Computerized Adaptive Tests for Rapid and Accurate Assessment of Autism

Importance: Extant diagnostic and screening tools struggle to accommodate the diverse features of autism spectrum disorder (ASD) while balancing psychometric properties with respondent burden. Bifactor multidimensional item response theory (MIRT)-based computerized adaptive testing (CAT) can increase accuracy and accessibility of diagnostic assessments. Objective: To develop, calibrate, and validate the CAT-Autism tool for children and adolescents. Design, Setting, and Participants: This diagnostic study used National Institute of Mental Health Data Archive data available February 2024 from neurodevelopmental studies including children and adolescents with item-level responses for measures relevant to domains related to autism phenotype features. Based on reported scores from clinical assessments or measures, individuals were grouped by diagnosis as neurotypical and ASD (all levels of severity). Data were analyzed from February 2024 to February 2026. Main Outcomes and Measures: Valid CAT-based estimates in 10 subdomains (restricted to those in sample completing 200 items or more) compared with full item-bank scores; predictive performance was evaluated in discrimination, calibration, and clinical utility analyses against clinician-assigned diagnostic status. Results: A total of 490 questions to parents and caregivers were considered for potential item bank inclusion. The final sample included 10 309 children and adolescents for calibration (mean [SD] age, 9.3 [6.4] years; 7716 male [74.9%]); 2055 children and adolescents diagnosed as neurotypical (19.9%) and 8254 with ASD (80.1%). Of 490 items, 424 fit the bifactor structure with loadings higher than 0.3 on the primary dimension (ASD). Correlation between observed and estimated item-category proportions was r = 0.98, indicating exceptional fit of the model to the data. With age and sex included as external predictors, the CAT-Autism model yielded outstanding diagnostic predictive accuracy for differentiating neurotypical from ASD results with an AUC of 0.95 (95% CI, 0.92-0.98) for 1-to-5-year-olds and 0.94 (95% CI, 0.92-0.97) for 6-to-18-year-olds. Correlation between full item-bank scores and CAT scores (mean, 13 items; range, 6-45 items) was r = 0.95 for 1-to-5-year-olds and r = 0.94 for 6-to-18-year-olds. Conclusions and Relevance: In this development and validation of our prediction model, CAT-Autism surpassed full item-bank scores with an average of 13 questions; these findings demonstrated that adaptively administering a small, statistically optimal subset of items can yield comparable (or better) results while reducing respondent burden. Next steps will be to test implementation parameters and confirm efficacy of CAT-Autism in clinical and community settings.