Elastic SCAD SVM Cluster for The Selection of Significant Functional Connectivity in Autism Spectrum Disorder Classification

Abstract

In the study of functional connectivity for autism spectrum disorder (ASD), the correlation is calculated from the magnetic resonance imaging data for many different pairs of brain regions. Due to the huge number of brain regions exists, the correlation matrix that served as the input for the classifier in machine learning is of high dimensionality. The fact that the correlation is calculated based on all brain regions shows that the correlation matrix might contains irrelevant functional connectivity for the study of ASD. To solve these problems, a framework based on penalized support vector machine (SVM) cluster is proposed, which will select the significant functional connectivity from the original functional connectivity, to be used as the input for several penalized SVM in the cluster, each of the penalized SVMs generated a set of significant feature IDs. A significant functional connectivity matrix is generated to be used as the input features for the final SVM. By comparing to the existing methods that used single SVM, the results show that the proposed method has greatly improved the classification performance in terms accuracy, specificity and sensitivity. Additionally, the selected features are proposed as the regions of interest in brain to study ASD. Through biological validation of these regions, it shows that there might be linking between motor and social and communicative abilities in ASD, in which this suggestion is also supported by other studies related to ASD.