OR WAIT null SECS
Machine learning algorithms successfully predicted visual and anatomic outcomes and dosing frequency in patients with macular edema secondary to CRVO.
New findings suggest machine learning algorithms had success in the prediction of visual and anatomic outcomes in patients with patients with macular edema secondary to central retinal vein occlusion (ME-CRVO).
It additionally successfully predicted dosing frequency with high accuracy, except for central subfield thickness (CST) in this patient population undergoing treatment with aflibercept
The data were presented at the American Society of Retina Specialists 40th Annual Scientific Meeting.
In an interview with HCPLive, Yasha Modi, MD, Associate Professor, NYU Langone Health, discussed the creation of the machine learning algorithm to make predictions on visual acuity, CST, and treatment frequency using the COPERNICUS and GALILEO studies.
“What made the COPERNICUS and GALILEO study optimal was that the study designs mandated six monthly injections and then they were crossed over into a PRN window,” he said. “That gave us the opportunity to ask if we could predict what would happen to individuals in the PRN window.”
The investigators trained an AI model on 80% of the data, Modi said and inputted data on each patient. This data included baseline demographics, visual acuities, laboratory analyses, and then each of the visual acuity and CST at every monthly visit.
“The algorithm was actually very good at predicting visual acuity,” Modi said. “Not only the absolute visual acuity six months after switching over to PRN treatment, but also predicting change in visual acuity and also reasonably good at predicting change in central subfield thickness, but that was less reliable than visual acuity.”
He noted that retina specialists often do not have good prognostic factors in making the determination between number of injections necessary for patients. He added that the algorithm may have performed better than clinicians in predicting these factors, but the integration of AI is only as good as the individuals who are creating the technology.
“It's important to realize that these are valuable as potential assist tools, but not to supplant clinical decision making,” Modi said. “So we think that this is a good proof of concept to build on and we hope to continue to do more work on this in the future.”
In doing so, the investigators hope to optimize AI algorithms in larger and more diverse data sets to help build more confidence in the reproducibility of the model and introduce it into a clinical setting.
“As that happens, then we start to feel more excited, if you will, about trying to incorporate this into a sort of assist device in the clinical setting,” Modi said.