Three-Variable Score Flags Which Fundus Images Need Urgent Eye Review

A simple score built from three routine clinical measurements—age, systolic blood pressure, and fasting plasma glucose—can identify which fundus images are most likely to contain findings that warrant referral to an eye specialist, according to a study published in the European Journal of Medical Research.

The motivation is a practical bottleneck. When fundus photographs are taken during routine health examinations, limited ophthalmologist reading capacity can delay interpretation of the acquired images. A score that estimates the probability of referral-warranted findings from data already collected at the health check could allow reading queues to be prioritized, so that image sets with the highest likelihood of clinically important findings move to the front of the line while lower-probability sets are interpreted later.

The study included 19,221 participants, with a mean age of 58.9 years and 13,723 participants (71.4%) men. Among them, 2,709 (14.1%) had referral-warranted findings on interpretation of their fundus images.

The three predictors each carried independent information. In the final model, the adjusted odds ratio was 1.31 per 5 years of age (95% CI, 1.28–1.34), 1.05 per 10 mmHg increase in systolic blood pressure (95% CI, 1.02–1.08), and 1.12 per 1 mmol/L increase in fasting plasma glucose (95% CI, 1.08–1.16). In other words, older age, higher systolic blood pressure, and higher fasting plasma glucose were each associated with higher odds of referral-warranted findings, and the model combined these three variables into a single probability estimate.

The bootstrap-corrected C-statistic was 0.7478 (95% CI, 0.739–0.757), the Brier score was 0.1070 (95% CI, 0.104–0.110), and the calibration slope was 0.992 (95% CI, 0.965–1.020). These figures indicate moderate discrimination—the model’s ability to separate image sets with and without referral-warranted findings—and good internal calibration, meaning the predicted probabilities aligned well with observed event rates in the development data.

The model also held up in a temporal test. In that temporal-validation sample, the C-statistic was 0.744 (95% CI, 0.723–0.765), the Brier score was 0.110 (95% CI, 0.105–0.115), the calibration intercept was 0.036, and the calibration slope was 0.957—figures closely matching the internal-validation results.

To show how the score could translate into workflow decisions, the researchers grouped participants by predicted probability after the model was fixed. In post hoc workload groups with predicted probabilities below 10%, 10% to below 18%, and 18% or above, the observed proportions of referral-warranted findings were 5.6%, 14.0%, and 29.1%, respectively. This stepwise gradient means the lowest-risk group carried roughly one-fifth the event burden of the highest-risk group, providing a concrete basis for deciding which image sets should be read first.

Notably, although BMI improved likelihood-based fit, its inclusion did not change the C-statistic or Brier score to three decimal places, supporting the choice of the leaner three-variable model. The integer-score version, which rounds the model into a point-based tool for bedside use, had a corrected C-statistic of 0.7466 and a calibration slope of 0.984, essentially matching the full model.

The authors conclude that age, systolic blood pressure, and fasting plasma glucose stratified interpretable fundus image sets by the probability of referral-warranted findings, with moderate discrimination and good internal calibration. They state that the score merits evaluation for prioritizing interpretation of acquired images, but they are explicit that external validation and prospective workflow assessment are required before deployment.

Because the score relies entirely on data collected routinely at health examinations—age, a blood pressure reading, and a fasting glucose value—it could be computed immediately after image acquisition without additional testing or specialist input. That makes it a candidate for triage logic in screening programs where reading capacity, not image quality, is the limiting factor. The next steps the authors identify—external validation in other settings and prospective assessment of how the score performs inside an actual reading workflow—will determine whether this three-number rule can move from retrospective modeling into operational use for prioritizing fundus image interpretation.

Subject of Research: Development and internal validation of a three-variable clinical score to prioritize interpretation of fundus images for referral-warranted findings in adults aged 50 or older

Article Title: Development and internal validation of a three-variable score to prioritize interpretation of acquired fundus images for referral-warranted findings in adults aged 50 years or older

Article References: Zhang, X., Ma, Y., Niu, Y., Xiong, D., Wu, C., Lai, X., Cai, X., & Yang, S. (2026). Development and internal validation of a three-variable score to prioritize interpretation of acquired fundus images for referral-warranted findings in adults aged 50 years or older. European Journal of Medical Research. https://doi.org/10.1186/s40001-026-05306-8

Image Credits: AI Generated

DOI: 10.1186/s40001-026-05306-8

Keywords: fundus imaging, prediction model, referral-warranted findings, logistic regression, internal validation, ophthalmology, health screening, triage score, systolic blood pressure, fasting plasma glucose, bootstrap validation, temporal validation