CARE-ASD · Clinician-Aligned Reasoning and Explanation Framework for Autism Spectrum Disorder Screening

Autism (ASD) affects 1 in 36 children (2.7%) in the US, with a 700% increase in diagnosis since 1998. Early diagnosis is essential, as interventions during early brain development significantly improve outcomes. However, the current gold-standard assessment, the Autism Diagnostic Observation Schedule (ADOS), is time-consuming, requires expert clinicians, and causes long waits that delay support. Machine learning (ML) can support more scalable and efficient screening, but most existing models rely only on perceptual pattern recognition and act as “black boxes”, generating predictions without explaining how they were reached. This project addresses this limitation by separating perception (ML-based detection of behavioural cues: gestures, attention shifts, interactions) from reasoning (their clinical interpretation within the ADOS framework using clinician-defined rules). This makes the system transparent and clinically meaningful. The main contributions are: (1) the release of a large-scale dataset densely annotated with gestures and ADOS items