GCN based Bio-Inspired Classifier for Autism Spectrum Disorder

Abstract

Autism spectrum disorder is a diverse neurological state with long-lasting and in most instances lifetime implications for individuals. The early identification and intervention are crucial in mitigating the impact of this disorder, necessitating the development of an objective diagnostic method. This study proposes a novel diagnostic approach that utilizes the data extracted from resting state Functional Magnetic Resonance Imaging (rs-fMRI) and critical phenotypic data of each individual. Recursive feature elimination with Grey Wolf Optimization (GWO) is employed for identifying the optimal attributes from the fMRI data. The selected attributes are then inputted into a Graph Convolution Network (GCN) along with the demographic and basic clinical information for categorization purposes. By utilizing a bioinspired optimization algorithm, the likelihood of identifying the optimal feature subset is enhanced. The study compares the performance of the GCN obtained from the GWO feature selection using both the wrapper and filter approaches. The feature set selected through the GWO wrapper approach demonstrates improved accuracy, achieving 73.86%, along with an AUC of 0.817 when inputted into the Graph Convolution Network. These detections emphasise the significance of an objective and accurate ASD diagnosis method with a limited feature set.