A functional system-informed graph neural network framework to quantify interpretable brain dysfunction in ASD

Autism Spectrum Disorder (ASD) is one of the most common neurodevelopmental disorders affecting patients from childhood to adulthood. Yet, its pathological mechanism has not been conclusively established, and its diagnosis mainly depends on subjective assessments of clinical or behavioral measures. In the human brain, connectivity is ubiquitous between brain regions that share functional properties, and groups of brain regions spontaneously assemble into large-scale systems as per their functions. In this study, we propose functional system-informed graph neural network (FS-GNN), a framework using functional magnetic resonance images (fMRI) to diagnose ASD and unveil new perceptions of ASD-related dysfunctions by fully leveraging the topological and functional information underlying the brain connectome. Specifically, we introduce a learnable positional encoding approach for brain regions of interest (ROIs) concerning their natural locations and functional interactions. The large-scale brain systems are integrated as prior knowledge into the graph representation learning to aid the model in identifying clusters of functionality. A graph readout with system-driven regularization is also developed to automatically weigh the ROIs in respect of their contribution to the classification. Experimental results on a multi-site database known as Autism Brain Imaging Data Exchange (ABIDE) validate the efficacy of FS-GNN by outperforming prevalent machine learning and GNN approaches, reaching 75.02 % accuracy, 73.22 % precision, and 71.64 % recall in ASD diagnosis. The brain dysfunctions detected by our model from both ROI and system levels achieve high agreement with previous fMRI-derived evidence of ASD biomarkers. The results demonstrate the strength of our proposed FS-GNN in discovering interpretable and trustworthy neural patterns for a more precise diagnosis of ASD.