Early identification of autism spectrum disorder by multi-instrument fusion: A clinically applicable machine learning approach

Abstract

Autism spectrum disorder (ASD), developmental language disorder (DLD), and global developmental delay (GDD) are common neurodevelopmental disorders in early childhood; however, the differential diagnosis of these disorders is difficult because of overlapping symptoms. Drawing on a cohort of 2004 children with ASD, DLD, or GDD, this study developed machine learning classifiers using decision trees, support vector machines, eXtreme gradient boosting (XGB), logistic regression, and neural networks by combining several easily accessible behavioral and developmental assessment instruments. The best-performing XGB model was further simplified into a two-stage decision model (TS-DM) to achieve better interpretability. Model performance was tested and compared with that of 12 pediatricians on an external dataset of 60 children. The accuracies of the resident pediatricians, senior pediatricians, TS-DM, and XGB were 53.3%, 66.7%, 75.0%, and 78.3%, respectively. Machine learning has the potential to identify these three neurodevelopmental disorders by integrating information from multiple instruments and thereby may increase our understanding of the roles of different behavioral and developmental characteristics in the different diagnoses.