Diagnostic prediction of autism spectrum disorder using complex network measures in a machine learning framework

Objective imaging-based biomarker discovery for psychiatric conditions is critical for accurate diagnosis and treatment. Using a machine learning framework, this work investigated the utility of brain's functional network topology (complex network features) extracted from functional magnetic resonance imaging (fMRI) functional connectivity (FC) as viable biomarker of autism spectrum disorder (ASD). To this end, we utilized resting-state fMRI data from the publicly available ABIDE dataset consisting of 432 ASD patients and 556 matched healthy controls. Upon standard pre-processing, 3D + time fMRI data were parcellated into 200 functionally homogenous regions, and whole-brain FC network using Pearson's correlation was obtained from corresponding regional mean time series. A battery of complex network features were computed from the FC network using graph theoretic techniques. Recursive-Cluster-Elimination Support Vector Machine algorithm was employed to compare the predictive performance of three independent feature sets, (i) FC, (ii) complex network measures, and (iii) both combined. The study found that FC could diagnose ASD with 67.3 % accuracy and graph measures with 64.5 % accuracy, while the combined feature set could diagnose with 70.1 % accuracy (all accuracies were significantly different, p < 10−30). The most discriminative imaging features were mainly from lateral temporal, occipital, precuneus (all reduced in ASD) and orbito-frontal (elevated in ASD) regions. We concluded that network topology (graph measures) carried some unique information about ASD pathology not available in bivariate connectivity (FC), and that using both together provided better prediction than using individual measures. Future prediction studies could incorporate multiple fMRI analysis strategies within their framework to achieve superior prediction performances.