AHA Scientific Sessions 2026 Abstract · Nov 2, 2026
Prognostic Value of Higher-Eye Score and Between-Eye Discordance in Retinal Artificial Intelligence for 10-Year Atherosclerotic Cardiovascular Disease Prediction in the UK Biobank
By Nam D, Thakur S, Nusinovici S et al.
Background: Deep-learning retinal artificial intelligence (AI) cardiovascular (CV) risk scores are typically derived as the bilateral mean of the two eye-specific predictions, an aggregation convention adopted without direct prognostic evaluation. Whether an isolated elevation in one eye reflects unrecognized CV risk, and whether the higher-eye score or between-eye discordance adds prognostic information beyond the mean, remains unclear. Hypothesis: Using the higher-eye score, reporting both eye scores jointly, or quantifying between-eye discordance would improve 10-year atherosclerotic CV disease (ASCVD) prediction beyond the bilateral mean. Methods: We analyzed 48,061 UK Biobank participants aged 40-69 with bilateral fundus imaging at baseline; the outcome was 10-year incident ASCVD (myocardial infarction, ischemic stroke, or CV death; 1,203 events). Per-image scores from the retinal-AI CV risk algorithm (Dr. Noon CVD) were aggregated to the bilateral mean, the higher- and lower-eye scores, and between-eye discordance metrics (absolute difference and proportional log-ratio difference). All Cox proportional-hazards models were adjusted for age, sex, smoking, systolic blood pressure, total cholesterol, HDL cholesterol, and baseline lipid-lowering therapy. Each alternative representation was compared against the bilateral mean using Harrell's C-index, likelihood-ratio tests, and category-free net reclassification improvement (NRI). Results: Scores from the two eyes were highly correlated (left-right r=0.77), and the bilateral mean and the higher-eye score were near-identical (r=0.98). The two yielded equivalent risk stratification (Harrell's C-index, p=0.83). Across five between-eye absolute-difference bands (<0.05 to ≥0.20), 10-year ASCVD incidence rates showed no monotonic trend (1.4 to 2.9 per 1,000 person-years). Neither discordance metric nor separate reporting of both eye scores added information beyond the mean (absolute-difference hazard ratio 2.39, 95% CI 0.61-9.27; all p≥0.21). Findings were consistent in statin-treated participants (n=6,858; statin х AI-score interaction p=0.42). Of participants classified low-risk by the bilateral mean (validated thresholds 0.31/0.41/0.51), only 0.06% (7 of 11,852) were re-categorized to high or very-high risk by the higher-eye score, with no events. The NRI was 0.034 (95% CI –0.099 to 0.121, p=0.55). Conclusions: In this general-population cohort, the bilateral mean of a retinal-AI CV risk score captured essentially all eye-derived prognostic information for 10-year ASCVD. Neither the higher-eye score nor between-eye discordance improved prediction beyond it, and this held among statin-treated participants. These findings provide direct prognostic support for the prevailing bilateral-mean convention and indicate that between-eye asymmetry does not flag unrecognized risk.