A 17-year-old student from New Jersey, the United States, has attracted attention in the scientific community after developing an artificial intelligence (AI) tool that analyses retinal images to screen for signs of autism spectrum disorder (ASD) and attention-deficit/hyperactivity disorder (ADHD).
Edward Kang developed RetinaMind, an AI-powered system that examines images of the retina to identify patterns potentially associated with the two neurodevelopmental conditions.
The idea emerged around three years ago when Kang came across a study by researchers at the Chinese University of Hong Kong exploring the use of retinal images to identify autism.
Intrigued by the possible connection between the eye and the brain, the Bergen County Academies student began investigating how retinal images could provide clues about brain development.
“I thought it was interesting and very unintuitive that you could use something like the eye to understand what is happening inside the brain,” he told Smithsonian Magazine.
The retina and the brain develop from the same embryonic tissue, suggesting that certain differences in brain development may also be reflected in the eye.
How RetinaMind works
Despite having no professional programming background when he started, Kang taught himself coding and machine learning through online classes and internet resources.
He subsequently developed RetinaMind by training an AI model using publicly available retinal image datasets.
The system analyses fundus images, which capture the back of the eye, to identify microscopic patterns associated with autism, ADHD and neurotypical individuals.
According to Kang, RetinaMind has achieved an accuracy rate of approximately 89% in its reported testing.
The tool also generates AI heat maps to highlight areas of retinal images that influence its predictions, offering a way to understand how the model reaches its conclusions.
Kang used Gradient-weighted Class Activation Mapping (Grad-CAM), a technique that helps identify the parts of an image that contribute most strongly to an AI model’s prediction.
This approach addresses a key challenge in medical AI: understanding why a system produces a particular result rather than treating its predictions as a black box.
Kang also developed retinal cell models to investigate possible biological explanations for the differences identified by the AI, including genes that could warrant further research.
Research points to possible genetic links
One finding that attracted attention involved the ABCA4 gene, which plays an important role in retinal health.
Studies using cell models found lower expression of the gene in the autism model compared with controls, suggesting a possible link between biological pathways and the retinal differences observed.
Kang said the findings could help researchers better understand why retinal development may differ in people with neurodevelopmental conditions.
“My autism retinal cell model showed lower ABCA4 expression compared with controls,” he said, explaining that the finding could point towards differences in biological processes affecting the retina.
He added that the list of genes identified through his research could help address unanswered questions about retinal development in people with neurodevelopmental disorders.
However, these findings remain preliminary and require further investigation before firm conclusions can be drawn about their significance.
Young scientist wins US$175,000 award
Kang’s work earned him second place in the 2026 Regeneron Science Talent Search, one of the United States’ leading science competitions for young researchers.
The achievement came with an award of US$175,000 (approximately RM714,210).
His project stood out for combining computer engineering, machine learning, biology and medical research, while attempting to investigate the biological mechanisms behind the AI’s predictions rather than focusing solely on accuracy.
Kang hopes RetinaMind could eventually help identify neurodevelopmental conditions earlier, allowing people to access support and intervention sooner.
“My hope is that RetinaMind will enable earlier diagnoses of neurodevelopmental disorders than is currently possible,” he said, adding that earlier identification could help improve quality of life for people with autism and ADHD.
Can AI diagnose autism and ADHD through the eyes?
Despite its promising results, RetinaMind remains at an early stage of development and requires further research and clinical validation before it can be considered for widespread medical use.
Autism and ADHD are complex conditions with a wide range of characteristics, and retinal differences identified by an AI model may not be unique to either condition. Other biological factors or health conditions could also influence the patterns detected.
An accuracy rate of 89% in a particular test also does not automatically mean the tool will perform equally well across different populations or in real-world clinical settings.
Kang acknowledges these limitations and hopes to refine RetinaMind so that it can eventually distinguish between different conditions and their characteristics more precisely.
The goal, he said, is not to replace doctors but to provide another tool that could assist with early screening.
If future studies confirm its reliability and clinical usefulness, AI-powered retinal imaging could offer a new approach to screening for neurodevelopmental conditions.
For now, RetinaMind represents an encouraging example of how a young researcher’s curiosity and the combination of AI and biology could contribute to future developments in healthcare.
