Development and Validation of Prediction Models for the Diagnosis of Autism Spectrum Disorder in a Korean General Population.

Objective: Delays in autism spectrum disorder (ASD) diagnosis and treatment are significant clinical problems that can be addressed by timely, community-based assessment. This study examined tools for identifying ASD in community settings using machine learning (ML) models. Method: This study analyzed population-based cross-sectional studies (2005-2017) of ASD in South Korea. A community sample of 62,083 children was screened using the Autism Spectrum Screening Questionnaire (ASSQ) and teacher/caregiver referrals. Caregivers completed the Behavior Assessment System for Children-2nd Edition (BASC-2) and the Social Responsiveness Scale (SRS). Screen positives were offered a comprehensive clinical evaluation. Among the first-graders in regular elementary schools who completed the diagnostic evaluation (N = 746), supervised ML models (generalized linear model with elastic net regularization [GLMNET], classification and regression tree, random forest, and gradient boosting [GB]) were developed and validated for classification of ASD. Models were developed in the single questionnaire and combined questionnaire datasets, using questionnaire responses and demographic and developmental information. Results: ASD was diagnosed in 46.2% of children (median age, 6.8 years [interquartile range, 6.5-7.1 years]; 71.7% boys). Among single questionnaire models, the BASC GB model demonstrated the best discrimination ability (area under the curve 0.80, 95% CI 0.75-0.83). Area under the curve of the GLMNET model with combined ASSQ, BASC-2, and SRS was the highest, 0.82 (95% CI 0.77-0.89); the predicted risk of ASD by the GB model of combined questionnaires agreed the best with the observed risk of ASD compared with other ML models. Conclusion: Caregiver questionnaire ML models showed future promise for identifying children with ASD in community settings.