A Systematic Survey of Vision Transformer-Based Autism Detection Using Full-Body Movement Data from Augmented Reality Gameplay

Abstract

Early detection of autism spectrum disorder (ASD) increasingly relies on objective movement-based biomarkers. This Systematic survey explores the confluence of three promising areas: Augmented Reality (AR) gameplay for eliciting naturalistic full-body movement, skeleton tracking for data extraction, and Vision Transformer (ViT) models for analysing spatiotemporal movement patterns to detect ASD. This survey aims to summarize the current research landscape by considering available full-body skeleton datasets and transformer-based approaches, such as pure ViTs, spatiotemporal transformers, graph transformer variants, hybrid CNN-ViT models, and physics-informed approaches for AR systems that elicit ASD-relevant motor and social behaviors, and to organize them into a taxonomy. Our results reveal that models from the transformer learning family are better than traditional deep learning methods at all aspects of modelling coordination, time irregularity, and joint-dependency patterns related to ASD; however, this is limited due to a lack of AR gameplay data sets, pose-estimation accuracy, heterogeneity in task design, and a lack of clinically validated benchmarks. Building upon these results, we present a reference end-to-end pipeline for AR-collected movement analysis taking into account standardized task design, robust skeleton preprocessing, motion tokenization, and classifier through transformers and targeted suggestions for future research such as the development of shared AR-based datasets, multimodal fusion approaches (e.g., gaze + skeleton), self-supervised vital transformer pretraining with large motion datasets, physics-informed modeling, and clinically aligned evaluation protocols to ensure the development of scalable, robust and clinically meaningful vit-based autism detection systems.