Researchers at Yonsei University in Seoul gathered photos of the retinas, the light-sensing tissue at the back of the eye, of 958 children, about half with autism spectrum disorder and half without, then trained a deep learning model on the images.
Tested on photos it had not seen, the model distinguished children with autism from those without with an area under the curve of 1.00, a perfect score on that statistical measure, and kept that score using only the small image patch containing the optic disc.
A perfect score on one dataset invites caution rather than celebration: the children were closely matched by age and sex, and more varied outside populations have not yet been tested. If the result holds up elsewhere, a retinal photo could flag autism risk well before a behavioral diagnosis is typically made.
Read the original study
Kim et al., JAMA Network Open, 2023 · doi.org/10.1001/jamanetworkopen.2023.47692
A retina photo carried a signature of behavior.
Source
- Kim, J. H., Hong, J., Choi, H., Kang, H. G., Yoon, S., Hwang, J. Y., Park, Y. R., & Cheon, K.-A. (2023). Development of Deep Ensembles to Screen for Autism and Symptom Severity Using Retinal Photographs. JAMA Network Open, 6(12), e2347692. https://doi.org/10.1001/jamanetworkopen.2023.47692