Social Whole-Body Movement Aids the Objective Classification of Children With Autism and Typical Development by Implementing a Machine Learning Approach

Autism spectrum disorder (ASD) is characterized by high prevalence and a time-consuming diagnostic procedure that greatly depends on clinicians’ experience. An objective and efficient approach to identifying ASD is urgently called upon. In current clinical practice, although the diagnosis of ASD is primarily based on the individual’s behavioral presentation, few studies have examined the feasibility of using behavioral indices to identify ASD, and none has investigated whether behavioral features extracted from social whole-body motor behavior during a face-to-face interaction could aid an accurate detection of ASD. This study aimed to address this question by applying machine learning (ML) algorithms to classify children with ASD and typical development (TD). Twenty children with ASD and twenty-one children with TD were videotaped in a face-to-face conversation. A motion energy analysis technique extracted the whole-body movement time series. Features computed were the power of eighty-six frequencies obtained by fast-Fourier transform. Support vector machine and leave-one-out cross validation were implemented. Results showed that a maximum classification accuracy of 97.56% (specificity = 100%, sensitivity = 95.00% ) could be achieved with three features. The contactless and calibration-free approach proposed in this study may help facilitate an objective, effective and efficient diagnosis of ASD.