Machine learning approach to identify resting-state functional connectivity pattern serving as an endophenotype of autism spectrum disorder

Abstract

Endophenotype refers to a measurable and heritable component between genetics and diagnosis and exists in both individuals with a diagnosis and their unaffected siblings. We aimed to identify a pattern of endophenotype consisted of multiple connections. We enrolled adult male individuals with autism spectrum disorder (ASD) endophenotype (i.e., individuals with ASD and their unaffected siblings) and individuals without ASD endophenotype (i.e., pairs of typical development (TD) siblings) and utilized a machine learning approach to classify people with and without endophenotypes, based on resting-state functional connections (FCs). A sparse logistic regression successfully classified people as to the endophenotype (area under the curve=0.78, classification accuracy=75%), suggesting the existence of endophenotype pattern. A binomial test identified that nine FCs were consistently selected as inputs for the classifier. The least absolute shrinkage and selection operator with these nine FCs predicted severity of communication impairment among individuals with ASD (r=0.68, p=0.021). In addition, two of the nine FCs were statistically significantly correlated with the severity of communication impairment (r=0.81, p=0.0026 and r=-0.60, p=0.049). The current findings suggest that an ASD endophenotype pattern exists in FCs with a multivariate manner and is associated with clinical ASD phenotype.