This High Schooler Developed an A.I. Tool to Diagnose Autism and ADHD Using the Retina

Why It Matters

Early, objective screening could cut months from diagnosis, enabling timely interventions that improve outcomes for millions of children with autism or ADHD.

Key Takeaways

• RetinaMind achieves 89% diagnostic accuracy for autism and ADHD
• Ensemble learning and GradCAM improve model reliability and explainability
• $175,000 Regeneron prize spotlights AI‑driven neurodevelopmental diagnostics
• Identified ABCA4 gene as potential link between retina and autism

Pulse Analysis

Autism spectrum disorder and attention‑deficit hyperactivity disorder affect roughly one in 54 children and seven million U.S. youths respectively, yet clinicians still rely on behavioral questionnaires and lengthy observation periods to confirm a diagnosis. The absence of a physiological biomarker has long hampered early detection, which research shows can dramatically improve long‑term outcomes. Recent advances in optical coherence tomography have revealed subtle retinal anomalies in neurodivergent individuals, sparking interest in the eye as a window to brain development. Against this backdrop, Edward Kang’s RetinaMind offers a non‑invasive, image‑based alternative that could reshape screening protocols.

Kang built RetinaMind on a convolutional neural network that ingests high‑resolution retinal scans and outputs probability scores for autism, ADHD, or typical development. By employing ensemble learning—running multiple models on the same image—and aggregating their predictions, he lifted overall accuracy to about 89 percent, a notable achievement for a proof‑of‑concept. The system also integrates Grad‑CAM heat maps, granting clinicians visual insight into which retinal regions drive the decision, thereby addressing the ‘black‑box’ criticism of many AI diagnostics. Simultaneously, Kang’s exploratory work on the ABCA4 gene provides a molecular hypothesis linking retinal changes to neurodevelopmental pathology.

The commercial potential of a rapid, eye‑based screening tool is substantial. Payers and school‑based health programs are eager for cost‑effective diagnostics that truncate the months‑long referral cascade, while ophthalmic device manufacturers could embed the algorithm into existing OCT platforms. However, regulatory approval will require large‑scale, multi‑ethnic validation studies to prove specificity beyond other neurological conditions. If these hurdles are cleared, RetinaMind could catalyze a new class of biometric tests for mental health, encouraging further investment in AI‑driven ophthalmic biomarkers and accelerating the shift toward precision pediatric care.