Peer-reviewed article

From clinical and typical separation to autism spectrum disorder and global developmental delay differentiation: Interpretable eye-tracking-based machine learning.

Gang Zhou, Xiaobin Zhang, Xingda Qu, Xinyao Hu, Qiongling Peng et al. · 28 Aug 2026

Abstract

In clinical practice, objectively distinguishing children with autism spectrum disorder (ASD) from those with typical development (TD) and other neurodevelopmental conditions with overlapping symptoms, such as global developmental delay (GDD), is critical for improving developmental outcomes. This study aimed to investigate the feasibility of eye-tracking-based machine learning (ML) models in distinguishing children with developmental concerns (ASD + GDD) from TD children, and in further differentiating ASD from GDD. A total of 168 toddlers (45 ASD, 56 GDD, and 67 TD; aged 18-48 months) viewed a socially dynamic "hide-and-seek" video. Both conventional area-of-interest (AOI)-based features and combined features derived from AOI combinations were used to train five ML classifiers. A two-stage classification framework was adopted, and SHapley Additive exPlanations (SHAP) analysis was used to interpret feature contributions. Combined features consistently outperformed isolated AOI features. For TD versus clinical classification, the random forest model achieved the highest accuracy of 80.36% (sensitivity = 84.16%, specificity = 74.63%). In differentiating ASD from GDD, the k-Nearest Neighbors model achieved the highest accuracy of 79.21% (sensitivity = 71.11%, specificity = 85.71%). SHAP analysis indicated that cross-regional attention features contributed substantially to model performance. These findings suggest that eye-tracking-based ML models may provide a promising approach for early screening of developmental conditions and ASD-GDD differentiation. Combined features reflecting cross-regional attention distribution demonstrated higher discriminative value than conventional AOI measures. Future studies should validate these findings in larger, multicenter, and more balanced samples with richer dynamic eye-tracking representations.