Machine Learning-Based Early Prediction Model for Autism Spectrum Disorder in Infants Using Acoustic Feature.
Ten minutes of baby sounds can predict autism at nine months with 93 percent accuracy.
01Research in Context
What this study did
The team recorded 10-minute Still-Face videos of 88 babies.
They pulled 39 sound features from each clip.
A computer model learned which sound pattern matched later autism diagnosis.
What they found
The model flagged autism with 93 percent accuracy.
It worked just as well for boys and girls.
The test spotted risk as early as nine months.
How this fits with other research
O'Brien et al. (2026) hit 99.6 percent accuracy using mom’s weight and income data, but only at birth.
The new acoustic test gives nearly the same power months after birth, when signs first show.
Marsack-Topolewski et al. (2025) reached 97-99 percent with parent forms across all ages.
The baby cry model narrows that power to infants and needs no parent paperwork.
Ochi et al. (2024) showed the same idea works in adults, proving voice markers hold up across life stages.
Why it matters
You can’t start ABA until you have a diagnosis.
A cheap mic and ten minutes could move that clock up by a full year.
If you run infant groups, tape a Still-Face session and send the file for scoring.
Earlier intake means earlier therapy and better long-term gains.
Want CEUs on This Topic?
The ABA Clubhouse has 60+ free CEUs — live every Wednesday. Ethics, supervision & clinical topics.
Join Free →Add a voice recorder to your infant assessment kit and save a Still-Face clip for scoring.
02At a glance
03Original abstract
This study aimed to create a machine learning-based predictive model for early detection of autism spectrum disorder (ASD) in infants using acoustic features. Conducted as a prospective cohort at Nanjing Medical University from 2019 to 2024, infants aged 9-18 months from an ASD sibling cohort participated. Behavioral and vocalization data were gathered during the Still-Face Paradigm, with ASD diagnoses confirmed at 36 months through ADOS and ADI-R assessments. Researchers extracted 4368 acoustic features from the recordings and applied LASSO regression for dimensionality reduction, identifying 39 key features. A support vector machine (SVM) classifier was then developed, tested with four kernel functions-linear, radial basis function, polynomial, and sigmoid-via tenfold cross-validation. The final sample included 88 infants, 28 of whom were diagnosed with ASD. The sigmoid kernel yielded the best results, achieving a 92.86% sensitivity, 93.33% specificity, and a 93.18% accuracy. Notably, spectral and energy-related features were significantly higher in ASD infants (p < 0.01). These findings suggest that acoustic features can serve as early, noninvasive biomarkers for ASD, and the SVM model demonstrates significant promise for early screening and intervention efforts.
Autism research : official journal of the International Society for Autism Research, 2026 · doi:10.1002/aur.70179