Sacha Bors

While sex as a biological variable has gained increasing attention in recent years, its effect in shaping retinal vascular phenotypes and their genetic architecture remains poorly characterized. Given known sex differences in vascular biology and disease risk, elucidating these effects may provide insights into sex-specific mechanisms of systemic vascular disease. Here, we investigated sex differences in retinal vascular phenotypes in the UK Biobank (N = 68k), which are emerging as non-invasive biomarkers for ocular, cardiovascular, and neurodegenerative diseases. We used both interaction and sex-stratified analyses in disease association and genome-wide association studies, including the X chromosome. Our analyses identified sex-differential associations between retinal traits and cardiometabolic risk factors as well as disease outcomes. Several phenotypes showed prognostic relevance for mortality and cardiovascular endpoints in Cox models, predominantly in one sex, illustrating the limitations of sex-agnostic analyses. Genetic analyses indicated modest differences in genetic architectures, with females showing a greater number of associated variants and gene-level signals, and evidence for sex-specific trait–gene associations. These findings suggest sex-differential genetic and environmental contributions to microvascular morphology, possibly reflecting divergent biological pathways and risk profiles. Our study underscores the necessity of integrating sex-aware approaches in genetic and prognostic research to advance precision medicine and improve tailored risk stratification.

Retinal fundus images offer a non-invasive window into systemic aging. Here, we fine-tune a foundation model (RETFound) to predict chronological age from color fundus images in 71,343 participants from the UK Biobank, achieving a mean absolute error of 2.85 years. The resulting retinal age gap, i.e. the difference between predicted and chronological age, is associated with cardiometabolic traits, inflammation, cognitive performance, all-cause mortality, dementia, cancer, and incident cardiovascular disease. Genome-wide analyses identify genes related to longevity, metabolism, neurodegeneration, and age-related eye diseases. Sex-stratified models reveal consistent performance but divergent biological signatures: males have stronger links to metabolic syndrome, while in females, both model attention and genetics point to a greater involvement of retinal vasculature. Additional analyses indicate that retinal aging patterns in females vary across the menopausal transition, with postmenopausal females exhibiting higher retinal age gap values and clinical associations that more closely resemble those observed in males. Our study positions the retinal age gap as a biologically relevant and sex-specific phenotype associated with multiple aging-related diseases and outcomes beyond conventional risk factors, including chronological age.

Deep learning models based on Vision Transformers (ViTs) have shown strong performance in retinal fundus imaging, but their interpretability remains poorly understood. In particular, attention-based attribution methods are widely used to explain ViT predictions, despite limited evaluation of their faithfulness and biological relevance in medical imaging. Here, we systematically benchmark four attention-based interpretability methods for RETFound, a retinal ViT-based foundation model, that we previously fine-tuned to predict 17 retinal vascular phenotypes from UK Biobank fundus images ¹ . We compare raw attention, attention rollout, gradient-weighted attention rollout, and Chefer’s hybrid relevance-based method using both qualitative visualisation and quantitative evaluation frameworks. To assess attribution faithfulness, we perform perturbation-based deletion and insertion experiments, quantifying changes in model predictions as highly attended image regions are progressively removed or restored. To evaluate biological specificity, we run structure-aware analyses combining attribution maps with vessel segmentation and artery-vein labels through the Relative ratio of Attention Intensity (RAI) metric. Across models, attribution maps differed substantially depending on the selected interpretability method, highlighting the need for rigorous quantitative evaluation. Among the evaluated approaches, gradient-weighted attention rollout consistently achieved the strongest perturbation performance and produced attribution maps most closely aligned with the anatomical definition of the predicted retinal traits. Furthermore, vessel-type specific models systematically concentrate attention on the corresponding vascular structures despite being trained using only a single scalar value per image as supervision. These findings demonstrate that attention-based attribution methods capture biologically meaningful vascular representations, while also revealing method-dependent variability in attribution behaviour. This work provides a quantitative framework for evaluating interpretability methods in medical imaging with annotated segmentation and contributes toward more transparent and biologically grounded medical AI systems.