In a significant stride toward applying precision medicine to treat patients with diabetic macular edema (DME), Duke researchers developed a new algorithm that helps clinicians select the most effective first-line therapies based on patients’ specific disease conditions.

The algorithm is the result of a pilot study conducted through a partnership between the Duke Eye Center and the Pratt School of Engineering; results were published in Biomedical Optics Express in February 2020.

Based on a novel convolutional neural network architecture, the algorithm analyzes a single pre-treatment volumetric scan and accurately predicts whether a patient is likely to respond to anti-vascular endothelial growth factor (anti-VEGF) therapy—without the need for longitudinal treatment information, such as time-series optical coherence tomography (OCT) images, patient records, or other metadata.

The researchers say that knowing in advance whether a patient will respond to anti-VEGF treatment can help avoid unnecessary trial-and-error treatment strategies and ease the heavy treatment burden on patients. Anti-VEGF treatments for aggressive disease usually require eight or nine costly injections over the first year of treatment, presenting a significant issue for people who have other demands on their time.

“Most clinicians agree with what the first-, second-, and third-line treatments are, but there’s a lot of trial and error,” says Michael Allingham, MD, PhD, ophthalmologist and retinal specialist and co-author of the study. “If we could streamline our treatment approach from the outset, it would potentially cut down on some of that treatment burden, or at least help us prepare patients who are going to have a heavier treatment burden in advance.”

The retrospective study involved 127 participants who underwent three intravitreous anti-VEGF injections (ranibizumab 0.3 mg, aflibercept 2.0 mg, or bevacizumab 1.25 mg). Researchers used the algorithm to analyze OCT images taken before the injections, then compared the predictions to images taken afterward to confirm whether the therapy improved a participant’s condition. Based on the results, the team concluded the algorithm would have an 87% chance of correctly predicting which patients would respond to anti-VEGF treatment.

“Although we used a relatively small data set, and the study was technically very challenging, we were able to come up with an algorithm that was quite robust,” Allingham notes. “At Duke, we have very talented people with diverse experiences and areas of expertise, and we can put our heads together to attack a problem that's clinically important, like this one.” The researchers plan to confirm and extend the findings from this pilot study by performing a larger observational trial of patients who have not yet undergone treatment.

“This research represents a step toward precision medicine, in which such predictions help clinicians better select first-line therapies for patients based on specific disease conditions and adapt optimized dosage in followup visits," says Sina Farsiu, PhD, director of Duke’s Vision and Image Processing Laboratory and lead researcher/lead author of the study. "Our approach could potentially be used in eye clinics to prevent unnecessary and costly trial-and-error treatments and thus alleviate a substantial treatment burden for patients. While this first application was targeted at diseases of the eye, the algorithm could also be adapted to predict therapy response for many other types of systemic or organ-specific diseases."

“This type of approach can help us answer the question of how we can more efficiently make sure that all of our patients with diabetes are getting screened, and that we’re not missing people with ocular pathology,” says Allingham. “My hope is that this is the start of something larger and more impactful to help determine care and make a difference in our clinics. The real work lies ahead.”