The fundus photograph may already contain a glaucoma prognosis

Design

Retrospective cohort with automated deep-learning vasculometry and linear mixed-effects models

Population

1595 eyes (757 POAG, 838 glaucoma suspects) from 929 participants, mean 14.5 visual fields over 10.3 years

Primary outcome

Rate of visual field mean deviation change

Effect

Slope differed by 0.079 dB/year (95% CI 0.034 to 0.123) per SD of venous fractal dimension; mean overall change −0.15 dB/year

Baseline fundus photographs from 1595 eyes — 757 with primary open-angle glaucoma, 838 glaucoma suspects — were processed by an automated deep learning system to extract vessel geometry: fractal dimension, tortuosity, branching and vessel density. Those eyes had a mean of 14.5 visual fields over 10.3 years.

Three features survived model selection. Higher venous fractal dimension, higher arterial fractal dimension and higher arterial tortuosity at baseline were each associated with a less negative visual field slope, the largest being 0.079 dB/year per standard deviation of venous fractal dimension.

Against a mean field loss of −0.15 dB/year, that is about half the average rate of progression — meaningful at the level of a population, and nowhere near enough to tell an individual patient what their next decade holds. The appeal is that it needs no new equipment: the photograph is already being taken.

• Do not change monitoring intervals on vasculometry; it has not been validated for that.
• Keep baseline fundus photography in the glaucoma protocol — it may acquire prognostic value retrospectively.
• Intraocular pressure, field loss at presentation and disc appearance remain the basis of risk assessment.
• Ask whether any automated vessel metric offered to you has been validated outside its development cohort.
• Image quality drives these metrics; a poorly focused photograph will produce a confident wrong number.

Why it matters

It suggests the prognostic information clinicians want may already be sitting in images taken years ago.

Don't overread it

Retrospective single-cohort associations; the authors explicitly call for independent validation before clinical use.

The statistics, in plain English

These are associations between a baseline measure and the slope of subsequent field loss, adjusted for demographic and clinical factors and selected by a backward procedure using the Akaike information criterion — which means the final model was chosen from many, and the confidence intervals do not account for that selection. An effect of 0.079 dB/year per standard deviation is a group-level shift: two patients one standard deviation apart differ, on average, by about half the mean progression rate.

Read the rest in the app

You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.

Get Daily Doctor — free

Scan to keep reading on your phone. No account needed to start.