A teenager with no formal programming background has developed an artificial intelligence system that could one day change how autism and ADHD are identified. What began as a high school research project has grown into an award-winning proof of concept that uses something as ordinary as a photograph of the eye to detect patterns linked to neurodevelopmental conditions.
Seventeen-year-old Edward Kang’s invention, called RetinaMind, has drawn attention from scientists, educators, and healthcare professionals after demonstrating an accuracy rate of about 89 percent in distinguishing between autism spectrum disorder (ASD), attention-deficit/hyperactivity disorder (ADHD), and neurotypical development. While experts stress that the technology is still in its early stages and requires extensive clinical validation, the project highlights how artificial intelligence could support earlier screening for conditions that often take years to diagnose.
A High School Research Project Became an Award-Winning Innovation
Edward Kang, a senior at Bergen County Academies in Hackensack, New Jersey, did not set out to reinvent autism screening. His journey began in 2023 while searching through scientific papers for a school project.
One study immediately caught his attention. Researchers at the Chinese University of Hong Kong had demonstrated that subtle retinal differences could be associated with 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,” Kang explained.
Rather than stopping at understanding the published research, he challenged himself to improve it.
That decision eventually led to RetinaMind, an AI-powered system capable of analyzing retinal photographs and identifying patterns associated with both autism and ADHD. The project impressed judges at the 2026 Regeneron Science Talent Search, where Kang earned second place along with a $175,000 prize.
According to Maya Ajmera, president and CEO of Society for Science, the project stood out because it combined sophisticated artificial intelligence with biological research instead of relying on computer science alone.
“Edward’s project stood out for combining A.I. with lab-based biology, which gave it both computational sophistication and biological depth,” Ajmera said.
The recognition places Kang among the country’s most promising young scientists, but the broader significance lies in the healthcare challenge he chose to tackle.
Why Earlier Autism and ADHD Detection Matters
Autism spectrum disorder and ADHD are among the most common neurodevelopmental conditions affecting children.
The Centers for Disease Control and Prevention estimates that autism affects roughly one in 31 children in the United States, while ADHD impacts nearly seven million American children. Although awareness has increased substantially over the past decade, obtaining a diagnosis often remains a lengthy and complicated process.
Unlike many medical conditions that can be confirmed through blood tests or imaging studies, autism and ADHD are diagnosed through developmental assessments, behavioral observations, interviews with caregivers, and standardized psychological evaluations.
These evaluations are thorough because clinicians are assessing how a child communicates, learns, behaves, interacts socially, and responds to different situations over time.
Paul Lipkin, a neurodevelopmental pediatrician at the Kennedy Krieger Institute and professor of pediatrics at Johns Hopkins Medicine, explains that both autism and ADHD are behavioral phenotypes rooted in brain development.
“They are neurologically based conditions that are described by development of skills or by unusual or problematic behaviors,” Lipkin said.
This process is essential for making an accurate diagnosis, but it also contributes to long waiting periods in many communities.
Families frequently spend months, and sometimes much longer, waiting for specialist appointments. During that time, parents may notice developmental concerns without having clear answers about what their child is experiencing.
Early intervention has consistently been associated with improved long-term developmental outcomes, particularly for children with autism. Earlier access to speech therapy, occupational therapy, behavioral interventions, educational support, and family guidance can help children develop communication, social, and adaptive skills during important stages of brain development.
That is why researchers continue searching for faster and more objective screening tools that could identify children who may benefit from a comprehensive clinical evaluation.
RetinaMind was designed with that goal in mind.
“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,” Kang said.
Looking to the Eye for Clues About the Brain
At first glance, the idea sounds surprising. How could a picture of the retina reveal anything about conditions that primarily affect brain development?
The answer lies in biology.
The retina is technically part of the central nervous system. During fetal development, retinal tissue and brain tissue develop from the same embryonic structures. Because of this close relationship, scientists have long suspected that changes occurring in brain development might also leave subtle signatures within the eye.
Previous research has identified small structural differences in the retinas of some people with autism and ADHD.
These include slight variations in the thickness of retinal nerve fiber layers, differences within the macula, and other microscopic structural features.
The challenge is that these changes are far too subtle for the human eye to detect consistently.
A trained ophthalmologist cannot simply examine a retinal photograph and determine whether someone has autism or ADHD.
Artificial intelligence, however, excels at recognizing complex visual patterns that humans may never consciously notice.
Machine learning systems can examine thousands of images simultaneously, identifying combinations of tiny features that collectively point toward meaningful differences.
That capability became the foundation of RetinaMind.
Teaching Artificial Intelligence to Recognize Hidden Patterns
One of the most remarkable aspects of Kang’s project is that he taught himself the programming skills needed to build it.
“I don’t really come from a programming background,” he admitted.
He relied on online tutorials and virtual courses while learning machine learning from the ground up.
His first version was intentionally simple.
Using a convolutional neural network, or CNN, Kang recreated the type of deep-learning model described in the original research paper that inspired him.
Convolutional neural networks have become one of the most widely used forms of artificial intelligence for image recognition because they excel at identifying shapes, textures, edges, and spatial relationships inside photographs.
