From Observation to Early Screening: Automated Extraction of Behavioral Biomarkers for Autism Spectrum Disorder
Al-Nahyan, Shamsah Hamad Bin Tahnoon
Al-Nahyan, Shamsah Hamad Bin Tahnoon
Citations
Altmetric:
Supervisor
Department
Machine Learning
Embargo End Date
Type
Thesis
Date
2026
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
Autism Spectrum Disorder (ASD) has become a critical area of focus in pediatric healthcare, with prevalence rates rising to 1 in 36 children. While early intervention is universally recognized as the primary driver for improved long-term developmental outcomes, the diagnostic pathway remains fraught with delays. Current reliance on subjective behavioral observations and scarce specialist availability creates a significant "diagnostic gap." While computer vision offers a promising avenue for objective screening, existing solutions are largely confined to high-resolution, frontal-facing imagery collected in controlled laboratory settings. This limited ecological validity poses a challenge for developing scalable screening tools that can operate in "in-the-wild" environments.
In response to this challenge, this thesis introduces a novel, end-to-end computer vision pipeline designed to extract behavioral biomarkers directly from low-resolution, top-down surveillance footage. The proposed framework shifts the paradigm from active, structured testing to passive, naturalistic screening. The methodology is built upon a hierarchical classification architecture that processes raw CCTV footage through four distinct stages: frame extraction, person detection, facial alignment, and demographic-filtered emotion analysis. By leveraging YOLOv8 for subject localization (achieving 99.5% precision) and SCRFD for robust face detection, the system overcomes the severe occlusion and perspective distortions inherent in ceiling-mounted cameras.
Central to this work is the development of a domain-adapted Age Group Classification module based on EfficientNet-B0. To address the extreme class imbalance typical of public spaces, the model was fine-tuned using weighted loss functions, achieving a 98.0% accuracy in distinguishing children from adults. This hierarchical filtration ensures that downstream emotional analysis is performed exclusively on the target demographic, significantly reducing computational noise.
We further demonstrate the clinical utility of this pipeline by quantifying emotional phenotypes in a dataset of over 6,700 child face crops. The analysis reveals a distinct behavioral signature characterized by a predominance of Sad (42.2%) and Neutral (28.4%) affect, with a marked reduction in positive expressions (6.0% Happy). These findings align strongly with established clinical literature regarding the "flat affect" and social deficits associated with ASD. By successfully correlating automated metrics with clinical biomarkers, this thesis provides compelling evidence that non-invasive, CCTV-based analysis can serve as a viable, scalable, and objective screening tool to support early detection in naturalistic settings.
Citation
Al-Nahyan, Shamsah Hamad Bin Tahnoon, "From Observation to Early Screening: Automated Extraction of Behavioral Biomarkers for Autism Spectrum Disorder," M.S. Thesis, Machine Learning, MBZUAI, 2026.
Source
Conference
Keywords
Autism Spectrum Disorder, ASD, YOLO, SCRFD, SMOTE, CCTV