Truace
Subscribe

AI spotted 70.6% of glaucoma eyes at 95% specificity in a 9,837-person Dutch study

Study title: External validation of a deep-learning algorithm for perimetric glaucoma detection on color fundus photographs in the Rotterdam Study

Researchers tested a deep-learning algorithm that won the AIROGS glaucoma screening challenge on a new population: 19,088 eyes from 9,837 Rotterdam Study participants aged 50-80. Of those eyes, 201 (1.1%) had perimetric glaucoma. At 95% specificity, the algorithm detected 70.6% of affected eyes, with an ROC-AUC of 0.911.

Read the study abstract

Purpose To externally validate a deep-learning (DL) algorithm detecting perimetric glaucoma on color fundus photographs (CFPs) in a population-based cohort. Design External validation in a prospective population-based cohort. Participants Participants aged 50-80 years from three Rotterdam Study cohorts. Methods We applied the winning Artificial Intelligence for Robust Glaucoma Screening (AIROGS) challenge algorithm, trained on 101,442 macula-centered CFPs to detect referable presumed perimetric glaucoma. Perform…

Truvace record2 min read1 sourceEvidence: Moderate · 70/100 — Controlled or validated comparison · n = 9,837 · 1 independent sourcePeer-reviewedTRV-2026-2390 · cite or check it
External validation of a deep-learning algorithm for perimetric glaucoma detection on color fundus photographs in the Rotterdam Study"Medical Laboratory" by ben.dracup, CC BY 2.0.
How the two sides scoreContested
P 57P is the problem side, out of 100. It goes up when the problem is specific, when it measurably hurt people and when the sources behind it are strong.

A fixed formula sets both scores from the evidence and its sources. Reader votes don’t move them.

G 59G is the gain side, out of 100. It goes up when the gain is specific, when it measurably helped people and when the sources behind it are strong.

In brief

At participant level, sensitivity was 68.8% and ROC-AUC was 0.909. The algorithm had been trained on 101,442 macula-centered photographs, yet it performed no worse on disc-centered images. Correctly identified cases were often flagged years before the perimetric diagnosis.

The screening cost is substantial: 5.1 referrals for each detected case, with a precision-recall AUC of 0.343 because glaucoma is rare in this population. The authors conclude that fundus-photo deep learning could serve as a scalable triage tool, and that image acquisition standards matter for real-world use.

Main points

  1. The validation covered 19,088 eyes from 9,837 participants in three Rotterdam Study cohorts, all aged 50-80.
  2. Only 201 eyes (1.1%) had perimetric glaucoma, so positive predictive value was limited, with a PR-AUC of 0.343.
  3. Eye-level sensitivity at 95% specificity was 70.6% (95% CI 63.8-76.8); participant-level sensitivity was 68.8%.
  4. True positives had more advanced visual field damage than false negatives (median mean deviation -10.8 vs -5.8 dB).
  5. The algorithm showed no loss of performance on disc-centered photographs despite training only on macula-centered ones.
  6. Incomplete optic nerve head visibility was an important source of false positives.
  7. The overall screening yield was 5.1 referrals per detected glaucoma case.

The gain

In a population study of 19,088 eyes, a deep-learning algorithm detected 70.6% of eyes with perimetric glaucoma at 95% specificity, often flagging cases years before diagnosis.

The problem

The same algorithm needed 5.1 referrals per detected glaucoma case, missed less advanced cases, and produced false positives when the optic nerve head was only partly visible.

The full story

The study applied the AIROGS challenge-winning algorithm, built to detect referable presumed perimetric glaucoma, to participants in a prospective population-based cohort. Performance was assessed at eye and participant level with cluster bootstrap confidence intervals, and misclassified cases were reviewed qualitatively.

Results showed ROC-AUC values near 0.91 at both levels, but sensitivity of about 70% at 95% specificity. Cases the algorithm caught had worse visual field loss (median mean deviation -10.8 dB) than those it missed (-5.8 dB), so milder disease was more likely to slip through.

Further analyses compared macula- and disc-centered photographs in pairs, tracked detection over time using Kaplan-Meier analysis, and examined how image quality affected positive predictive value. The authors say their findings point to image acquisition standards as key to implementation.

What this doesn’t fix

Glaucoma was rare in the cohort (1.1% of eyes), giving a modest PR-AUC of 0.343, and incomplete optic nerve head visibility caused many false positives.

Sources

  1. Peer-reviewedOphthalmology2026-11-02

The debate