A Fusion-Based Machine Learning Approach for Autism Detection in Young Children Using Magnetoencephalography Signals

Abstract

In this study, we aimed to find biomarkers of autism in young children. We recorded magnetoencephalography (MEG) in thirty children (4–7 years) with autism and thirty age, gender-matched controls while they were watching cartoons. We focused on characterizing neural oscillations by amplitude (power spectral density, PSD) and phase (preferred phase angle, PPA). Machine learning based classifier showed a higher classification accuracy (88%) for PPA features than PSD features (82%). Further, by a novel fusion method combining PSD and PPA features, we achieved an average classification accuracy of 94% and 98% for feature-level and score-level fusion, respectively. These findings reveal discriminatory patterns of neural oscillations of autism in young children and provide novel insight into autism pathophysiology.