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

Limited ophthalmologist reading capacity can delay interpretation of fundus images acquired during routine health examinations. We developed a score using routine clinical data to prioritize interpretation according to the probability of referral-warranted findings in adults aged 50 years or older whose image sets were ultimately classified as interpretable. This single-center study of prediction-model development and internal validation included adults aged 50 years or older with completed fundus screening and a final interpretable result. Multivariable logistic regression estimated the probability of referral-warranted findings after image acquisition and predictor availability but before ophthalmologist interpretation. Timestamps were used to verify this sequence. Low-level predictor missingness was handled using 20 multiply imputed datasets. Internal validation used 500 bootstrap resamples, and a secondary temporal analysis fitted the model in 2021–2024 and evaluated it without recalibration in 2025. Complete-case, functional-form, body mass index (BMI)-inclusive, alternative-predictor, integer-score, subgroup, threshold, and missing-not-at-random sensitivity analyses were performed. Among 19,221 participants (mean age, 58.9 years; 13,723 [71.4%] men), 2,709 (14.1%) had referral-warranted findings. In the three-variable model, adjusted odds ratios were 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). The bootstrap-corrected C-statistic was 0.7478 (95% CI, 0.739–0.757), Brier score 0.1070 (95% CI, 0.104–0.110), and calibration slope 0.992 (95% CI, 0.965–1.020). In the 2025 temporal-validation sample (n=3,835; 571 events), the C-statistic was 0.744 (95% CI, 0.723–0.765), Brier score 0.110 (95% CI, 0.105–0.115), calibration intercept 0.036, and calibration slope 0.957. Across post hoc workload groups with predicted probabilities of <10%, 10% to <18%, and ≥18%, observed event proportions were 5.6%, 14.0%, and 29.1%. Although BMI improved likelihood-based fit, its inclusion did not change the C-statistic or Brier score to three decimal places. The integer score had a corrected C-statistic of 0.7466 and calibration slope of 0.984. 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. The score merits evaluation for prioritizing interpretation of acquired images, with external validation and prospective workflow assessment required before deployment.