Combining voice and language features improves automated autism detection.
Heather MacFarlane, Alexandra C Salem, Liu Chen and 2 others
PMID 35460329WHAT IT FOUND
Combining automated voice and language features from recorded conversations classified children already diagnosed with autism more accurately than either set alone.
The study was not intended for diagnosis or screening.
Key findings
01The combined voice and language model classified ASD status with accuracy 0.8671, sensitivity 0.8977, and specificity 0.8286, substantially better than either single-modality model.
02The combined model's area under the ROC curve, a measure of how well it separates ASD from non-ASD, was 0.9205, significantly higher than the voice model's 0.7800 and the language model's 0.8748.
03Nine participants with ASD (10.2%) were missed by the combined model, and they had the highest means for IQ 105.1, communication units 108.9, and words 627.2 among subgroup means.
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 study was not intended for diagnosis or screening; it used already-diagnosed children to validate voice and language measures. The sample was limited to children aged 7 to 17 with normal IQ who spoke fluent English and could complete ADOS-2 Module 3. The paper states that these findings cannot currently be generalized to different contexts. The authors state they do not yet know the reliability or short-term stability of the measures. The analysis relied on manual transcription, and transcription reliability results were not available. Only nine missed participants with ASD were available for further subgroup comparisons. The small number made further comparisons less reliable. The groups were not tightly matched, which the authors note could affect the analysis.
Declared interests
The article is marked as NIH extramural research support. The supplied text does not include an author conflict-of-interest declaration.
The easy way to misread this
Do not read the 0.8671 accuracy as a validated clinical diagnostic test. The study used already-diagnosed children aged 7 to 17 with normal IQ who spoke fluent English, was not intended for diagnosis or screening, and the authors state they do not yet know the reliability or short-term stability of the measures.