Discover how AI technology is revolutionizing the diagnosis of autism through retinal photographs, achieving 100% sensitivity and specificity. Dr. Michael Fulks explores groundbreaking technology in this informative blog.
I have been reviewing the promising technology of reading ECGs by artificial intelligence (AI) when I came across this amazing article by Jae Han Kim where AI was used to diagnose autism (better labeled autism spectrum disorder or ASD), with 100% specificity and 100% sensitivity just by looking at a retinal photograph. From my experience doing multivariate analysis of structured data (test results), such a combination of sensitivity and specificity is near impossible, and the association of ASD and retinal photos was non-intuitive to me, so I dug into this further.
The optic nerve is really part of the CNS rather than a peripheral nerve and diffuse changes to the brain may be reflected in it as seen on imaging the retina. Building on other earlier studies (see article references), Kim’s group looked at 4- to 19-year-olds diagnosed with ASD who had retinal photos taken and those without any brain diseases who previously had retinal photos taken utilizing deep machine learning. In a large test cohort separate from the training cohort, the AI evaluation proved both 100% sensitive and specific for ASD as well as having moderate (AUC .74) predictive ability for severity of the ASD. Heat mapping showed that the crucial information was derived from the area of the optic disk and (based on other studies) this may relate in part to reduced nerve fiber layer thickness in the ASD.
The authors acknowledge that whether this would work as well at younger ages where the diagnosis of ASD is more often uncertain but more active development of the eye structures is ongoing or whether images could routinely be taken at ages younger than 4, or how other neurologic conditions might confuse the result is unknown. But for even a skeptic like me, this is another example of just how powerful AI can be in using unstructured data such as retinal images or ECG recordings in assigning a correct diagnosis/outcome. AI can have a much higher accuracy (plus being faster and cheaper) than human experts can ever achieve but with the limitation that how it makes these decisions may be near impossible to tease out.
About the Author
Michael Fulks, MD, Consulting Medical Director, is board-certified in internal and insurance medicine. After leaving practice, he served as a medical director, creating or editing several underwriting manuals and preferred programs. More recently, Mike has consulted for CRL participating in its mortality research on laboratory test results, BP and build, and in the development of risk-scoring tools for laboratory and non-laboratory data.