Key Takeaways

  • G-PROG uses a single baseline color fundus photograph to predict structural glaucoma progression over the following 2 to 5 years
  • The retrospective study analyzed 161,827 fundus images from 13,913 patients and externally tested the model across 5 international datasets, where maximum AUCs ranged from 0.74 to 0.86
  • Researchers said prospective evaluation is still required to determine whether the model can improve glaucoma risk stratification, monitoring, and allocation of clinical resources

A deep learning model could help clinicians identify patients at increased risk of glaucoma progression using a single baseline color fundus photograph, according to findings from a large, international retrospective study published in The Lancet Digital Health.1

The model, known as G-PROG, was developed to predict structural glaucoma progression over 2 to 5 years. Researchers said earlier identification of patients likely to experience faster progression could eventually help clinicians tailor monitoring and treatment intensity according to individual risk.

The study was led by Ruben Hemelings and colleagues and included investigators from centers in Europe, Singapore, and China. The analysis included 161,827 fundus images obtained during 127,962 visits involving 13,913 patients, representing 128,021 eye-years of follow-up.

Determining the rate at which an individual patient's glaucoma will progress generally requires longitudinal follow-up. G-PROG was designed to approach the problem differently by extracting prognostic information from the appearance of the eye at baseline. The researchers trained and internally validated G-PROG using data from UZ Leuven in Belgium. Independent datasets from glaucoma centers in Brussels and Liège, Belgium; Tampere, Finland; Mainz, Germany; and Hangzhou, China were then used for external testing.

G-PROG uses color fundus photographs to predict the trajectory of structural glaucomatous damage. Progression in the study was defined using the slope generated by G-RISK, a previously validated deep learning model that quantifies glaucomatous optic nerve damage from color fundus photographs. The investigators calculated the slope using longitudinal G-RISK predictions within individual eyes over follow-up periods ranging from 2 to 5 years.

The team evaluated 20 configurations of G-PROG, varying criteria including the number of available visits, image quality, intervals between visits, and baseline G-RISK values. Model performance was assessed using area under the receiver operating characteristic curve (AUC), coefficient of determination, and explained variance score.

Eighteen of the 20 model configurations produced statistically significant AUC results, according to the study.

In internal validation, G-PROG achieved a maximum AUC of 0.98 (95% CI, 0.97-1.00) across the 2- to 5-year follow-up intervals. Among glaucomatous eyes with a baseline G-RISK score greater than 0.6, the maximum AUC was 0.92 (95% CI, 0.85-0.98).

For external validation, investigators averaged predictions from the 8 best-performing model configurations. Maximum AUCs across the 5 independent test datasets ranged from 0.74 to 0.86, suggesting that the model retained predictive ability when applied to populations outside its development center. Researchers also compared the G-RISK slope used to define progression with established clinical measures. The measure showed agreement with visual field mean deviation slope and average retinal nerve fiber layer thickness slope, reaching maximum AUCs of 0.82 and 1.00, respectively.

The ability to estimate future progression from a baseline fundus photograph could have implications for how glaucoma services allocate monitoring and treatment resources, the study authors said.

In describing the research, Mr. Hemelings noted that clinicians can assess factors such as age, intraocular pressure, and existing damage at an initial visit, but determining a patient's actual rate of progression currently requires follow-up over time. The goal of G-PROG is to estimate that future trajectory from an image available at baseline.

If prospective studies confirm its clinical utility, such an approach could potentially help identify higher-risk patients for closer monitoring or earlier intervention while allowing lower-risk patients to follow less intensive surveillance schedules. That application, however, was not tested in the retrospective study.

The investigators emphasized that prospective evaluation is needed to establish whether G-PROG can meaningfully improve clinical risk stratification and resource allocation in glaucoma care.

Reference

1. Hemelings R, et al. Prediction of structural glaucoma progression from baseline fundus photographs using deep learning: a retrospective multicentre study. Lancet Digit Health. 2026;8(8):101032. doi:10.1016/j.landig.2026.101032.