Four articles. Four minutes. | Issue #25
Artificial Intelligence: Language models disagreed most with specialist consensus in glaucoma suspects and mild disease
Health Systems: Electronically monitored glaucoma drop adherence fell 10 percentage points in the first month after coaching ended
Diagnostics: iCare and Goldmann showed wide limits of agreement despite a small average difference
Medication: Statin use was associated with less recorded treatment escalation in open-angle glaucoma
Patterns and Predictors of Misclassifications by Large Language Models in Multimodal Glaucoma Diagnosis
Methods
This retrospective analysis used clinical data, disc photographs, OCT, and 24-2 fields from 230 eyes of 131 adults in a University of Southern California registry (2016–2022). GPT-5 Pro, Gemini 3 Pro, and 2 glaucoma specialists classified each eye as glaucoma suspect or manifest glaucoma; errors were disagreements with the majority diagnosis of 3 separate specialists.
Results
Disagreement rates were 19.6% for GPT-5 Pro and 16.5% for Gemini 3 Pro; GPT-5 Pro exceeded one comparison specialist's 10.9% rate (P=0.001). Errors clustered near borderline structural findings. Review of 25 shared model errors found overcalling of isolated OCT abnormalities and missed focal structural damage. Disagreement rates fell for all evaluators when analysis was restricted to unanimous reference diagnoses.
Conclusion
Borderline and discordant findings remained difficult for both models, and the study did not establish diagnostic equivalence with specialists.
Our Angle
In clinic, the eyes where AI could be most helpful are also where this study found the most uncertainty: glaucoma suspects and mild disease with discordant OCT, disc, and field findings. Shah and colleagues found shared errors in both directions: overcalling isolated OCT abnormalities and missing focal structural damage. These are familiar clinical traps. I would still review scan quality, myopia and optic nerve anatomy, and the field together, and look for serial structural or functional change before deciding that a borderline eye has manifest glaucoma. Clinical and historical context is equally important for risk stratification. The models were given age and IOP, but not family history or central corneal thickness; in practice, these factors can materially change how I interpret the same borderline OCT or field findings and how closely I monitor the patient.
The 36.1% rate of 2-to-1 reference-panel decisions is itself clinically informative: many cases were genuine diagnostic gray zones, so disagreement with the majority is not the same as an objectively wrong diagnosis. A useful near-term role for these models may be as a second reader that highlights discordance or cases needing expert review. However, this selected cohort required complete, gradable multimodal testing, and the study did not test whether clinicians make better decisions with AI assistance. Before changing treatment or surveillance based on an LLM output, we need prospective studies comparing clinicians with and without AI assistance in more typical real-world patients, including those with incomplete or poor-quality testing.
Health Systems
Learning How to Implement Glaucoma Coaching in Academic Medical Centers from a Successful Clinical Trial using Implementation Science
Design: Mixed-methods implementation analysis within a 2-center randomized clinical trial
Journal: Ophthalmology Glaucoma, October 2026
Authors: Carey et al.
The Support, Educate, Empower (SEE) glaucoma program was delivered consistently at 2 academic centers, but drop adherence declined after coaching ended. This mixed-methods analysis enrolled 236 adults reporting poor adherence after screening 2,057; participants valued the coaches' support, while staff described training, staffing, and workflow needs. Among 76 intervention participants who chose extended monitoring, adherence fell 10.0 percentage points in the first month after coaching ended and another 3.6 points in the second. A reimbursement model projected a $19,272 annual surplus per coach before additional labor and facility costs. Continued coaching was not compared with stopping, and improved IOP control or delivery outside academic centers remains unproven.
Diagnostics
Comparison of Intraocular Pressure Measurements Using iCare Rebound Tonometry and Goldmann Applanation Tonometry: A Systematic Review and Meta-Analysis
Design: Systematic review and meta-analysis
Journal: Journal of Glaucoma, October 2026
Authors: Sachdeva et al.
A small average difference between iCare and Goldmann did not translate into close agreement for individual readings. The review included 33 studies; 31 studies with 5,593 eyes yielded a mean difference of -0.51 mmHg for iCare minus Goldmann. Reported pooled agreement limits were -6.56 to 5.57 mmHg, and differences between iCare models were inconclusive. Only 4 cohorts had mean IOP above 21 mmHg, and the authors rated the evidence very low certainty. Most studies were cross-sectional and used clinic-operated devices, so the review does not establish agreement over time or the performance of iCare HOME.
Medication
Statin use is associated with lower risk of open-angle glaucoma incidence and treatment escalation
Design: Retrospective matched cohort study
Journal: British Journal of Ophthalmology, October 2026
Authors: Lin et al.
Statin users had fewer recorded open-angle glaucoma diagnoses and treatment escalations than matched nonusers in a retrospective TriNetX study. In the incidence analysis of 210,701 patients per group, diagnoses were recorded in 2.3% versus 2.6% within 5 years. A separate analysis of 29,598 patients with glaucoma per group found 5-year MIGS rates of 1.7% versus 3.8%, with fewer selective laser trabeculoplasties, incisional surgeries, and topical prescriptions as well. Associations persisted at 10 years, although results varied by individual statin. IOP, visual fields, optic nerve imaging, reasons for surgery, adherence, and statin dose were unavailable, so the study cannot establish slower progression or a causal benefit. More importantly, follow-up for treatment escalation began at statin initiation for users but at glaucoma diagnosis for controls; excluding users who had already escalated treatment before starting a statin may preferentially select a more stable statin cohort and bias the association toward benefit.
Glaukos reports that adding an intracameral travoprost implant to cataract surgery lowered mean diurnal IOP an additional 3.6 mmHg at three months in a randomized phase 4 trial.
Three patients with open-angle glaucoma had no reported serious adverse events through day 56 after investigational intravitreal epigenetic gene therapy plus oral doxycycline in a phase 1 trial.
A subretinal implant study reports device-assisted vision gains at three years in geographic atrophy, with efficacy assessed in 18 of 38 enrollees without a control group.
A 14-hospital US study found rising ophthalmic transfer and consultation demand, with consultation trends assessed at seven academic institutions and implications for on-call coverage.
The Marcus Gunn pupil is named after Robert Marcus Gunn, but the swinging flashlight test used to detect it was introduced decades later by:
A. Paul Levatin
B. Irene Loewenfeld
C. William Kestenbaum
D. Edmund Landolt
See answer at bottom of newsletter
Get the next issue in your inbox
The Open Angle is a weekly glaucoma research newsletter for busy eye care professionals.
Four articles worth knowing, in under four minutes.
Edited by Jella An, MD, MBA and Jason Dossantos, MD.
The Open Angle is an educational editorial product. It is not medical advice. Readers should review original sources before changing practice.
TRIVIA ANSWER EXPLANATION
The Marcus Gunn pupil is named after Robert Marcus Gunn, but the swinging flashlight test used to detect it was introduced decades later by:
A. Paul Levatin
Levatin introduced the swinging-flashlight method for comparing afferent pupillary responses. Alternating the light between the eyes makes a relative difference in afferent input easier to recognize.
The affected eye may still constrict to light, so the comparison is more informative than asking whether a reaction exists. The test's developer and the clinician named in Marcus Gunn pupil are different people. The eponym identifies the finding, while the later technique provided a convenient way to demonstrate it.