17-year-old’s AI screens autism and ADHD with 89% accuracy

Signals Inbox·July 7, 2026·Mental Health

The signal is exciting, but the jump from science fair breakthrough to real medical diagnosis is massive.

Q1What actually happened?

Edward Kang, a 17-year-old student from New Jersey, built RetinaMind, an AI system that analyzes retinal images to screen for autism and ADHD. The headline number is 89% accuracy, and the project won a $175,000 prize at the Regeneron Science Talent Search.

Q2Why would the retina say anything about autism or ADHD?

Because the retina is not just a camera surface. It is nervous system tissue. It develops from the same early embryonic tissue as the brain, and it contains neurons, blood vessels, and structural patterns that can reflect broader neurodevelopmental differences. The bet here is that tiny retinal patterns may act like a cheap window into the brain. Not a perfect window, but maybe a useful one.

Q3So is this better than neurologists?

No, and that framing is a bit too spicy. Neurologists and clinicians are not failing because they cannot look hard enough at a retina. Autism and ADHD are diagnosed through behavior, development history, interviews, rating scales, and clinical judgment. RetinaMind is interesting because it could become a faster first-pass tool. It does not replace the full clinical process yet.

Q4What does 89% accuracy actually mean?

It means the model performed well on the dataset it was tested on. But accuracy alone can hide a lot. We need to know the sample size, the mix of autism versus ADHD versus control cases, the age range, image quality, ethnicity, other eye conditions, and whether the model was tested on totally new patients from different clinics. A model can look amazing in a controlled dataset and become much messier in the real world.

Q5Why is this still a big signal?

Because neurodevelopmental diagnosis is slow, expensive, and often inaccessible. Families can wait months or years for an autism or ADHD evaluation. Retinal imaging is already common in eye care, relatively cheap, and non-invasive. If this approach keeps working across larger studies, it could become a simple early warning layer inside schools, pediatric clinics, or routine eye exams.

Q6What would need to happen before this becomes medical?

It needs external validation. That means larger datasets, different hospitals, different cameras, different age groups, and blind testing where the model sees patients it has never been tuned on. Then comes regulatory review, clinical workflow design, and a clear rule for what doctors should do with the output. The hard part is not making a cool model. The hard part is proving it works when the world gets messy.

Q7Could this work for both autism and ADHD?

Maybe, but that is one of the most important questions. Autism and ADHD can overlap, and many people have both. A model that separates autism, ADHD, both, and neither would be much more useful than a model that only says neurodevelopmental disorder detected.