Design: External validation in a prospective population-based cohort
Population: 19,088 eyes of 9,837 adults aged 50–80, Rotterdam Study
Primary outcome: Sensitivity at 95% specificity for perimetric glaucoma
Effect: 70.6% (63.8–76.8); AUC 0.91; 5.1 referrals per case detected

Dutch investigators applied a deep-learning algorithm, trained on more than 100,000 macula-centred fundus photographs, to 19,088 eyes of 9,837 people aged 50–80 in the population-based Rotterdam Study. Perimetric glaucoma was present in 201 eyes (1.1%).

At 95% specificity, sensitivity was 70.6% at eye level, with an AUC of 0.91; the algorithm generated about five referrals per glaucoma case found. True positives had more advanced field loss than missed cases. Among correctly flagged people with earlier photographs, 54% had been flagged four to seven years before diagnosis. Disc-centred photographs performed better than macula-centred ones, and poor optic disc visibility was a major source of false positives.

Opportunistic screening on diabetic retinopathy photographs is the obvious use, including in Indian programmes. The algorithm misses earlier disease and needs good disc images, so it is a triage aid, not a diagnosis.

Ensure the optic disc is fully visible on screening photographs; poor disc views drive false positives
Treat an algorithm flag as a reason for full glaucoma assessment, not a diagnosis
Remember the algorithm misses earlier disease; a negative result does not exclude glaucoma
Consider adding glaucoma triage to existing fundus photography programmes

Why it matters

Shows how an existing photograph could find glaucoma years earlier in people never examined for it.

The statistics, in plain English

An AUC of 0.91 is good discrimination, but because glaucoma affected only about 1% of eyes, most positive flags are false; five referrals per true case is the practical cost. Missing about 30% of cases, mostly earlier ones, limits its use as a rule-out.