A Multi-Site Machine Learning Classification of Autism Spectrum Disorder Using ICA Spatial-Temporal Features from Multi-Site rs-fMRI Data

Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder characterized by social interaction difficulties and repetitive behaviors. This study investigated functional brain networks associated with ASD using resting-state functional magnetic resonance imaging (fMRI) data from the Autism Brain Imaging Data Exchange (ABIDE) repository. A large dataset of 1750 subjects were employed for individual and group-level analyses. Group Independent Component Analysis (GICA) was used to extract independent brain networks. After noise removal, a greedy matching strategy identified the most relevant networks for ASD. We parcellated the brain into seven networks and subsequently focused on the five most prominent ones: Visual Salience Network (VSN), Default Mode Network (DMN), Sensorimotor Network (SMN), Dorsal Attention Network (DAN), and a Control Network. Classification accuracies were calculated using three features: spatial, temporal, and combined (spatial + temporal). The combined features yielded the highest accuracy, with VSN exhibiting the strongest performance (87.73%) compared to the other networks (DMN: 84.03%, SMN: 83.98%, DAN: 76.07%, Control: 75.28%). This study demonstrates that the VSN plays a crucial role in ASD pathophysiology. Furthermore, our network-based analysis using the largest combined dataset to date achieves higher classification accuracy compared to previous studies.