Initially, the model focused on distinguishing between autism and neurotypical development.
That was only the beginning.
Kang believed a clinically useful screening tool should do more than simply recognize one condition.
Instead, it should distinguish among multiple neurodevelopmental disorders that often share overlapping characteristics.
He expanded the system to recognize ADHD as well, creating a model capable of classifying three possible outcomes instead of only two.
Improving accuracy became his next challenge.
Rather than depending on one AI model, Kang implemented an approach known as ensemble learning.
Instead of asking a single algorithm to interpret each retinal image, several different models analyze the same photograph independently.
Each model reaches its own conclusion before their predictions are combined into a final consensus.
“You feed them the same retinal image and ask them to predict autism or ADHD, and then you take their predictions and combine them,” Kang explained.
This voting approach helps reduce errors that may occur when one individual model makes an incorrect prediction.
The combined decision tends to be more reliable because it reflects agreement across multiple systems rather than relying on one computational opinion.
That strategy helped RetinaMind reach an overall diagnostic accuracy of approximately 89 percent.
Making Artificial Intelligence More Transparent
Artificial intelligence has often been criticized for functioning like a “black box.” A model may produce the correct answer, but clinicians still want to understand how it reached that conclusion before trusting it in healthcare settings.
Kang addressed that concern by incorporating a technique called Gradient-weighted Class Activation Mapping, commonly known as Grad-CAM.
Rather than simply generating a diagnosis, the technology creates a visual heat map highlighting the retinal regions that influenced the prediction.
The colored overlays reveal which structures attracted the AI’s attention while analyzing each retinal image.
According to Kang’s research, the system frequently focused on areas including the macula, optic disc, fovea, and surrounding retinal blood vessels.
These visual explanations make the system easier for researchers to evaluate while helping determine whether the AI is relying on meaningful biological information instead of random image artifacts.
For healthcare professionals, explainable AI is becoming an increasingly important requirement as machine learning moves closer to clinical practice.
Understanding why a model reaches its conclusion can strengthen confidence in its recommendations while identifying situations where additional research is needed.
RetinaMind’s ability to generate these heat maps represents another step toward making AI-assisted screening tools more transparent and clinically useful.
Connecting Computer Predictions to Real Biology
Building an accurate artificial intelligence model was only part of Edward Kang’s ambition. He also wanted to understand why the retina might reveal clues about neurodevelopmental conditions in the first place.
To answer that question, he expanded his research beyond computer science and into cell biology.
Beginning in late 2024, Kang developed laboratory models of retinal cells to investigate whether biological changes associated with autism could explain the patterns his AI was detecting. Instead of treating RetinaMind as a system that simply produced predictions, he wanted to uncover the biological mechanisms behind those predictions.
“I really began working more on the cell biology side,” Kang said, “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.”
His research identified a dozen candidate genes that may help explain differences in retinal development among people with autism.
One of the most significant discoveries involved the ABCA4 gene.
This gene produces a protein responsible for helping remove toxic byproducts from retinal cells. In Kang’s laboratory model, retinal cells associated with autism showed lower levels of ABCA4 expression than healthy control cells.
“My retinal cell autism model showed less ABCA4 expression compared to the control,” Kang explained. “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.”
The finding does not establish that reduced ABCA4 causes autism. Instead, it provides one possible biological explanation for why artificial intelligence may be able to detect subtle retinal changes linked to neurodevelopment.
Researchers will need additional studies to determine how these genes interact with brain development and whether similar biological patterns appear across larger and more diverse populations.
Still, by combining laboratory biology with artificial intelligence, Kang strengthened the scientific foundation of his work and moved beyond simply creating an accurate computer model.
A New Way of Thinking About Autism Screening
RetinaMind is not designed to replace psychologists, pediatricians, neurologists, or developmental specialists.
Instead, Kang envisions it as a screening tool that could help identify children who may benefit from a comprehensive evaluation much earlier than they otherwise would.
The process itself is relatively straightforward.
A standard retinal photograph is uploaded into the system. The AI analyzes thousands of tiny visual characteristics that are too subtle for clinicians to detect consistently. It then estimates the probability that the image belongs to someone who is neurotypical or someone with autism or ADHD.
To make its predictions easier to interpret, RetinaMind also generates a heat map that visually highlights the retinal regions influencing its decision.
This transparency could prove valuable if similar systems eventually move into healthcare settings.
Unlike many current assessments that depend heavily on behavioral observation over time, retinal imaging is fast, non-invasive, and already widely used in ophthalmology clinics.
If future clinical studies validate RetinaMind’s performance across larger populations, similar technology could eventually become an additional screening option that complements existing evaluations rather than replacing them.
That distinction is important.
Medical experts emphasize that autism and ADHD involve far more than biological measurements alone.
Developmental history, communication skills, behavior, learning, family observations, and clinical expertise will remain essential parts of diagnosis.
Artificial intelligence may eventually help identify children who need further evaluation, but the final diagnosis would still require experienced healthcare professionals.
Experts See Promise but Urge Careful Validation
While RetinaMind has generated considerable excitement, researchers are equally careful about discussing its current limitations.
