Autism screening for children based on appearance features across multiple paradigms

Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder characterized by deficits in social communication and behavior . Current ASD screening methods often focus on single paradigms or isolated features, limiting their screening accuracy. In this work, we propose a novel method for ASD screening by integrating appearance features such as gaze points, facial expressions, and head pose across multiple paradigms. The multiple paradigms including blank-overlap, person-gaze and exogenous-cueing are designed to elicit atypical behavioral patterns in ASD children. The extracted features from video data are then classified using a Long Short-Term Memory (LSTM) model. Our method achieves a classification accuracy of 0.932 and a sensitivity of 0.947 in differentiating ASD from typically developing (TD) children. The code and dataset related to this paper are available at https://github.com/theolsy/ASDTD . • We propose a multimodal ASD screening method combining gaze points, facial expressions, and head pose features. • Our study highlights the effectiveness of task-specific paradigms in eliciting distinct ASD behavioral patterns. • The proposed fusion of appearance features significantly improves classification performance. • We effectively utilize temporal–spatial dynamics for robust ASD classification. • Our method achieves high screening performance, with 0.932 accuracy and 0.947 sensitivity.