A Machine Learning Strategy for Autism Screening in Toddlers.
Luke E K Achenie, Angela Scarpa, Reina S Factor and 3 others
PMID 30985384WHAT IT FOUND
In data from 14,995 toddlers, a computer program classified 99.72% of test cases correctly.
It was less likely than the paper screener to identify true cases. It is not yet ready to replace clinician screening.
Key findings
01In data from 14,995 toddlers, the best model used 18 M-CHAT-R items and classified 99.72% of test cases correctly.
02The model was less likely to identify true cases (sensitivity 0.738 vs 0.854) but more likely to correctly clear non-cases (specificity 0.999 vs 0.993).
03The model had a higher positive predictive value than the paper version (0.789 vs 0.475), but the authors note a high false-negative rate and call for testing in pediatric practices.
STILL TO COME
How it was doneWhat they found
Read the rest of this summary
You get three full summaries a month, free, and we do not ask for a card. Search, the TL;DRs and your library stay unlimited either way.
What it does not show
This is a model analysis of archival screening data, not a test of the method in a clinic or a study of whether it improved diagnosis or intervention. The model used only the initial M-CHAT-R items and did not include the follow-up interview questions. The authors note a high false-negative rate, meaning the model may miss true cases. American Indian, Hispanic, and Bi/Multiracial participants were excluded from subgroup analyses because there were not enough participants. Participants were mostly from urban and suburban areas, so rural populations were not well represented. The comparison with the paper M-CHAT-R/F used the same data and different metrics, and the machine learning model is a black box. The authors acknowledge that race is a complex social construct and that sociodemographic factors such as socioeconomic status or education were not fully disentangled.
Declared interests
Funding is listed as NIH extramural and non-U.S. government support. Diana L. Robins is co-owner of M-CHAT, LLC, which licenses commercial use of the M-CHAT-R/F. The other authors declared no conflicts.
The easy way to misread this
Do not conclude that automated scoring can replace the M-CHAT-R/F follow-up interview. The model's sensitivity, its ability to identify true cases, was 0.738, lower than the paper screener's 0.854. It also was not tested in real clinical use.