Ensemble Modeling of Multiple Physical Indicators to Dynamically Phenotype Autism Spectrum Disorder

Area of Science:

• Neuroscience
• Developmental Psychology
• Computer Science

Background:

• Autism Spectrum Disorder (ASD) is a neurodevelopmental condition impacting social communication.
• Early detection of ASD is vital for effective intervention.
• Mobile technology offers potential for scalable, naturalistic data collection for diagnostics.

Purpose of the Study:

• To evaluate the feasibility of using mobile-captured home videos for early Autism Spectrum Disorder (ASD) detection.
• To develop and validate computational models for analyzing behavioral signals from videos.
• To assess the predictive performance of different behavioral modalities and their fusion.

Main Methods:

• A dataset of 688 feature-rich videos was curated from the GuessWhat mobile game.
• A two-step pipeline involving video filtering and feature engineering was implemented.
• Unimodal Long Short-Term Memory (LSTM) models were trained on eye gaze, head position, and facial expressions, followed by late-stage fusion.

Main Results:

• Unimodal models achieved AUCs of 86% (eye gaze), 78% (head position), and 67% (facial expression).
• Late-stage fusion of unimodal outputs significantly improved performance, reaching a test AUC of 90% (95% CI: 0.84-0.95).
• Distinct behavioral channels provided complementary value in ASD detection.

Conclusions:

• Mobile-captured videos can be effectively used to detect clinically relevant behavioral signals for Autism Spectrum Disorder (ASD).
• AI-driven analysis of behavioral patterns from naturalistic videos shows significant promise for scalable and early autism phenotyping.
• Further research is needed to enhance generalizability and inclusivity for real-world application.