Applying machine learning to facilitate autism diagnostics: pitfalls and promises.

Bone et al. (2015) tried to repeat earlier claims that a computer program could spot autism in minutes.

They used bigger, balanced data sets instead of the small samples from the first reports.

The team wrote a methods paper that walks readers through why the flashy results fell apart.

The earlier machine-learning tools could not copy their own success.

When tested on new, larger groups the accuracy dropped and the tools failed.

The authors warn that quick-screen algorithms still need expert eyes to back them up.

Marsack-Topolewski et al. (2025) now says a four-model ensemble reaches 97-99 % accuracy.

This looks like a direct update to Daniel’s 2015 failure, but the newer study used tighter data cleaning and age-stratified sets.

Sahai (2025) and Kremkow et al. (2022) both agree the field is moving fast, yet they repeat Daniel’s warning: most apps are still proof-of-concept, not clinic-ready.

So the 2015 flop still matters—it keeps newer teams honest about replication before marketing.

You will hear sales pitches for AI screeners. Daniel et al. give you the questions to ask: Was the model retested on new, balanced data? Are the kids like the ones on your caseload? Until vendors show that level of proof, keep your ADOS and clinical judgment in the driver’s seat.