Paul Lipkin believes retinal imaging has genuine potential, but he also notes that autism and ADHD are highly complex conditions with significant overlap among many neurological disorders.
“Any retinal differences identified may not be specific for these conditions, but instead of some brain-based neurologic condition generally,” Lipkin explained.
In other words, some retinal features detected by artificial intelligence could reflect broader differences in neurological development rather than autism or ADHD specifically.
That possibility underscores why additional clinical testing remains essential.
Artificial intelligence models often perform well during early research but require validation across thousands of patients representing different ages, ethnic backgrounds, medical histories, and imaging equipment before they can be considered reliable clinical tools.
Researchers must also determine whether similar accuracy can be maintained outside carefully controlled research environments.
Kang shares these concerns.
Rather than presenting RetinaMind as a finished diagnostic solution, he consistently describes it as a proof of concept with room for substantial improvement.
“Right now, my model just makes a blanket diagnosis of either autism spectrum disorder or attention deficit hyperactivity disorder,” Kang said. “But within these kinds of disorders, it’s a very wide spectrum of different kinds of conditions.”
His next objective reflects a deeper understanding of clinical care.
Instead of simply identifying autism, he hopes future versions of RetinaMind will distinguish among different presentations and levels of support needs.
“The more specific information we can get out of the model,” he explained, “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.”
That vision aligns with the broader movement toward personalized medicine, where treatments can be tailored to each patient’s individual needs rather than relying on broad diagnostic categories alone.
Seeing Autism Care Up Close Changed His Perspective
As Kang’s technical skills developed, another experience transformed how he thought about autism research.
During his senior year, he volunteered at the Rutgers Center for Autism Research, Education and Services (RUCARES), part of the Rutgers Brain Health Institute.
The internship allowed him to observe clinicians working directly with children and families affected by autism.
Instead of seeing autism only through datasets and scientific papers, he witnessed the day-to-day reality of diagnosis, treatment, and individualized care.
“My volunteer internship at RUCARES continues to be a special honor,” Kang said. “Since last fall, I have had the opportunity to actually see the autism treatment practices I had only read about in articles before.”
Watching specialists personalize therapies for each child influenced the direction of his research.
“Getting to see first-hand how personalized treatments can be gave me direction for my research project,” he said, “helping me realize that future diagnostic tools for autism should aim to not only identify the disorder, but also distinguish between various subtypes of patients that could guide treatment.”
The internship also reinforced his interest in neuroscience.
“More than anything, though, I have had the privilege of connecting and conversing with professionals in the field, who have warmly shared their knowledge and excitement for their work. RUCARES has fueled and validated my love for neuroscience.”
According to Craig W. Strohmeier, a licensed psychologist and behavior analyst at the RUCARES Severe Behavior Program, students gain valuable experience by observing clinicians collect data, review treatment plans, and participate in applied research projects.
Those experiences gave Kang a broader appreciation for the role technology should play in healthcare.
Rather than replacing clinicians, better tools should help them deliver more personalized care.
Curiosity Continues to Drive the Research
Winning one of the nation’s most prestigious science competitions would be enough to satisfy many young researchers.
For Kang, it marked the beginning of even bigger questions.
Throughout the two-year project, he discovered that each answer opened new areas of investigation.
“The truth is, I have way more questions now than I started with,” he said. “That can be frustrating to think about sometimes, but I believe true scientific spirit comes from tackling those questions head-on and loving every step of the process.”
He also says the most meaningful conversations are not with judges or fellow researchers.
Instead, they happen with families living with autism or ADHD.
“I am always most grateful when I get to discuss my work with families affected by ASD or ADHD,” Kang said. “The thought of making some contribution to their lives is my strongest motivation.”
This perspective reflects an important shift occurring throughout medical artificial intelligence.
The ultimate goal is not simply building smarter algorithms. It is developing tools that improve patient care, reduce delays in diagnosis, and help families access support sooner.
What RetinaMind Could Mean for the Future
Artificial intelligence is becoming part of many areas of healthcare, from reading medical scans to identifying disease risks that might otherwise go unnoticed.
RetinaMind represents another example of how machine learning could eventually support clinicians by finding meaningful patterns hidden within ordinary medical images.
For families affected by autism and ADHD, earlier recognition can open the door to earlier interventions, educational planning, and support services during critical stages of development.
Much more research must happen before retinal AI screening becomes part of routine clinical practice. Larger studies, independent validation, and regulatory review will all be necessary before technologies like RetinaMind can move beyond research settings.
Even so, Kang’s work demonstrates how curiosity, persistence, and interdisciplinary thinking can lead to remarkable innovation.
A project that started with a teenager reading scientific papers has grown into research that combines artificial intelligence, genetics, neuroscience, and ophthalmology in ways that even experienced scientists find promising.
As Kang prepares to attend the Massachusetts Institute of Technology, RetinaMind stands as more than an award-winning science project. It offers a glimpse of how future healthcare may combine advanced technology with human expertise to identify developmental conditions earlier, understand them more deeply, and ultimately help children receive the support they need when it can make the greatest difference.
