Predicting Autism Spectrum Disorder Based on Resting-State fMRI and Graph Theory

Machine learning methods have been extensively employed in the diagnosis of autism spectrum disorder to comprehend the neurophysiological underpinnings. Meanwhile, resting-state functional connectivity has been considered a potentially significant and promising tool that furnishes biomarkers to diagnose diseases. In this study, we first extract static and dynamic functional connectivity patterns from functional magnetic resonance imaging data obtained from the Autism Brain Imaging Data Exchange I. Subsequently, graph theory is utilized to individually determine the optimal features from these groups of functional connectivity for autism spectrum disorder and typical control classification. The groups of selected features are then separately trained on support vector machines, resulting in several classification models. Finally, these models vote on whether a test subject has autism spectrum disorder. Our proposed method achieves a better autism spectrum disorder identification performance with an accuracy of 82.87% and provides measures of certain regions of interest that can serve as alternative reference biomarkers for classification.