Abstract
AIMS: Both Life's essential 8 (LE8) and retinal photographs are closely related to cardiovascular diseases (CVDs). We aimed to develop a novel deep learning-based tool for CVD risk stratification, termed RetiLE8, by predicting LE8 based on retinal photographs. METHODS AND RESULTS: This study was based on the UK Biobank, a prospective cohort study. Retinal photographs from the UK Biobank were used to train a deep learning model to predict LE8 scores, generating the RetiLE8 scores. Cox proportional hazards models were used to estimate the association of RetiLE8 with all-cause mortality, CVD mortality, and CVD events. Model performance was compared with that of LE8 and the Pooled Cohort Equations (PCE), a contemporary risk estimation tool, using Harrell's concordance index (C-index) and continuous net reclassification improvement (NRI). Retinal photographs from 10 798 participants and 25 750 participants of the UK Biobank were utilized for the development and validation of RetiLE8, respectively. RetiLE8 showed a modest correlation with LE8. One standard deviation (SD) increase in the RetiLE8 score was associated with 13% (95% CI, 7-19%) lower all-cause mortality risk and 10% (4-14%) lower CVD event risk, independent of the LE8 score and covariates. The RetiLE8 score showed similar discrimination to the LE8 score and the PCE in predicting outcomes and significantly improved risk stratification beyond both tools. CONCLUSION: The RetiLE8 score may serve as a complementary tool for CVD risk stratification with existing tools. Prospective studies implementing this tool in clinical practice are warranted to evaluate its utility in real-world settings.