Exploring the Potential of Parent Report for Autism/Developmental Screening at the 18-Month Visit

PURPOSE: To develop optimal algorithms of parent-administered items to improve the detection of autism and other developmental disorders at the 18-month visit. METHODS: at scheduled 18-month pediatric visits via an online system. Ninety-six children with positive screens and 314 matched controls completed additional items, including the POSI; items from the FYI and POEM data banks, and the MacArthur-Bates Communicative Development Inventory (MCDI) with short-form vocabulary. Diagnostic testing was conducted using the ADOS-2 Toddler Module and the Mullen. We evaluated a set of machine learning (ML) models predicting autism and Developmental Delay (DD). Model training used tree-based modeling under the gradient boosting framework with feature selection via the Boruta method with Shapley values and Bayesian hyperparameter optimization and synthetic data based on ~50% of our authentic data (n = 201), resulting in a synthetic training dataset of 25,000 cases and a synthetic validation dataset of 12,500 cases, each with > 90% accuracy. We evaluated the performance of top autism/Delay models using the authentic holdback sample (n = 202). RESULTS: The resulting model includes expressive vocabulary and items representing joint attention. CONCLUSIONS: The model appears to be approximately twice as sensitive to autism as the M-CHAT-R-F and twice as sensitive as the ASQ-3 for DD. The TADAS model shows promise as being the only screen for that age with the generally recommended performance of over .7 for both sensitivity and specificity for both autism and DD.