Breaking

Teen Builds AI Reporting 89% Accuracy in ASD/ADHD Screening From Retina

July 7, 2026 at 11:32 EDT

Healthcare
Research & Papers
Foundation Models

Edward Kang, a 17-year-old high school student in New Jersey, has built an AI tool called RetinaMind that screens for autism spectrum disorder (ASD) and attention-deficit/hyperactivity disorder (ADHD) from a single retinal image, reporting about 89% accuracy. The tool placed second at the 2026 Regeneron Science Talent Search, earning a $175,000 award.

2026 Regeneron Science Talent Search · 2nd Place

A 17-Year-Old Built an AI That Screens for Autism & ADHD From a Single Eye Photo

Edward Kang, a New Jersey high schooler, taught himself machine learning to create RetinaMind — a tool that reads a retinal image and reports ~89% accuracy in flagging neurotypical, ASD, or ADHD profiles. It's a striking proof-of-concept, not a clinical product.

~89%

Reported accuracy across the three-way classification

$175K

Award for placing 2nd at the Regeneron STS

Age of the self-taught student creator

How the tool decides

A convolutional neural network with ensemble learning turns one fundus photo into three probabilities, then picks the highest.

Retinal image
(fundus photo)

→

CNN + ensemble
models

→

3 probabilities
Neurotypical / ASD / ADHD

→

Grad-CAM heatmap
shows why

The catch: a very small validation set

One reported run used only ~353 images — heavily skewed toward neurotypical cases. Each block below ≈ 25 images.

203

Neurotypical

103

ASD

ADHD

Why it impressed

Detects ASD and ADHD simultaneously from one image
Backed by biology: an in vitro retinal cell model and candidate genes like ABCA4
Non-invasive, objective signal for a diagnosis that can take months or years

Why to stay cautious

Proof-of-concept only — no clinical use or large-scale validation
Retinal changes aren't specific to ASD/ADHD and may overlap other conditions
Accuracy tends to drop when separating ASD from ADHD

Bottom line: a promising student research effort that pairs AI with real biology — but clinical application would still require large-scale validation and far better specificity.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.