Abstract
<p>Restricted and repetitive behaviors (RRBs) are a common class of behavior in autistic children, and their assessment is typically based on presence and severity. While previous research has primarily focused on detecting RRBs (i.e., presence or type), automated classification of severity levels has received limited attention. The present study aimed to categorize RRB severity in 58 autistic children using clinician-labeled data based on the Autism Diagnostic Observation Schedule (ADOS). A machine learning model was evaluated under two strategies: session-based and participant-based analyses. The model achieved high accuracy for session-based training (Cohen’s kappa = 0.82) whereas accuracy decreased for participant-based training (Cohen’s kappa = 0.54), indicating reduced generalization to new children. Confusion matrix analysis showed clearer class separation in the session-based analysis and greater overlap among ratings in the participant-based analysis. These findings suggest that the automated rating of RRBs is feasible, but that inter-individual variability remains a challenge for participant-independent applications.</p>