A teenager from New Jersey has developed an artificial intelligence (AI) tool that diagnoses autism spectrum disorder (ASD) and attention deficit hyperactivity disorder (ADHD) using retinal images.
During his research for a school project, 17-year-old Edward Kang came across an intriguing 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,” the high school senior at Bergen County Academies recently told Smithsonian Magazine.
Ultimately, Kang’s research efforts led him to develop an AI tool called RetinaMind.
How ASD and ADHD are typically diagnosed
As the fastest-growing neurodevelopmental disorder in the country, ASD affects around three per cent of children in the United States. Nearly 7 million children in the US experience ADHD, which has become one of the most common childhood disorders.
ASD and ADHD are both considered to be “behavioural phenotypes” with no biomarkers. Derived from brain functions, they are also usually accompanied by other brain-based developmental or behavioural issues.
“Both are neurologically based conditions that are described by development of skills or by unusual or problematic behaviours,” says Paul Lipkin, a neurodevelopmental paediatrician at Kennedy Krieger Institute and professor of paediatrics at Johns Hopkins Medicine.
Currently, there are no physical tests to diagnose ASD and ADHD, so medical professionals often use developmental and behavioural tests like the American Psychiatric Association’s Diagnostic and Statistical Manual, the Autism Diagnostic Observation Schedule, and Conners Rating Scales to diagnose the two respective disorders. “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.”
Though studies have found that early intervention, especially in the case of autism, can result in better long-term effects for children, early diagnosis of these disorders is difficult.
The road to RetinaMind
Three years ago, Kang set out to make the Chinese University of Hong Kong researchers’ model even more accurate.
Enrolling in a few online classes and teaching himself the basics of machine learning, Kang was able to develop the first iteration of his model. He created the first basic version of the Convolutional Neural Network (CNN), a type of deep-learning model that is primarily designed to classify images, and a replica of the one used in the study he had found. He says the process boiled down to “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.” This model became a baseline against which he measured new and improved versions.
A diagnostic tool should be able to differentiate between various disorders instead of simply detecting whether a person has one, Kang explains. This led him to add ADHD to his next prototype. “Distinguishing between neurotypical individuals and those with autism is not very difficult, and existing studies have already achieved close to 100% accuracy,” he says. Identifying distinct disorders is a much harder task and one that is clinically important.
To further improve the model’s accuracy and effectiveness, Kang also employed a few more advanced computational techniques. For example, he used a technique in which different models are given the same task, called ensemble learning. “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. He then calculates an average from their prediction. This means the results are more reliable, Kang explains, because using multiple models and a voting approach “tends to be more accurate, and performance can improve.”
The retinal difference in people with autism and ADHD
To understand the underlying biological mechanisms and foundations that cause retinal differences in people with autism and those with ADHD, Kang began working more on the cell biology, “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 explains.
Kang used gradient-weighted class activation mapping, or GradCAM, an explainable AI technique that identifies the specific regions of an image most useful to the model in making a prediction. This allowed him to explore the inner workings of the CNN and to identify which region of the initial input image the model took into consideration to complete its task. Kang explains that, “in this case, that would mean which part of the retina was important for making a diagnosis of autism and ADHD.”
Previous research identified many retinal features that differ, on average, in people with autism or ADHD. Specialised tools, like optical coherence tomography scans, can detect differences in the length, thickness, and depth of the macula, retinal nerve fibre layers, and other regions. These differences are very subtle and overlap heavily with the normal range seen in neurotypical individuals, presenting a challenge for clinicians. They cannot simply look at a retinal image and diagnose autism or ADHD.
In his research, Kang 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 explains. “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.”
Somewhat counterintuitively, the goal of RetinaMind is to use a retinal image to predict something unrelated to the eye. Once the AI tool has the image, it will analyse it, breaking down percentages of its confidence that a retinal image indicates whether a patient is neurotypical or has autism or ADHD.
“The diagnosis with the highest confidence becomes the official diagnosis of the model,” Kang explains. To support a diagnosis, the model produces a heat map visualisation of the retinal image, highlighting in red the key parts that led to the diagnosis. With an accuracy rate of about 89%, RetinaMind could help answer the complex question of why retinal development differs in people with neurodevelopmental disorders, Kang says.
Solving urgent global challenges
At the 2026 Regeneron Science Talent Search, Kang’s invention won second place and an award of US$175,000 (RM713,825).
“Edward’s project stood out for combining AI with lab-based biology, which gave it both computational sophistication and biological depth,” says Maya Ajmera, president and CEO of Society for Science. A nonprofit dedicated to expanding scientific literacy, the Society hosts the Talent Search, which is the oldest and most prestigious science, technology, engineering and math competition for high school students in the United States. At a time when getting a diagnosis can take months or even years, notes Ajmera, early screening could make a major difference for a lot of families. “He focused on real-world challenges – on autism and ADHD.”
Kang didn’t just build a model, Ajmera explains, “he also explored the underlying gene changes, which strengthened the scientific rigour and helped explain why the patterns might exist.”
The future of RetinaMind
While Lipkin is excited at the potential of the project to lead to earlier diagnosis, he’s quick to caution that autism and ADHD are developmental and behavioural conditions rooted in the brain with much overlap between them and other conditions. “Any retinal differences identified may not be specific for these conditions,” he notes, “but instead of some brain-based neurologic condition generally.”
Agreeing with Lipkin’s concerns, Kang says that his model makes a blanket diagnosis of ASD or ADHD. “But within these kinds of disorders, it’s a very wide spectrum of different kinds of conditions,” he says. Already thinking of next steps to train the model to distinguish between levels of autism, Kang explains that the more specific information a model can produce, the more effective it is at guiding treatment, ultimately ensuring that a child will receive the right amount of support. “I think that can be something powerful for the future.”
“My hope is that RetinaMind will enable earlier diagnoses for neurodevelopmental disorders than currently possible,” says Kang. Ideally, he adds, it could unlock earlier treatment and “a higher quality of life for the millions of patients of autism and ADHD around the world.” – Inc./TNS
