A New Diagnostic Lens for Neurodevelopment
Edward Kang, a 17-year-old student at Bergen County Academies in New Jersey, has developed an artificial intelligence tool called RetinaMind that analyzes retinal images to identify signs of autism and ADHD with approximately 89% accuracy. The project earned Kang a $175,000 award at the 2026 Regeneron Science Talent Search, a competition for high school students in the United States.
Decoding Retinal Patterns
RetinaMind functions by processing images of the retina to generate confidence percentages for three categories: neurotypical, autism, or ADHD. The system employs ensemble learning, a technique that combines predictions from multiple computational models to increase reliability. According to project documentation, the software generates a heat map that highlights specific retinal regions—such as the macula and nerve fiber layers—that influenced the AI’s diagnostic prediction.
The tool identifies subtle structural patterns, including variations in the depth and thickness of retinal tissue, which can be difficult for human clinicians to detect during an isolated analysis. Kang’s research suggests these patterns may overlap with neurotypical ranges, necessitating the high-precision analysis offered by machine learning.
From Chinese University Research to Genetic Discovery
Kang began the project three years ago after discovering research from the Chinese University of Hong Kong regarding the link between retinal imaging and neurodevelopmental conditions. He initially built a basic convolutional neural network (CNN) as a baseline before expanding the model to differentiate between autism and ADHD.
Beyond the computational model, Kang has investigated the underlying biology of these retinal differences. Using autism cell models, he identified twelve candidate genes potentially linked to retinal variations, including ABCA4, a gene associated with proteins involved in retinal detoxification.
The Clinical Divide
Currently, medical professionals rely on behavioral and developmental assessments, such as the DSM-5, ADOS, and Conners Rating Scales, to diagnose autism and ADHD. There are no physical exams to diagnose autism and ADHD.
Paul Lipkin, a pediatrician specializing in neurodevelopment at the Kennedy Krieger Institute and a professor of pediatrics at Johns Hopkins Medicine, notes that while the AI shows promise, autism and ADHD are fundamentally neurological and developmental conditions rooted in brain function. Lipkin cautioned that retinal differences may reflect broader neurological characteristics rather than being specific to a single disorder.
Refining the Spectrum
While RetinaMind provides a significant step in identifying neurological markers, Kang acknowledges the current limitations of the system. The model currently provides a general indication of autism or ADHD. His next phase of development involves training the system to distinguish between mild, moderate, and severe manifestations within the autism spectrum.
Maya Ajmera, president and CEO of the Society for Science, highlighted the project for its integration of computational sophistication with laboratory biology. The project’s success in the 2026 Regeneron Science Talent Search reflects an ongoing focus on using AI to address complex, multi-faceted global health challenges.
Keep reading