Classification of Autism Disorder Using Functional Connectivity Networks Obtained Through Sparse Inverse Covariance Estimation

Abstract

In recent years there have been a growing number of studies investigating the efficacy of using brain connectivity networks in autism diagnosis. Many of these studies have used correlation analysis to estimate functional connectivity structure; however, this approach is problematic because it ignores the confounding effects of other brain regions. A promising area of research is using partial correlations to infer the connectivity structure. There are several methods that have recently been proposed by the machine learning community for estimating partial correlations in high-dimensional settings, such as Sparse Inverse Covariance Estimation (SICE) using the Graphical Lasso algorithm. In this study, we apply SICE in developing a connectivity-based classifier for autism disorder, and we evaluate its predictive accuracy on a set of 73 adolescents from the Autism Brain Imaging Data Exchange (ABIDE) dataset. Using SVM for our classifier, we obtained a classification error rate of 26%, with a false positive rate of 31% and a false negative rate of 22%.