WG-Mamba: A Wavelet-Graph Mamba Network with MultiScale Spatiotemporal-Spectral Fusion for Major Depressive Disorder Diagnosis

Abstract

Major Depressive Disorder (MDD) is a mental disorder negatively impacting the lives of numerous patients worldwide. Currently, functional magnetic resonance imaging (fMRI) data are widely used for computer-aided MDD diagnosis. However, fMRI samples exhibit high-dimensional complexity and dynamic dependency, leading to suboptimal classification accuracy in traditional deep learning networks. To address these challenges, we propose WG-Mamba, a novel Wavelet-enhanced Graph Mamba network with multi-scale spatiotemporal-spectral fusion architecture, for feature extraction and classification of fMRI data between MDD patients and healthy controls (HCs). First, the frequency block is designed to decompose fMRI signals into multi-scale frequency features and filter noise through wavelet transform. Second, the spatial block utilizes a graph neural network with a brain region adjacency matrix to model cross-regional dependencies, capturing spatial features of the brain network. Third, the temporal block is designed based on Mamba to extract temporal dynamics. Finally, a dynamic weight mechanism is designed to achieve deep coupling of multi-scale features. The five-fold cross-validation on the REST-meta-MDD dataset shows WG-Mamba achieves 83.72% classification accuracy. The model also exhibits strong interpretability by identifying core brain regions associated with MDD, including the prefrontal cortex, anterior cingulate cortex, and limbic system, providing novel insights into MDD neurobiological mechanisms.