A new deep learning model claims it can spot which glaucoma patients are heading toward rapid vision loss years before doctors could confirm it with the current standard test.
Researchers describe the system, called GLAM (Glaucoma Longitudinal Analysis Model), in an arXiv preprint posted September 25, 2026 (arXiv:2609.29256, https://arxiv.org/abs/2609.29256). GLAM uses attention-based fusion to read sequences of Humphrey 24-2 visual field tests - the same scans clinics already run - plus five clinical features, and estimates how quickly a patient's mean deviation score will worsen. Tested on 4,276 patient-eyes from the open-access University of Washington Humphrey Visual Field dataset, it predicted progression rate with a mean error of 0.139 dB per year and an R^2 of 0.927, a 73.5% reduction in error compared to a standard ridge-regression baseline. It flagged fast progressors with an AUC of 0.990.
That accuracy matters because glaucoma is the leading cause of irreversible blindness, and the current method - ordinary least-squares regression on repeated visual field tests - typically needs six to ten scans spread over several years before a slope estimate is trustworthy. A model that reaches similar or better confidence from fewer visits, using data clinics already collect, could mean earlier treatment decisions for the patients most at risk of losing vision fast.
Worth noting: this is one retrospective analysis on one public dataset, not a clinical trial, and the paper is not yet peer reviewed. Regression-based glaucoma prediction has had strong retrospective numbers before without changing how ophthalmologists actually manage patients - the harder test is whether GLAM holds up across different clinics, scanners, and populations, and whether regulators and doctors trust a black-box slope estimate over a line on a chart.