Computational Measurement of Social Gaze During Naturalistic Conversations in Autism

Standardized, granular measurement of autistic behaviors, such as social gaze during interactions, is needed for a range of clinical applications including diagnosis and detecting clinical change. Computational approaches show promise in automatically measuring social behaviors within natural settings. This study aims to automatically measure social gaze features from videos of dyadic conversations, characterize autism-related differences, and perform individual-level diagnostic classification. 46 autistic Participants and 36 neurotypical Participants, aged 8-29 years, engaged in a brief video-recorded conversation with a research staff member (Partner). A deep learning AI model trained to detect whether each partner was looking at the other achieved 89% cross-validated accuracy. Comparing these automatic gaze measurements, autistic Participants spent less time looking at Partners and engaging in mutual gaze than neurotypical Participants did. They also initiated mutual gaze less frequently and had shorter mutual gaze episodes, but did not differ in mutual gaze counts. An AI-derived social gaze summary score correlated specifically with ADOS-2 Social Affect scores and not Restricted and Repetitive Behavior scores. Cross-validated machine learning using gaze features predicted diagnostic group with 73% accuracy. This study provides a framework for automatically quantifying social gaze behaviors, with potential for enhancing diagnostic precision and tracking therapeutic progress in autism.