An ensemble multitask deep-learning model trained on fundus photographs from 27,214 disease-free adults estimated retinal age with mean absolute error as low as 2.78 years in internal validation and 3.39 years in a primary external cohort, with reduced performance in a larger heterogeneous dataset. The model used glycated hemoglobin as an auxiliary training signal but required only images at inference. A larger retinal age gap (predicted minus chronological age) was associated with diabetes, cardiac disease, and stroke after age- and sex-matched analyses, supporting retinal age as a potential noninvasive biomarker of systemic ageing, though findings are observational and limited by population and data constraints.