Camera-based Gait Kinematic Features Analysis and Recognition of Autism Spectrum Disorder.

Abstract

The atypical development in children with autism spectrum disorder (ASD) may cause varying degrees of gait deficits, characterized by uncoordinated and peculiar postures. However, these symptoms are often ignored due to their subtlety. This study aimed to quantify the atypical gait pattern in ASD and explore the feasibility of a gait-based method for ASD recognition. Firstly, we collected natural walking videos from 38 ASD children and 30 health control (HC) children, then extracted gait kinematic parameters using a skeleton model, including joint swing angle and amplitude features, to analyze subtle changes among ASD children. Subsequently, the potential correlation of these features with the clinical severity of ASD was analyzed, and several machine learning models were constructed for recognition. The results showed, compared to HC group, ASD group had a significant decrease in step length, speed, leg swing angle and coordination, along with a significant increase in head angle. Moreover, significant correlations were observed between these features and both Autism Behavior Checklist (ABC) and Clancy Autism Behavior Scale scores, except for the coordination, which only exhibited significant correlation with ABC score. For recognition, the Random Forests achieved the best recognition performance with an accuracy of 0.84 and an F1 score of 0.86. Overall, this study reveals the atypical gait pattern of ASD children, and proposes a novel gait-based recognition model for future auxiliary evaluation.