A Graph Neural Network Approach to Study the Human Connectome in Male Individuals with Autism Spectrum Disorders

Graph Neural Networks (GNNs) have emerged as powerful tools for analyzing brain graphs, gaining increasing attention in neurological and psychiatric disorder research. By modeling the brain as a graph, where nodes represent anatomical regions and edges represent functional connections based on specific parcellations, GNNs enable the integration of multiparametric data, providing a richer understanding of brain structure and function. In this study, we focused on the classification of Autism Spectrum Disorders (ASD) using data from the ABIDE database (only males, age-range 5–35 years). We extracted morphological features from structural MRI and time-series from resting-state fMRI, using the Destrieux atlas for brain parcellation. These data were combined to construct functional connectome-based graph representations. The classification task was formulated as an atlas-based binary whole-graph classification to distinguish ASD from typically developing (TD) individuals. We applied a SAGPooling-GraphSAGE model, analyzing the impact of edge thresholding, which controls graph sparsity and the number of nodes selected through pooling. Using nested cross-validation, our model achieved an AUC of (72.2 ± 1.8)%, demonstrating effective binary classification performance. Moreover, the model identified critical brain subgraphs and connections that contribute to classification, highlighting their relevance. Our results confirm that GNNs offer reasonable performance in brain graph analysis and provide valuable insights into ASD-related brain alterations.