Machine Learning Could Change Understanding of AMD Progression, Severity

Ryan Black

Algorithm able to correctly predict drusen progression with encouraging accuracy.

Speaking in a press conference at the Annual Meeting of the Association for Research in Vision and Ophthalmology in Baltimore, Maryland, Hrvoje Bogunovic, PhD, of the University of Vienna presented potentially-massive findings regarding age-related macular degeneration (AMD) prediction in a calm, understated manner.

Bogunovic is a computer scientist by trade, and his team’s work involves using machine learning to detect patterns in drusen progression in order to predict who will develop severe AMD, when they will develop it, and where in the retina it will occur.

Using optical coherence tomography (OCT) imaging in eyes from patients originally enrolled in the HARBOR ranibizumab trial, the group was able to identify drusen in the eyes at baseline and months 1 through 4. Drusen size and number are a hallmark of the disease, and drusen regression in intermediate AMD is an indicator of increasing disease severity, though it is “extremely difficult to know…in which eyes such drusen regression will appear,” according to Bogunovic, noting that all progression is individual.

The team developed “a machine learning algorithm predicting the conversion to advanced AMD on an individual basis, using the extracted imaging biomarkers as well as known genetic risk factors of AMD (34 single-nucleotide polymorphisms) as input features,” according to the study. The biomarkers that they found most important were drusen size, drusen OCT appearance, and change of drusen size from baseline, while age itself was not nearly as predictive.

The study featured OCT images of retinas from 38 patients and applied their algorithm. Bogunovic showed examples from the algorithm’s predictive imaging system, which was able to simulate an OCT scan at 1 year from baseline. The generated mapping images produced were about 74% consistent with the actual scans at 1 year.

Speaking of the work, Bogunovic said the impact was potentially twofold. For patients, “By being able to predict the development of their disease we’ll be able to adjust their scheduling regimen on an individual basis, it will also allow us to potentially catch the onset of severe AMD very early on, which is extremely important in order to save their vision.” And for pharmaceutical companies, he says, the ability to more accurately identify those at risk for severe AMD will help them design their clinical trials, “Potentially by having a suitable amount of subjects it will make the trials more precise and shorter in duration.”

Important to understand in the long run, Bogunovic says, is how the accuracy of this system relates to that of current methods being used in treatment. “Every doctor has their own sort of idea of how to advise patients,” he said, but in the long run the goal could be to create an open cloud-based system to input progression results in order to eventually sharpen the algorithm’s predictive prowess. It also doesn’t necessarily end with AMD.

“Nevertheless, the entire sort of methodology that we are developing in terms of the image analysis tools are applicable to other diseases of the retina, so for example we are trying to apply similar models to understand progression of diabetic retinopathy, particularly diabetic macular edema,” he said.