Sparsifying machine learning models identify stable subsets of predictive features for behavioral detection of autism.
Sebastien Levy, Marlena Duda, Nick Haber and 1 others
PMID 29270283WHAT IT FOUND
A small set of behavioral items from the standard autism observation test, including unanswered items, separated children with autism from children without autism in these datasets.
This was model testing only, not a clinic-ready triage tool.
Key findings
01For module 3, the reduced-10 feature set included language items, social interaction items, repetitive behavior, gender, and features recording that the examiner could not test a behavior.
02Logistic regression using the reduced-10 feature set achieved AUC ROC 0.95 and balanced accuracy 0.90 on the test set.
03For module 2, the reduced-5 feature set classifiers reached AUC ROC almost 0.88, and the authors said generalization may be limited.
STILL TO COME
How it was doneWhat they foundWhat it means for SLPs
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
The testing set was 20% of the same aggregated data, not a new clinical population. The control groups were small: 70 non-ASD subjects in module 2 and 273 in module 3. Controls were children suspected of autism who did not meet diagnostic cutoffs, not typical community controls. Age and gender were not well balanced, especially in module 2. The authors said generalization to new data may be limited. The paper discusses possible future use for triage, but did not test a triage workflow.
Declared interests
Funding named: Beckman Coulter Foundation and Hartwell Foundation. The publication types also list NIH and non-U.S. government research support.
The easy way to misread this
Do not read the high test-set scores as proof that these models can diagnose autism in practice. They were built on retrospective ADOS records with few controls, and the authors said generalization may be limited.