Christopher A. Clark
ORLANDO, Fla. – Machine learning may have the potential to detect changes in the retinal structure before a patient with diabetes develops diabetic retinopathy, according to a study presented here at the American Academy of Optometry.
Christopher A. Clark, OD, PhD , presented a study here at an academy-sponsored press conference that used 8,760 spectral domain OCT images of diagnosed diabetics with a wide range of HA1c values but without retinopathy to train and test a machine learning model.
“We already know that in subclinical levels there are vascular changes much earlier on, before they develop diabetic retinopathy,” Clark said.
According to the abstract, 2,920 images were reserved for testing the model, and the remaining 5,840 were used to train it. Five hundred features were used at each HA1c level to train and test 18 models.
The model improved in its ability to classify into higher vs. lower groups in the absence of retinopathy as HA1c levels increased, according to the study. The model that took advantage of multiple groups of features was the best one, detecting the presence or absence and spacing of capillaries, the researchers said.
“The ability of the model to predict at all levels, including relatively low HA1c ones, suggests that there are other retinal features occurring at subclinical levels,” according to the abstract.
Clark said at the press conference that the potential exists to identify future patients at highest risk for developing retinopathy, reducing visits for diabetics with otherwise healthy retinas and detecting diabetes in patients currently not diagnosed with it. – by Nancy Hemphill, ELS, FAAO
Clark CA, et al. Detecting retinal changes using machine learning in diabetes without retinopathy. Presented at: American Academy of Optometry meeting; Orlando, Fla.; Oct. 23-27, 2019.
Disclosure: Clark reported no relevant financial disclosures.