AI-Based Strategies for Predicting Glaucoma Progression

Big data and artificial intelligence (AI) hold strong potential for improving glaucoma patient management by enabling individualised assessment and prediction of disease progression. However, validating these AI-based prognostic approaches will be essential before they are ready for clinical implementation, according to Cecilia S Lee MD.

“A major challenge that we face in glaucoma management today is that progression is highly variable,” Dr Lee explained. “Therefore, we can miss the rapid progressors, and surgery is done only because there is evidence of progression versus prevention.”

Interested in addressing this challenge, Dr Lee and the research team worked on training deep learning networks to forecast future 24-2 Humphrey Visual Fields (HVFs).¹ They extracted about 1.7 million perimetry points from more than 32,000 HVFs collected over 20 years with the aim of predicting a future HVF from a single baseline test. This research used time as the undeniable ground truth, meaning the training relied on a single visual field as the input and a future visual field at a specified time as the output.

“We were surprised by how well the models performed,” Dr Lee said. “What was very interesting was the study showed deep learning can learn structural spatial patterns that are carried when glaucoma progresses.”

As a next step, Dr Lee and colleagues developed a policy-driven multimodal model that reflects clinical practice by fusing information from disc photos and OCT to predict the visual field.²

They found that when the two modalities were fused and the strategy was able to decide which modality was more appropriate for its contribution to visual field prediction, the model performed better than when it relied on a single input.

The team is also implementing big data and AI to discover new modifiable factors for glaucoma development and gain new insights into which patients will experience glaucoma progression and why individuals progress the way they do—an exciting area in which various systemic biomarkers and environmental factors have already been reported.

Dr Lee said these initiatives require access to large, longitudinal multimodal data sets to train predictive models. In this realm, Dr Lee and colleagues are working on AI-READI (Artificial Intelligence Ready and Exploratory Atlas for Diabetes Insight), a US National Institutes of Health Bridge2AI programme that is recruiting about 4,000 participants and gathering a broad spectrum of data to optimise for AI/machine learning analysis. Included in the data set are OCT and OCT angiography volumes, macular and optic disc imaging, glucose measurements, and other health metrics as well as sleep and stress measurements and environmental data.

Dr Lee presented at the 2026 ASCRS meeting in Washington, DC.

Cecilia S Lee MD, MS is the Jane Hardesty Poole Endowed Professor in the Department of Ophthalmology and Visual Sciences, Washington University, St Louis, Missouri, US. She is also Program Director of the AI-READI project. cslee@wustl.edu

1. Wen JC, et al. PLoS One, 2019 Apr 5; 14(4): e0214875.

2. Kihara Y, et al. Ophthalmology, 2022; 129(7): 781–791.