Abstract
Objectives: Formal autism diagnosis typically follows initial concern by more than two years, and current screening tools rely on subjective caregiver reports. Motor differences in walking gait are well documented but not used in standard screening. This pilot study examined whether gait features from 2D walking videos could distinguish autistic from non-autistic children.
Methods: Using OpenPose, a 2D pose estimation system, we analyzed front- and side-view walking videos of 72 children aged 3–12 years (30 autistic, 42 non-autistic). From joint coordinates, 163 biomechanical gait features were computed and reduced to eight via random-forest-based selection. Three classifiers were compared: logistic regression, random forest, and support vector machine (SVM).
Results: On a stratified held-out test set, the SVM achieved the strongest performance (AUC = 0.935; accuracy = 0.810; sensitivity = 0.889; specificity = 0.750), a profile consistent with screening use. Random forest reached moderate performance but paired high specificity with low sensitivity, while logistic regression performed least effectively, suggesting linear models capture this signal poorly.
Conclusions: These pilot findings support the feasibility of 2D video-based gait analysis as an accessible, objective complement to existing autism screening, though validation in larger, more diverse samples is needed before any clinical application.
Keywords: Autism; gait analysis; screening; 2D pose estimation; motor development; machine learning