A neural network approach to the classification of autism.
A 1993 neural net showed computers could out-score stats at autism diagnosis, launching today’s AI screening race.
01Research in Context
What this study did
The team built a computer brain to spot autism. They fed it answers from parent interviews. The goal was to beat old-style statistics at telling who has autism and who does not.
Kids already diagnosed with autism or intellectual disability took part. The computer learned patterns in their parent-interview data. Then it tried to sort new kids into the right group.
What they found
The neural net hit 92% accuracy. Old-style discriminant analysis only reached 85%. The computer brain also handled new cases better.
That 7-point jump mattered in 1993. It showed machines could help clinicians decide.
How this fits with other research
Marsack-Topolewski et al. (2025) now tops this score. Their four-model ensemble reaches 97-99% across toddlers to adults. The 1993 result is still good, but it has been superseded.
Bone et al. (2015) sounds like bad news. They could not copy earlier machine-learning wins. The difference is data size and balance. Small, tidy sets let 1993 shine. Bigger, messier sets expose flaws.
Liu et al. (2016) swapped parent talk for eye tracking. Their ML hit 88.5%, conceptually replicating the 1993 idea with face-scan data instead of words.
Why it matters
This paper is the grandparent of AI screening tools you see today. It proved parent-interview data plus machine learning could beat classic stats. When you read new studies claiming 99% accuracy, remember they stand on this 1993 shoulder. Use the history to ask hard questions about sample size and real-world balance before you buy the next app.
Want CEUs on This Topic?
The ABA Clubhouse has 60+ free CEUs — live every Wednesday. Ethics, supervision & clinical topics.
Join Free →Ask any AI-vendor rep: ‘What was your replication sample size and balance?’
02At a glance
03Original abstract
A nonlinear pattern recognition system, neural network technology, was explored for its utility in assisting in the classification of autism. It was compared with a more traditional approach, simultaneous and stepwise linear discriminant analyses, in terms of the ability of each methodology to both classify and predict persons as having autism or mental retardation based on information obtained from a new structured parent interview: the Autistic Behavior Interview. The neural network methodology was superior to discriminant function analysis both in its ability to classify groups (92 vs. 85%) and to generalize to new cases that were not part of the training sample (92 vs. 82%). Interrater and test-retest reliabilities and measures of internal consistency were satisfactory for most of the subscales in the Autistic Behavior Interview. The implications of neural network technology for diagnosis, in general, and for understanding of possible core deficits in autism are discussed.
Journal of autism and developmental disorders, 1993 · doi:10.1007/BF01046050