Artificial Intelligence-Guided Endpoint Selection for Neuroprotection Trials in Glaucoma.

Objective

Standard Automated Perimetry (SAP) is the primary method for monitoring glaucoma progression and an established functional endpoint in clinical trials. A ≥7 dB loss in at least five prespecified test locations has been proposed as a potential endpoint for glaucoma trials, but identifying such vulnerable points in advance remains a major challenge. We developed an artificial intelligence model that uses baseline SAP data to predict the locations most likely to progress, enabling a targeted approach to endpoint selection in neuroprotection trials.

Design

Retrospective cohort study to predict progression.

Subjects

A total of 82,449 SAP tests from 14,237 eyes of 10,533 subjects with open-angle glaucoma across three independent datasets: the Bascom Palmer Ophthalmic Registry, the Duke Glaucoma Registry, and the University of Washington Humphrey Visual Field.

Methods

A graph attention network was trained on Bascom Palmer Ophthalmic Registry data, split at the patient level into development and internal validation datasets, to predict progression risk at each SAP location using baseline total deviation values and anatomically informed graph connectivity. For each eye, the five highest-risk points (High-5) were identified and compared with global mean deviation (MD) and the five lowest-risk points (Low-5). The model was applied without retraining to the Duke Glaucoma Registry and University of Washington Humphrey Visual Field cohorts for external validation.

Main outcome measures

Model ranking performance, neighborhood hit rate, rates of change at High-5, Low-5, and MD, discrimination by area under the receiver operating characteristic curve, and time to repeatable ≥7 dB decline.

Results

In progressing eyes, mean slopes at High-5 were -2.05 to -2.32 dB/y, four to five times steeper than Low-5 (-0.45 to -0.66; P < .001) and approximately twice that of MD (-0.93 to -1.14; P < .001). Area under the receiver operating characteristic curve for discriminating progressors from nonprogressors ranged from 0.883 to 0.937 for High-5, outperforming Low-5 (0.668-0.731; P < .001) and MD (0.871-0.911; P = .003). Nearly all progressors reached the High-5 ≥7 dB threshold during follow-up.

Conclusions

The artificial intelligence-based model identified the most vulnerable visual field locations from baseline data. The High-5 metric provides a proof-of-concept framework consistent with regulatory concepts of functional endpoints and shows improved sensitivity to progression, supporting more efficient neuroprotection trials and personalized glaucoma monitoring.