Fine-tuning RETFound reveals retinal age gap markers and their clinical and genetic divergence between the sexes

Medical research based on vision foundation models demonstrates that fundus photography is more than a diagnostic tool for visual impairment, serving instead as a non-invasive biological window into systemic ageing across the entire body. In a study published in Nature Communications, researchers from the University of Lausanne and the Swiss Institute of Bioinformatics, in collaboration with European research centres, fine-tuned the RETFound foundation model to accurately estimate chronological age from colour fundus photographs of 71,343 participants in the UK Biobank, achieving a mean absolute error of no more than 2.85 years.

The central finding of the study was calculating what is known as the retinal age gap, defined as the mathematical difference between the model's predicted age and an individual's actual chronological age.Analyses showed that a widening of this gap is directly associated with increased risks of cardiometabolic traits, inflammatory markers, cognitive decline, cardiovascular disease, dementia, and cancer, as well as all-cause mortality.Genome-wide association analyses further reinforced these findings, uncovering genetic pathways linked to longevity, metabolic processes, neurodegenerative diseases, and age-related ocular disorders.

The most notable feature of the model's results is the clear biological divergence between sexes despite identical statistical prediction accuracy. Sex-stratified models revealed that the retinal age gap in males correlates strongly with metabolic syndrome and metabolic disorders, whereas the model's attention and genetic analyses in females point toward retinal vasculature. Furthermore, the study demonstrated that retinal ageing patterns in women shift substantially across menopause, with post-menopausal women exhibiting higher retinal age gap values alongside clinical correlations that increasingly resemble the patterns observed in men.

This scientific advance moves computer vision models from specialised ophthalmic screening into preventative clinical triage within healthcare systems. In regional markets, whether in primary care centres across the Gulf states facing growing burdens of metabolic syndromes and diabetes, or in public health systems across Egypt and the Levant, adopting foundation models for retinal screening offers a low-cost, scalable mechanism for early assessment of cardiovascular and neurological risks without requiring complex imaging scans or expensive genetic sequencing.

The fundamental shift lies in moving beyond one-size-fits-all medical models and constructing diagnostic pathways that account for biological sex differences and stages of hormonal change.Technical and medical teams responsible for integrating artificial intelligence into hospitals must recognise that a single visual biomarker carries metabolic implications in one group and vascular indicators in another, requiring calibrated clinical decision criteria before deploying algorithms across early detection programmes.