B3D: A Decision-Analytic Machine Learning Study for Autism Screening in Toddlers Using Python

This session presents a decision-analytic machine learning study of autism screening in toddlers using Q-CHAT data. Rather than treating screening as a pure classification problem, the study evaluates how interpretable models can support referral decisions under uncertainty. It compares item-level questionnaire responses with total-score approaches and examines discrimination, calibration, and decision-curve utility. Findings suggest that preserving item-level response patterns may improve referral quality while maintaining transparency, offering a practical framework for more clinically meaningful autism screening research. In simplified terms, the study challenges existing screening procedures and identifies an approach that successfully refers 4 out of every 100 more children for autism screening by looking at the value of individual responses to questions vs current cumulative scoring methodologies.

Learning Outcomes:

Participants will be able to explain how item-level screening responses can provide additional predictive value beyond a total screening score in autism referral modeling.

Participants will be able to interpret discrimination, calibration, and decision-curve analysis as complementary measures of whether a screening model is clinically useful.

Participants will be able to identify key limitations in applying machine learning to autism screening datasets, including prevalence enrichment, transportability, and measurement bias.

Speakers