Dynamic coarsened spatio-temporal graph convolutional networks for fMRI classification of addiction-induced sleep disorders

Abstract

Abstract Substance Use Disorder (SUD) is frequently characterized by persistent sleep disturbances that hinder cognitive recovery. Accurately identifying these disruptions requires methods capable of tracking both spatial interactions and the temporal evolution of brain activity. While functional magnetic resonance imaging (fMRI) provides offer valuable insights into broad neural patterns, conventional fMRI classification methods typically rely on static connectivity and predefined knowledge. These conventional approaches fail to account for the complex, time-varying nature of brain networks, thereby posing significant challenges for accurate fMRI classification. To overcome these limitations, this paper proposed the Dynamic Coarsened Spatio-Temporal Graph Convolutional Network (DC-STGCN), a data-driven framework designed to models the brain as a time-evolving directed graph and extracts highly distinct data features. The framework achieves this by integrating two key components: a dynamic causal learning module that maps evolving connections over time, and a noise-reducing Gaussian mixture model. Experimental results indicate that the proposed approach outperforms existing methods in classifying fMRI data related to addiction-induced sleep disturbances, offering a robust solution for high-dimensional fMRI analysis in clinical neurology.