Developing an integrated algorithm to support autism diagnostic decisions: Model performance and statistical fairness

Editorial reading guide · Respect

Has an autism assessment model been checked for children like ours?

Overall validation performance was high but fairness criteria varied, and false positives were more common in advantaged groups. The authors warn that a model can integrate useful information while still risking unequal classification.

Study context

The study included 639 Illinois early-intervention toddlers assessed through telehealth, with caregiver and provider ratings and an expert diagnostic reference.

Age / stage: Toddlers aged 15–36 months; mean 30.4 months

Read with three questions

1. Which fairness measures differed despite high average accuracy?
2. Why might group differences in false positives suggest underclassification elsewhere?
3. How does an early-intervention sample differ from general population screening?

Keep the limits in view

The sample already received early intervention, had high autism prevalence, and excluded major medical or genetic complications. Internal split validation and subgroup results do not establish fairness or predictive values in other populations.