Autism Detection in Children using Common Spatial Patterns of MEG Signals

Autism exhibits a wide range of developmental disabilities and is associated with aberrant anatomical and functional neural patterns. To detect autism in young children (4-7 years) in an automatic and non-invasive fashion, we have recorded magnetoencephalogram (MEG) signals from 30 autistic and 30 age-matched typically developing (TD) children. We have used a machine learning classification framework with common spatial pattern (CSP)-based logarithmic band power (LBP) features. When comparing the LBP feature to the conventional logarithmic variance (LV) spatial pattern, CSP + LBP (92.77%) has performed better than CSP + LV (90.66%) in the 1-100 Hz frequency range for distinguishing autistic children from TD children. In frequency band-wise analysis using our proposed method, the high gamma frequency band (50-100 Hz) has shown the highest classification accuracy (97.14%). Our findings reveal that the occipital lobe exhibits the most distinct spatial pattern in autistic children over the whole frequency range. This study shows that spatial brain activation patterns can be utilized as potential biomarkers of autism in young children. The improved performance signifies the clinical relevance of the work for autism detection using MEG signals.

I. Introduction

Autism spectrum disorder (ASD) is a complicated neurodevelopmental condition that impairs the brain’s capacity to process information. Inabilities in social interaction and communication, which typically manifest as repetitive habits and restricted interests, are generally identified as ASD. Its estimated prevalence rate among children varies from 0.23% in India, 1.7% in the UK to 2.5% in the USA [1]. Compared to the conventional diagnosis tools [2], [3], brain imaging-based approaches are being explored as potential diagnostic technologies [4] to detect autism automatically and non-invasively, as well as to discover potential biomarkers underlying the ailment. For example, the spatial pattern of brain activity recorded by fMRI is shown to be a biomarker of ASD symptoms [5]. However, fMRI is difficult to record in young children [6]. Furthermore, its temporal resolution is poor (in the order of seconds), and it is an indirect measure of brain activity. On the other hand, electroencephalogram (EEG) and magnetoencephalogram (MEG) are direct indicators of neuronal activity, are noninvasive, and offer excellent temporal resolution (in the order of milliseconds). Therefore, they are more suitable to work with young children. In this study, we select MEG over EEG to identify early neural markers of autism in young children because (i) it is reference-free, (ii) has a better spatial resolution, and (iii) is more suitable for young children (e.g., less prep. time).