Deep learning-derived retinal age gap and its associations with lifestyle, systemic, and ocular health in a health screening cohort

The study built a multi-task deep learning model predicting retinal age from 29,530 fundus images (7,535 participants); the bias-corrected multi-task model trained on mixed data performed best, with a mean absolute error of 2.656 years and a Pearson correlation of 0.921 in cohort A and 2.529 years and 0.938 in cohort B. Whereas prior work centered on the RAG concept, this study provides a scalable, noninvasive retinal age prediction model with quantified performance in a real-world health screening cohort. Trained and evaluated on a large fundus image set (29,530 images), reporting mean absolute error and correlation, with bias correction and mixed-data training identified as the best configuration among compared models.

Higher RAG was significantly associated with smoking (ex-smokers beta = +0.46 years; current smokers beta = +0.50 years) and clinical diabetes (+2.52 years); both survived false discovery rate and Bonferroni correction, and the diabetes association persisted across all sequential covariate-adjustment sets. Because RAG's associations with real-world health determinants were previously unclear, this study systematically evaluated lifestyle, socioeconomic, and systemic factors in cohort A (n = 5,606) and reported associations that remain significant after multiple correction. Associations survived both false discovery rate and Bonferroni correction, and the diabetes association held across all sequential covariate-adjustment sets, indicating some robustness to model specification.

In cohort B (n = 1,810), RAG was significantly higher in eyes with age-related macular degeneration (beta = +0.60 years) and cataract (beta = +1.86 years) than in normal controls, both surviving correction. The study extends RAG associations from systemic and lifestyle factors to specific ocular diseases, providing cohort evidence linking retinal age gap to ocular health. Based on comparison within an independent ocular disease sub-cohort (n = 1,810), with both disease associations surviving multiple correction.

Married participants had lower RAG (beta = -0.46 years), significant after false discovery rate correction only; hypertension and hyperlipidemia were not associated with RAG. This adds evidence on socioeconomic factors in relation to RAG while indicating that not all cardiometabolic factors are associated with RAG, delineating the boundaries of the association. The marital status association was significant only after false discovery rate correction and did not survive stricter correction; hypertension and hyperlipidemia were null findings, indicating selective associations.

The study delineates RAG's intended setting: the authors state that RAG currently suits population-level characterization better than individual-level risk stratification, and that whether it reflects biological aging requires longitudinal validation against established aging biomarkers. Results come from a health screening cohort, with associations covering lifestyle (smoking), systemic factors (diabetes), and ocular diseases (age-related macular degeneration, cataract), while hypertension and hyperlipidemia showed no association, indicating a selective scope of applicability.

Readers should still watch: as an imaging-derived age-prediction residual, RAG's correspondence to biological aging awaits longitudinal validation; the marital status association was significant only after false discovery rate correction, weaker in robustness than the smoking and diabetes findings; the null results for hypertension and hyperlipidemia suggest selective associations; and because this is abstract-level information lacking figures and full statistical detail, assessing model architecture, the specific composition of covariate sets, or confidence intervals for effect sizes would require consulting the original text.