Development and validation of a machine learning-based tool to predict autism among children.
Hospital birth data can flag autism risk early, but newer models double the accuracy with the same inputs.
01Research in Context
What this study did
Betts et al. (2023) built a computer model that spots autism risk in babies. They fed the model routine hospital data from birth and the first weeks of life. The goal was a cheap, early warning system that any clinic could run.
What they found
The model reached an AUC of 0.73. That means it caught most later autism cases, but also flagged some babies who turned out fine. It is good enough to be a first-pass screen, not a final diagnosis.
How this fits with other research
O'Brien et al. (2026) used the same idea—birth records plus machine learning—but hit 0.997 AUC. Their secret was adding mother’s weight and family income, and building separate boy and girl models. The jump from 0.73 to 0.997 shows the target paper’s method has already been outpaced.
Gur et al. (2024) swapped hospital data for regular check-up notes. Their sex-tuned model caught about two-thirds of autism cases. This match tells us the exact data source matters less than tuning the model to boys versus girls.
Bone et al. (2015) warned that flashy ML autism tools often fail when tested on new kids. Steven et al. answered that call by using a fresh, unseen dataset, yet the later 2026 paper still beats it. The field is moving fast.
Why it matters
You now have a clear upgrade path. If you want an early-warning screen, ask your IT team for perinatal data plus mother BMI and zip-code income, then run sex-split models. The 2026 recipe is open, beats the 2023 tool, and still uses data you already store. Start there, not here.
Want CEUs on This Topic?
The ABA Clubhouse has 60+ free CEUs — live every Wednesday. Ethics, supervision & clinical topics.
Join Free →Email your data team: can they pull maternal BMI and census income with the birth file? If yes, skip Steven et al. and build the 2026 model instead.
02At a glance
03Original abstract
Autism is a lifelong condition for which intervention must occur as early as possible to improve social functioning. Thus, there is great interest in improving our ability to diagnose autism as early as possible. We take a novel approach to this challenge by combining machine learning with maternal and infant health administrative data to construct a prediction model capable of predicting autism disorder (defined as ICD10 84.0) in the general population. The sample included all mother-offspring pairs from the Australian state of New South Wales (NSW) between January 2003 and December 2005 (n = 262,650 offspring), linked across three health administrative data sets including the NSW perinatal data collection (PDC); the NSW admitted patient data collection (APDC) and the NSW mental health ambulatory data collection (MHADC). Our most successful model was able to predict autism disorder with an area under the receiver operating curve of 0.73, with the strongest risk factors for diagnoses found to include offspring gender, maternal age at birth, delivery analgesia, maternal prenatal tobacco disorders, and low 5-min APGAR score. Our findings indicate that the combination of machine learning and routinely collected admin data, with further refinement and increased accuracy than achieved by us, may play a role in the early detection of autism disorders.
Autism research : official journal of the International Society for Autism Research, 2023 · doi:10.1002/aur.2912