AI-Driven Detection of Autism Spectrum Disorder Utilizing Home Video Analysis
Overview
This study developed an AI-based screening system using short home-recorded videos to detect autism spectrum disorder (ASD) in young children. The system analyzed videos from 510 children, achieving an area under the ROC curve of 0.83 and 75% accuracy, demonstrating potential to complement clinical evaluations and enable earlier ASD identification.
Background
Autism spectrum disorder (ASD) is a neurodevelopmental condition characterized by social communication differences and repetitive behaviors, affecting approximately 0.6% of the global population. Early diagnosis is critical for effective intervention but is often delayed due to resource-intensive traditional assessments and limited access to specialists, especially in low- and middle-income countries. Conventional diagnostic tools require trained professionals and are time-consuming, while caregiver-report screening tools have limited accuracy. Home videos provide a naturalistic context for observing behaviors but manual coding is laborious and variable. AI and machine learning offer scalable, objective alternatives for automated home video analysis to improve early ASD detection.
Data Highlights
Metric
Value
Sample Size
510 children (253 ASD, 257 typically developing)
Age Range
18–48 months
Number of Hospitals
9 (South Korea)
Video Protocols
3 tasks (name-response, imitation, ball-playing), each under 1 min
Area Under ROC Curve (AUC)
0.83
Accuracy
0.75
Key Findings
The AI system used three brief home video tasks to elicit natural child behaviors relevant to ASD screening.
Deep learning models extracted task-specific features from videos, combined with demographic data via machine learning classifiers.
The ensemble classifier achieved an AUC of 0.83, indicating good discrimination between ASD and typically developing children.
Overall classification accuracy was 75%, supporting the system's potential as a screening tool.
This fully automated approach reduces reliance on specialized clinical resources and observer bias.
Home video analysis captures behaviors in naturalistic settings, enhancing ecological validity compared to clinical assessments.
Clinical Implications
The AI-driven home video screening tool offers a scalable, accessible method to support early ASD identification, particularly in settings with limited specialist availability. By enabling earlier detection, it may facilitate timely referral and intervention, improving developmental outcomes. This approach complements traditional clinical evaluations and could reduce diagnostic delays inherent in current practices.
Conclusion
Automated analysis of short home-recorded videos using AI demonstrates promising accuracy for early ASD detection, providing a practical adjunct to conventional diagnostic methods. This technology has the potential to enhance early identification and intervention efforts globally.
References
Global Burden of Disease Study 2021 -- ASD Prevalence and Impact
Centers for Disease Control and Prevention (CDC) -- ASD Diagnosis Age Data
