Graph Neural Networks (GNNs) are powerful tools for analyzing brain graphs. GNNs are attracting growing interest in the study of neurological and psychiatric disorders. Based on specific parcellation, graphs can model the brain, with nodes as anatomical regions and edges representing functional connections. Combining multiparametric information in the same graph enriches representations, offering a more comprehensive view of brain function and structure.

We focused on the analysis of Autism Spectrum Disorders (ASD) and analyzed ABIDE data (males, 5–40 years) extracting morphological features from structural MRI (sMRI) and time-series from resting-state fMRI (rs-fMRI) based on the Destrieux atlas. This data was used to construct graph-structured representations based on the functional connectome. The problem was framed as an atlas-based binary whole-graph classification task, aiming to distinguish ASD from typically developing (TD) individuals based on differences in their brain graphs. Then, we identified key subgraphs for classification.

We examined the performance of a SAGPooling-GraphSAGE model, focusing on the effects of two parameters: the edge threshold, which determines the number of connections in the brain graph, and the number of nodes used in the classification, which are extracted through the pooling mechanism.

Employing nested cross-validation, our results show a promising Area Under the Curve (AUC) score of 71.8±1.8, highlighting the model effectiveness in binary classification with brain graph parameters. Findings highlight anatomical regions and connections as key to model performance.

Our study demonstrates that GNNs have state-of-art performance in brain graph classification and identify key brain subregions, leveraging the functional connectome-inspired structure.