AI models show promise but face significant hurdles in predicting glaucoma progression, review finds

Artificial intelligence (AI) tools can predict glaucoma progression with moderate to good accuracy, but widespread clinical adoption remains limited by inconsistent study methods, poor transparency, and a lack of external validation, according to a new systematic review.

The review, which analyzed 46 reports from 43 unique studies, is among the first to map the landscape of AI-based glaucoma progression prediction with an eye toward what would be required for real-world clinical use.

“Similar to explorations of AI applications in other fields, AI studies for glaucoma predictions have seen significant growth in research attention and output in recent years, albeit with high heterogeneity and inconsistency in their approaches,” study authors, led by Yichuan G. Liang, Faculty of Medicine and Health at the University of Sydney in Sydney, NSW, Australia, stated. “It is therefore of interest to researchers, clinicians and developers to systematically map out the current landscape of glaucoma progression forecasting with AI in order to understand the heterogeneous design and outcomes in current approaches and identify key challenges and strategies towards clinical implementation. Although previous analyses explored major trends and limitations in this field, several key issues remain poorly addressed, such as interpreting AI model outputs, integrating AI into existing clinical pathways and consistent reporting items for future studies. Our review placed particular emphasis on these topics which are critical to the robust evaluation and safe implementation of AI tools.”

Why it matters

Glaucoma is the leading cause of irreversible blindness worldwide, with an estimated disease burden projected to exceed 111 million people by 2040. Its progression is highly variable — recent estimates suggest 1 in 8 eyes with glaucoma undergoing routine treatment will exhibit rapid progression beyond negative 1.0 decibels per year. No clinically applicable tool currently exists to guide personalized treatment based on forecasted progression risk.

“Early and accurate prediction of progression is critical to the maximal preservation of vision, as it could guide more intensive monitoring and treatment for high-risk patients,” the study authors stated.

What the studies found

Researchers screened more than 4,100 records from five databases — Embase, Web of Science, MEDLINE, arXiv, and Cochrane CENTRAL — ultimately including 43 unique study cohorts covering more than 202,207 subjects across six countries. The United States accounted for 28 of the 43 studies. All studies were retrospective. Thirty-six of the 46 articles were published within the last 4 years.

AI models were trained on three types of tasks. Binary classification models predicted whether a patient would progress or convert to glaucoma; numeric regression models predicted future values of clinical measurements such as visual field indices and retinal nerve fiber layer thickness; and survival models estimated the probability of progression at different time points.

For binary classification, models predicting conversion from ocular hypertension or glaucoma suspect status to diagnosed glaucoma achieved area under the curve scores ranging from 0.62 to 0.99. Models predicting progression in already-diagnosed glaucoma patients using biological criteria achieved AUC ranges of 0.68 to 0.93, while those using clinical event-based criteria such as escalation to surgery ranged from 0.65 to 0.99.

For numeric regression, predictive errors for visual field mean deviation ranged from 1.06 to 3.12 decibels for mean absolute error and 1.0 to 2.71 decibels for root mean square error. The review noted that errors increased with longer prediction timelines and more severe disease.

One study evaluated the zero-shot glaucoma predictive performance of ChatGPT 4.0 using baseline demographic, clinical, and visual field inputs. It achieved a sensitivity of 0.56 and specificity of 0.78 for predicting conversion from ocular hypertension to glaucoma at 1 year.

Key limitations found across studies

The review identified several recurring problems.

Only seven of the 43 studies validated their AI models on external datasets — data not used during training. One study that did conduct external validation saw its AUC drop from 0.67 on internal data to 0.49 on external data, a result the review authors describe as illustrating the risks of relying on internal validation alone.

Only 42% of studies received low risk of bias ratings across all assessment domains using the QUADAS-2 evaluation tool. The most common problems were in patient selection, where studies frequently used non-consecutive or non-randomized recruitment methods, and in the handling of missing data.

Twenty-one percent of studies provided open-source access to their AI code, and only 5% made their training and testing data publicly available. Forty percent of studies reported race and ethnicity data for their patient cohorts, despite evidence that AI models have shown race- and ethnicity-associated differences in glaucoma predictive performance.

The review also found that most studies did not report clinically actionable performance metrics. Using figures from one included study as an example, the authors calculated that a model with a reported sensitivity of 0.95 and specificity of 0.74 would yield a positive predictive value of only 18.4% when applied to a population with a rapid progressor prevalence of 5.8% — meaning the model would produce a high rate of false positives in typical clinical use.

None of the included studies explicitly modeled the effects of treatment, which the review authors identify as a significant limitation. Patients in the datasets had received varying interventions, but treatment history and outcomes were generally unreported, meaning models implicitly assumed either no treatment or uniform treatment responses.

Recommendations

The authors propose what they describe as the first glaucoma-specific list of recommended practices and reporting items for future studies. These include adopting consecutive or randomized patient recruitment, reporting comprehensive cohort characteristics including race and ethnicity, performing external and ideally multi-center validation, reporting clinically contextual metrics such as positive and negative predictive values, and disclosing error distributions rather than summary statistics alone.

The authors also recommend that future studies explore AI techniques capable of accounting for treatment effects as variables, and suggest approaches such as shadow deployment, recurring local validation, and human-in-the-loop clinical pathways as strategies for building clinician trust ahead of broader implementation.

Limitations of the review itself

The review was restricted to English-language publications, excluded conference abstracts, and did not perform a quantitative meta-analysis due to the degree of between-study heterogeneity. The authors acknowledge that publication bias may make the overall picture of AI performance appear more favorable than it is.

“Several challenges for clinical translation remain, including inconsistent reporting, limitations and heterogeneity in study design and poor AI generalisability and transparency. We encourage future studies to adopt robust study design and transparent reporting and propose the first glaucoma-specific list of recommended practices and reporting items for future clinical implementation,” the study authors stated.

Reference:

1. Liang YG, Fan L, Teixeira-Pinto A, Liew G, White AJR. A systematic review of AI for predicting glaucoma progression: challenges and recommendations towards clinical implementation. npj Digital Medicine. 2026;9:140. https://doi.org/10.1038/s41746-025-02321-7