We can use machine learning to detect changes in the retinal structure before the development of retinopathy, according to Christopher Clark, OD, PhD, who presented his recent research on the topic at the American Academy of Optometry meeting in Orlando. He and his team wanted to see if they could use machine learning to correlate changes in HbA1c levels with changes in retinal structure that are not detectable through current means such as traditional OCT and fundus photography.

The researchers determined that the fine k-nearest neighbors algorithm (KNN) model was the best at differentiating SD-OCT images by HbA1c at the time of a patient’s screening visit. The fine KNN model could detect unique structural changes in the SD-OCT image regardless of the HbA1c level, even at HbA1c levels of six.

The researchers trained and tested the machine learning model with 8,760 SD-OCT images from screening centers—5,840 images for testing and 2,920 for training. All images came from patients diagnosed with diabetes with a wide range of HbA1c but without retinopathy. The entire data set was analyzed from five to 10 HbA1c in 1% intervals (six levels in total). With increasing HbA1c levels, the KNN model improved its ability to classify into higher vs. lower groups even in the absence of retinopathy.

The artificial intelligence’s accuracy increased with increasing levels of HbA1c and was most accurate at a HbA1c level of six when using the testing image set. The model correctly predicted 1,099 above six HbA1c and 1,049 below. It categorized 771 images in the wrong level.

During is Academy lecture, Dr. Clark concluded that the model detects the presence or absence as well as spacing of capillaries. The model’s ability to predict changes at all levels, including relatively low HbA1c, suggests that there are other retinal changes occurring at subclinical levels. Dr. Clark hopes that using the model could potentially reduce unnecessary visits for diabetics with otherwise healthy retinas and detect systemic diabetes in patients currently not diagnosed with diabetes.