Identification of informative function connectivity in autism spectrum disorder using panalized support vector machine cluster

Resting-state functional connectivity extracted from magnetic resonance images (MRI) is one of the potential biomarkers for psychiatric disorders such as autism spectrum disorder (ASD) owing to its existence of abnormality between diagnosed patient and typical developed (TD) individuals. However, the functional connectivity is often extracted based on huge number of brain regions which makes it high dimensional. In the case of ASD, identification of the informative functional connectivity that relates to the disorder is a crucial task. Support vector machines (SVM) has been widely used in classifying normal and abnormal brain regions. But the use of conventional SVM is not effective enough in identifying informative brain regions. Thus, we used cluster of penalized SVMs to classify between ASD patients and TD individuals. 144 functional MRIs (fMRIs) retrieved from Autism Brain Imaging Data Exchange (ABIDE) that contains 70 ASD samples and 74 TD samples were used. The performance is evaluated in terms of accuracy, sensitivity, and specificity. Comparison is done with results from single SVM. Our method shows better classification accuracy (+25.06%), sensitivity (+28.29%) and specificity (+22.60%). Additionally, the top three selected informative functional connectivity namely superior frontal gyrus with inferior frontal gyrus, precentral gyrus with middle temporal gyrus, and paracentral lobule with postcentral gyrus have been verified with biological knowledge bases and are associated with ASD.