Predicting Visual Outcomes and Visual Improvement in Macular Edema Due to CRVO

Delayed treatment from diagnosis to first intravitreal aflibercept (Eylea) injection results in less visual improvement, and a machine learning algorithm can successfully predict visual and anatomic outcomes, according to presentations at the American Society of Retina Specialists 40th Annual Scientific Meeting, held July 13-16, in New York, New York.

Dilsher S. Dhoot, MD, an ophthalmologist with California Retina Consultants, presents the results of a post hoc analysis of patients with macular edema due to central retinal vein occlusion in the phase 3 COPERNICUS and GALILEO trials for aflibercept. The goal was to evaluate the impact of time since diagnosis to treatment and baseline best-corrected visual acuity (BCVA) on visual and anatomic outcomes.

“Understanding those factors that may affect visual and anatomic outcomes can inform treatment decisions and manage physician and patient expectations,” he said.

In the 2 trials, patients were randomly assigned 3:2 to either receive aflibercept 2 mg every 4 weeks or a sham control. In both trials patients were eventually switched to pro re nata (PRN), or as needed, dosing. The researchers analyzed patients based on the time elapsed between diagnosis and the first treatment: < 1 month, 1-3 months, > 3 months. In both trials, patients who received treatment more than 3 months after their diagnosis had the worst vision compared with patients who received treatment less than 1 month after diagnosis. However, the results of optical coherence tomography angiography were similar among the 3 groups.

“So, despite these anatomic improvements that are similar, the visual acuity gains are different based on time to diagnosis,” Dhoot explained.

When evaluating patients based on BCVA, they found patients with the worst vision had the greatest improvement in vision, although their absolute vision was less than patients who started with the best visual acuity. The results were similar across COPERNICUS and GALILEO.

“We hope that these findings will add to the body of literature that helps inform our treatment decisions in those patients with macular edema from central retinal vein occlusion,” he concluded.

When the trials switched patients to PRN treatment, the criteria follows expected guidelines of increases in central subfield thickness (CST), new fluid loss, or gain of 5 letters of visual acuity, noted Yasha Modi, MD, assistant professor of ophthalmology at New York University. However, the data are based on a population, and outcomes can look very different on a patient-by-patient basis.

He highlighted 3 separate patients with similar BCVA and CST with drastically different number of PRN injections and drastically different visual acuity outcomes.

“This begs the question: can we predict whether patients will require frequent PRN treatment? And what would their visual and anatomic outcomes potentially look like?” he wondered.

He presented the findings of a study on a machine learning (ML) model using the COPERNICUS and GALILEO data to predict visual and anatomic outcomes. Modi and his colleagues used random forest trees on 80% of the data from the trials and used the remaining 20 of the data to validate predict outcomes and determine the performance of the ML model.

The researchers found reasonable correlation between the model’s prediction of visual acuity at week 52 and what actually happened. There were similar findings looking at the change in BCVA. Modi noted that key predictive factors for change in BCVA were BCVA at baseline, week 20, and week 24.

The algorithm “starts to fall apart” when looking at CST because it is not able to accurately predict it, but it is able to predict change in CST. The key predictors for change in CST were CST at baseline and BCVA at baseline.

Another area they looked at were injection frequency. “This is a question all of our patients have, inquiring ‘how many injections am I going to require?’” Modi reported that the model was able to predict with reasonable accuracy those patients who are going to have 2 or less injections between weeks 24 and 48.

“A machine learning model, which is on a very small amount of data, can predict visual outcomes, changes in macular thickness, and dosing frequency with reasonably high accuracy,” he concluded. “Now, of course, this is a proof of concept. It needs to be tested in larger number of patients.”