Retinal Artificial Intelligence Reclassifies SCORE2-Based Cardiovascular Risk Across the ESC/EAS Statin-Indication...

Whether a retinal artificial intelligence (AI) risk score reclassifies SCORE2 cardiovascular risk stratification in statin-naive adults across the ESC/EAS statin-indication boundary is uncertain. We evaluated the retinal AI risk score for 10-year atherosclerotic cardiovascular disease (ASCVD) prediction.

In 40,672 statin-naive UK Biobank participants (aged 40-69, without diabetes), cardiovascular risk categories and treatment recommendation status (Class I or IIa vs neither) were defined using the 2025 ESC/EAS focused-update framework. Cox models estimated hazard ratios (HRs) adjusted for SCORE2 or for established risk factors. The incremental predictive value of adding the retinal AI risk score to SCORE2 was evaluated using 10-fold cross-validation.

Over a 10.0-year median follow-up (914 ASCVD events), the very high retinal AI risk group independently predicted ASCVD (SCORE2-adjusted HR 2.32 vs low risk; p<0.001). Adding the retinal AI risk score to SCORE2 reclassified risk across the ESC/EAS statin-indication boundary (statin-eligibility NRI [net reclassification improvement] 0.012, p<0.001); reclassification was statistically significant but modest (continuous NRI 0.445, p<0.001), without C-index or net-benefit gain. In the statin-indicated subgroup, retinal AI risk groups further stratified risk (SCORE2-adjusted HR 2.16; p<0.001). Statin-not-indicated participants in the high or very high retinal AI risk groups had numerically similar observed 10-year risk to all statin-indicated, untreated participants (p=0.758). There was no evidence of effect modification by clinically detectable retinal disorder (interaction p=0.93). Incremental value was greater in women than men (sex interaction p<0.001).

Adding the retinal AI risk score to SCORE2 produced modest bidirectional risk reclassification across the ESC/EAS statin-indication boundary; these findings warrant prospective validation.Abstract word count: 250 words.