AI has been taught to look for autism in children from photos with 100 percent accuracy
Korean scientists have found that deep learning models can screen children for autism spectrum disorder (ASD) and determine the severity of symptoms using only retinal photographs of their eyes.
Autism spectrum disorder (ASD) is characterized by two main areas of symptoms: impaired social communication and restricted and repetitive behaviors or interests.
According to the U.S. Centers for Disease Control and Prevention, autism occurs in one in 36 children, with boys being 3.8 times more common on average than girls. In Russia, according to the Ministry of Health, the prevalence of autism spectrum disorders is about 1% of the child population. However, this figure is likely to rise in the future because the proportion of children with ASD is increasing as Westernization normalizes.
Last year, researchers at Stanford University developed an artificial intelligence algorithm that learned how to diagnose autism. It analyzed data from magnetic resonance imaging (MRI), which records neural activity throughout the brain. The AI then studied the impulses, which are almost as unique and individualized as a person's fingerprints. But they still have group differences, allowing them to be categorized and sorted. In the end, among the 1,100 patients in the study, the AI was able to select with 82% accuracy a group of subjects in whom medics confirmed autism.
Recently, researchers from the Yonsei University College of Medicine in South Korea developed a method to diagnose ASD and the severity of symptoms in children using retinal images verified by an AI algorithm. The results were published in JAMA Network Open.
The researchers recruited 958 participants, with an average age of 7.8 years. They were divided into groups of 479 people, one included children with signs of the disorder, the other - with normal developmental indicators. In total, the scientists obtained 1,890 eye images, that is, 945 in each group.
The severity of symptoms was assessed using the ADOS-2 - the Autism Diagnostic Assessment Scale, which is considered the international "gold standard" of test diagnostics of autism spectrum disorders. Scientists also used the SRS-2 scale, which focuses on social skills: communication, motivation, and social cognition.
A convolutional neural network was trained using 85% of the retinal images and symptom severity test results to build models to identify ASD and symptom severity. The remaining 15% of images were retained for validation.
The AI could select images of children with an AUROC parameter of 1.00 to detect RAS. The AUROC value ranged from one to zero. A model whose predictions are 100% incorrect has an AUROC of 0.0; one whose predictions are 100% correct has an AUROC of 1.0, indicating that the AI's predictions are 100% correct. There was no noticeable decrease in mean AUROC even when 95% of the least important image regions, not including the optic disc, were removed.
"Our models showed promising performance in distinguishing children with RAS from children with typical development. Interestingly, these models maintained an average AUROC value of 1.00 using only 10% of the image containing the optic disc. This indicates that this area is critical for determining the differences between healthy children and those with abnormalities," the authors said.
The average AUROC for symptom severity was 0.74, with an AUROC of 0.7 to 0.8 considered "acceptable" and 0.8 to 0.9 considered "excellent."
"Our results showed that retinal photographs can provide additional information about the severity of RAS symptoms. We observed that feasible classification was only achievable for ADOS-2 scale scores, but not for SRS-2 scores. This may be due to the fact that the ADOS-2 is administered by a trained professional with sufficient time for assessment, whereas the SRS-2 is usually completed by a caregiver in a few minutes. Thus, the former more accurately reflected severity status than the latter," the researchers reported.
According to the authors, the AI-based model can be applied as an objective screening tool for children from the age of four. Since a newborn's retina continues to grow until that age, further research is needed to see if the tool would be accurate for younger children.
Previously, scientists used non-invasive brain imaging to find out why autistic people rarely make eye contact. It turned out that, unlike healthy people, patients with autism spectrum disorders have reduced activity of the dorsal part of the parietal cortex during eye contact, which can serve as a diagnostic marker of ASD.