Title: Leveraging 3D Temporal Analysis for Autism Spectrum Disorder Screening in Attention-Based Tasks

Abstract

Timely intervention is essential for supporting the social, cognitive, and academic growth of school-age children, making accurate Autism Spectrum Disorder (ASD) screening a critical priority. This process is particularly important for identifying cases that may have been overlooked in earlier years. However, existing screening methods often depend on subjective evaluations and two-dimensional (2D) analysis techniques, which are insufficient for capturing the specific spatial displacement patterns associated with ASD behaviors.

To address these limitations, this study introduces a novel framework for 3D temporal analysis, leveraging DECA (Detailed Expression Capture and Animation), a 3D modeling tool. This approach allows for the extraction of comprehensive head pose parameters—specifically translational components ($T_x, T_y, T_z$)—and facial expressions that remain consistent regardless of pose variations. The research utilized video data from a cohort of 39 participants, comprising 19 children with ASD and 20 typically developing (TD) children, aged between 7 and 12 years. These participants underwent Virtual Reality-Continuous Performance Test tasks while their data was recorded.

Temporal classifiers based on Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) architectures were trained using the extracted 3D features. The results indicated that GRU-based models yielded superior performance compared to other methods. Specifically, 3D head pose features achieved an accuracy of 83.9%, while 3D facial features reached 81.4%. These figures represent significant improvements over 2D baseline approaches, surpassing them by 10.7% and 7.5%, respectively.

Additionally, the study explored multimodal fusion by combining 3D head pose and facial features, utilizing Principal Component Analysis (PCA) for dimensionality reduction. This combined approach resulted in the highest observed accuracy of 84.6%, outperforming models that relied on single modalities. By providing a more objective and automated method for identification, this research lays the groundwork for addressing current diagnostic gaps in ASD screening for school-age populations.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC