A school project led Edward Kang to an idea he thought was “really unintuitive,” using the eye to read what is happening in the brain.

Three years ago, Kang came across a study by researchers at the Chinese University of Hong Kong that used retinal images to diagnose autism. “I thought it was fascinating and really unintuitive that you can use something like the eye to understand what’s happening in the brain,” says Kang to Smithsonian Magazine, now a 17-year-old high school senior at Bergen County Academies in Hackensack, New Jersey.

He set out to improve the researchers’ model and developed RetinaMind, an A.I. tool that diagnoses autism spectrum disorder and attention deficit hyperactivity disorder using retinal images.

RetinaMind has an accuracy rate of about 89 percent. Kang’s invention won second place and an award of $175,000 at the 2026 Regeneron Science Talent Search.

Autism spectrum disorder affects 1 in 54 children in the United States, and attention deficit hyperactivity disorder is experienced by nearly seven million children in the U.S.

“Both are neurologically based conditions that are described by development of skills or by unusual or problematic behaviors,” says Paul Lipkin, a neurodevelopmental pediatrician at Kennedy Krieger Institute and professor of pediatrics at Johns Hopkins Medicine.

Lipkin says the disorders are considered “behavioral phenotypes” with no biomarkers and are derived from brain functions. “In the case of development,” explains Lipkin, “those affected by autism and/or ADHD frequently have intellectual or learning as well as language disabilities and motor coordination problems.”

Early diagnosis is difficult. There are no physical tests to diagnose autism and ADHD, so medical professionals often use developmental and behavioral tests like the American Psychiatric Association’s Diagnostic and Statistical Manual, the Autism Diagnostic Observation Schedule and Conners Rating Scales.

“My hope is that RetinaMind will enable earlier diagnoses for neurodevelopmental disorders than currently possible, unlocking earlier treatment and, therefore, a higher quality of life for the millions of patients of autism and ADHD around the world,” says Kang.

To build the project, Kang taught himself how to code and learned the basics of machine learning. “I don’t really come from a programming background,” he confesses. “I looked at a lot of different tutorials online.”

He also enrolled in online classes and built an initial model based on a Convolutional Neural Network, or CNN, replicating the one used in the study that first caught his attention. “Trying to replicate what they’ve done and creating a very simplistic model that is just taking the image, getting the diagnosis and training the model based on how well it can predict that diagnosis,” Kang says.

He later added ADHD to the model because, he says, a diagnostic tool should distinguish between disorders instead of only identifying a single condition. “Distinguishing between neurotypical individuals and those with autism is not very difficult, and existing studies have already achieved close to 100 percent accuracy,” he says.

Kang also used ensemble learning, which combines the predictions of different models working on the same retinal image. “You feed them the same retinal image and ask them to predict autism or ADHD, and then you take their predictions and combine them,” says Kang. “It tends to be more accurate, and performance can improve,” he explains.

Since late 2024, Kang has also been studying the biology behind retinal differences in people with autism and ADHD. “I really began working more on the cell biology side,” he says, “creating an in-vitro or cell-based model of autism and studying what kinds of genes may be involved in why autism patients have retinal differences that can be detected to begin with.”

He used gradient-weighted class activation mapping, or GradCAM, to identify which parts of a retinal image mattered most to the model. “In this case, that would mean which part of the retina was important for making a diagnosis of autism and ADHD,” Kang explains.

Researchers have previously identified retinal features that differ on average in people with autism or ADHD, including differences in the length, thickness and depth of the macula, retinal nerve fiber layers and other regions. But those differences are subtle and overlap heavily with the normal range seen in neurotypical individuals, so a clinician alone cannot look at a retinal image and diagnose autism or ADHD.

Kang says his research identified a dozen potential candidate genes linking autism and retinal development. “One potentially interesting gene I identified is ABCA4, which encodes a protein responsible for detoxifying the retina,” he says. “My retinal cell autism model showed less ABCA4 expression compared to the control. This suggests that autistic patients may have less of this detoxifying protein, potentially leading to increased retinal toxicity and degradation, which could explain some of the observed retinal differences.”

Once RetinaMind receives an image, it analyzes it and gives percentages for its confidence that the patient is neurotypical or has autism or ADHD. “The diagnosis with the highest confidence becomes the official diagnosis of the model,” Kang says. The tool also produces a heat map of the retinal image highlighting in red the key parts that led to the diagnosis.

“Edward’s project stood out for combining A.I. with lab-based biology, which gave it both computational sophistication and biological depth,” says Maya Ajmera, president and CEO of Society for Science, which hosts the competition. “He focused on real-world challenges, on autism and ADHD.”

Ajmera says early screening could make a major difference for many families. “He didn’t just build a model,” she adds. “He also explored the underlying gene changes, which strengthened the scientific rigor and helped explain why the patterns might exist.”

Lipkin says the idea has promise, but he also warned that retinal differences may not be specific to autism and ADHD. “Any retinal differences identified may not be specific for these conditions, but instead of some brain-based neurologic condition generally,” he says.

Kang said he shares that concern. “Right now, my model just makes a blanket diagnosis of either autism spectrum disorder or attention deficit hyperactivity disorder,” says Kang. “But within these kinds of disorders, it’s a very wide spectrum of different kinds of conditions.”

He said he wants to train the model to distinguish between mild, moderate and severe autism. “The more specific information we can get out of the model,” he explains, “the more effective it is in terms of guiding treatment and making sure that the child is getting the right amount of support that they need.”

Read more from Smithsonian Magazine.