functional organization characteristics. Furthermore, to incorporate non-imaging information, we integrate embeddings learned from individual brain graphs with non-imaging features to build a subject-level population graph. This explicit modeling of inter-subject relationships effectively enhances classification performance. Experimental results on the publicly available REST-meta-MDD dataset demonstrate that GST-GCN achieves a classification accuracy of 91.17%, with all performance metrics outperforming state-of-the-art methods. Moreover, the framework reveals alterations in functional gradients and identifies crucial biomarker regions for MDD classification, thereby exhibiting significant neurobiological interpretability.