An artificial intelligence tool can detect autism spectrum disorder with 100 percent accuracy simply by scanning images of children's eyes, according to a new study.
If confirmed, it would be a major breakthrough in detecting the disease. But several autism experts told that the number was unrealistic and the result was probably “too good to be true.”
Autism is estimated to affect one in 36 children in the United States, but many children remain undiagnosed until later in childhood, depriving them of possible treatments.
If a technological solution could help reduce long wait times for autism specialists or other barriers to diagnosis, it could benefit millions of families.
A new AI tool can detect autism with 100 percent accuracy using retinal scans, its inventors say. Autism experts are unconvinced, saying results are 'too good to be true'
Autism is a disorder that is associated with altered brain development and the optic nerve connects the retina to the brain via a very short path.
So it stands to reason that brain differences could be reflected in the eyes.
Dozens of news outlets have picked up the news about the AI tool, developed by a team of researchers at Yonsei University in Seoul.
But experts say it's still too early to trust these results and that the research raises several warning signs — starting with the 100 percent accuracy number.
“Something is clearly wrong here,” Fred Shic, an autism researcher at the Yale School of Medicine, told . Shic researches eye tracking and imaging techniques in autistic children.
“There is no way this test is more accurate than that used by doctors,” he told . “This reliability is not 100 percent even among the best clinicians in the world.”
Other autism experts share Shic's skepticism.
“It just seems too good to be true,” Cathy Lord, distinguished professor of psychiatry at the University of California, Los Angeles, told .
Lord is co-developer of the Autism Diagnostic Observation Schedule-Second Edition (ADOS-2), the gold standard clinical tool used to assess the children in the new study.
She said she hopes other researchers will try to replicate the results by running the experiment again and comparing the results to this one.
“It seems worth trying to emulate, but I'm very skeptical,” she added.
has reached out to the study's authors and will update this story if we receive a response.
958 children took part in the study in question: 479 with autism and 479 without autism.
Both groups had the same split of boys and girls – 82 percent boys and 18 percent girls – which corresponds to the 4:1 gender ratio found in most countries.
The researchers fed images of children's retinas to train the algorithm, excluding children with other psychiatric conditions that could complicate or confuse the results.
Specialists examined the children using the ADOS-2 to confirm that they had autism and to assess how pronounced their autism traits were.
A deep neural network was trained to distinguish between children with and without autism using iris scans. They also learned how to correlate the severity of autism traits with retinal scans.
Retinal scans have been shown to be effective in screening for certain diseases such as Alzheimer's
When the AI tool was tested on a different group of children than the one it was trained on, it accurately recognized the children's diagnosis 100 percent of the time, according to the study, published in JAMA Network Open.
Additionally, the severity of autism could be determined with about 74 percent accuracy using retinal scans alone.
The idea of scanning the retina to detect autism is “intriguing and promising,” Geri Dawson, director of the Duke Autism Clinic, told . “Changes in the retina have also been used to predict Alzheimer’s disease.”
Several studies have shown that autistic and neurotypical people have significant differences in the nerves of the retina.
However, more research is needed to determine whether the differences between the two groups of children are due to autism or another factor.
“The differences could be an indicator of brain changes more generally associated with cognitive disabilities, for example,” Dawson said. “In addition, the authors note that many of the autistic participants were taking medications that could have affected the retina.”
The average intelligence quotient (IQ) of the autism group was 70, exactly at the limit for a diagnosis of intellectual disability.
However, the researchers didn't provide IQ scores for the non-autistic children, so this unaccounted factor makes the study even more complicated, Thomas Frazier, a professor of psychology at John Carroll University, told .
“That makes the comparison worth it [typically developing children] “For clinical purposes, it’s even less realistic,” he said.
But even that wouldn't ensure 100 percent accuracy, Lord said.
“If it were just about IQ, you would still expect less perfect results.”
Several research teams are working on smartphone or tablet-based apps to detect autism, but these apps focus on social attention rather than retinal scans
The AI model itself could also be the problem, several experts said.
Something other than the retina, which reveals a child's diagnosis, could somehow be included in the images, driving the AI to its unrealistic 100 percent value, Shic said.
“This could be as simple as a word describing the data source of specialized companies.” [autism] Clinics,” he said. There could also be minor changes in image quality between the two groups.
At an acoustic engineering conference in 2013, Shri Narayanan's lab entered a competition to develop a method for identifying autistic children from voice recordings.
They have achieved very good results, Narayanan, University Professor and Nikias Chair in Engineering at the University of Southern California, told
However, it turned out that they were caused by a hidden factor: sound quality.
Their system's performance was actually due to the difference between noisy main classrooms and quiet special education classrooms – a difference reflected in the voice recordings – “which incorrectly produced what appeared to be the correct answer,” he said.
“The promise of being able to screen and diagnose a clinical condition with data and new (AI) tools, while exciting and can be impactful, must be implemented with extreme care and caution,” Narayanan said.
There are already some apps in development that are intended to diagnose autism.
They tend to track where a child is looking rather than the actual structure of their eyes.
Because social communication problems are one of the core features of autism, researchers have tried to screen children for autism by asking whether they pay more attention to objects than people.
One such app, developed by a team at Duke University, predicted autism diagnosis with 90 percent accuracy in a 2021 study.
Using eye-tracking software and smartphone cameras, this app showed toddlers videos of people talking and playing with toys and detected whether toddlers paid more attention to the toys or people.
But any technology used to diagnose or treat a disease still must undergo review and clearance by the Food and Drug Administration, and few studies of new techniques ultimately make it that far.