Classification of autism in young children by phase angle clustering in magnetoencephalogram signals

Autism spectrum disorder (ASD) is a complex neu- rodevelopmental condition that appears in early childhood or infancy, causing delays or impairments in social interaction and restricted range of interests of a child. In this work, our goal is to classify autistic children from typically developing children using a machine learning framework. Here, we have used magnetoencephalography (MEG) signals of thirty age and gender matched children from each group. We perform a spectral domain analysis in which the features are extracted from both power and phase of large-scale neural oscillations. In this work, we propose a novel phase angle clustering (PAC) based feature and have compared its performance with commonly used power spectral density (PSD) based feature. It is observed that with Artificial Neural Network (ANN) classifier, PAC yields better classification accuracy (88.20±3.87%) than the PSD feature (82.13±2.11%). To investigate laterality of brain activity, we evaluate the classification performance of each feature type over all channels as well as over individual hemispheres. Using machine learning framework it is found that the discriminating PSD features are mostly from high gamma band i.e. 50-100 Hz frequency oscillations and the PSD features are dominant in right hemisphere. These findings are in line with studies carried before in other framework. However, PAC based feature in our study shows that the whole brain contains important attributes of autism. The discriminating PAC features are mostly from theta band (i.e. 4-8 Hz frequency oscillations) that signifies memory formation and navigation. In this study, it is found that impaired theta oscillations correlate with autistic symptoms. Overall, our findings show the potential of such signal processing and classification based study to aid the clinicians in diagnosis of ASD.