A lightweight, end-to-end explainable, and generalized attention-based graph neural network model trained on high-order spatiotemporal organization of dynamic functional connectivity to classify autistics from typically developing

Autism spectrum disorder (ASD) is a neurodevelopmental disorder characterized by deficits in social cognition, interaction, communication, restricted behaviors, and sensory abnormalities. The heterogeneity in ASD's clinical presentation complicates its diagnosis and treatment. Recent technological advancements in graph neural networks (GNNs) have been extensively used to diagnose brain disorders such as ASD, but existing machine learning models often suffer from low accuracy and explainability. In this study, we proposed a novel, explainable, and generalized node-edge connectivity-based graph attention neural network (Ex-NEGAT) model, leveraging edge-centric high-order spatiotemporal organization of dynamic functional connectivity streams between large-scale functional brain networks implicated in autism. Using the Autism Brain Imaging Data Exchange I and II datasets (total samples = 1,500), the model achieved 88% accuracy and an F1-score of 0.89. Additionally, we used meta-connectivity subtypes to identify subgroups within ASD samples using the rough fuzzy c-means algorithm. We also used connectome-based prediction modeling, which revealed critical brain networks contributing to predictions that accurately correlate with Autism Diagnostic Observation Schedule (ADOS) and full intelligent quotient (FIQ) scores. The proposed framework offers a robust approach based on previously unexplored higher order spatiotemporal correlation features of dynamic functional connectivity, which may provide critical insight into ASD heterogeneity and improve diagnostic precision.