Predicting autism traits from baby wellness records: A machine learning approach.

The team trained computer models to spot autism from regular baby-checkup files.

They used growth charts, vaccine dates, and nurse notes for 0-24-month visits.

Separate boy and girl models were built and tested on 1,400 later-diagnosed kids plus matched controls.

The best model caught 63-a large share of children who later received an autism diagnosis.

It also kept false alarms low, so most flagged babies truly needed follow-up.

Girls were slightly harder to catch, so the girl-only model helped close that gap.

Marsack-Topolewski et al. (2025) now report 97-a large share accuracy with a four-model ensemble, far above the 63-a large share seen here.

That newer tool covers toddlers to adults, so it replaces this baby-only approach for sheer hit rate.

Yet this paper stays useful: it proves the cheap data you already collect at well-baby visits hold signal, while N et al. need extra parent questions.

Smit et al. (2019) tried parent temperament surveys and could not predict individual risk; the present ML method turns the same age window into real, if modest, prediction.

You do not need new tests—just mine the charts you already file.

Flagging two out of three future cases before 24 months gives you a head start on referral, early intervention, and parent coaching.

Try running the free code on your clinic’s wellness database; even a large share sensitivity beats waiting until red flags pile up.