Artificial intelligence continues to push the boundaries of what retinal imaging can tell us – not just about the eye, but also about the whole body. A new Communications Medicine study – conducted by Ninomiya and colleagues at the Tohoku University Graduate School of Medicine, Japan – suggests that a single fundus photograph may offer a surprisingly accurate estimate of biological aging, with potential implications for systemic disease screening.
The concept of “retinal age” is gaining traction within the emerging field of oculomics. By comparing AI-predicted age from retinal images with a patient’s chronological age – known as the retinal age gap (RAG) – researchers aim to quantify biological aging and disease risk. While previous models have achieved reasonable accuracy, generalizability and clinical utility have remained challenges.
In this study, the authors developed a multitask deep learning model trained on more than 50,000 fundus images from over 27,000 healthy individuals. Unlike conventional single-task models, this system simultaneously learned to predict age and glycated haemoglobin (HbA1c), using the latter as an auxiliary training signal. Crucially, the deployed model requires only a fundus image, making it compatible with routine clinical workflows.
The results are notable. The model achieved a mean absolute error of 2.78 years in internal validation – an improvement over many previously reported systems. External validation showed similarly strong performance, with errors of 3.39 years in a hospital cohort, though accuracy declined in a larger, more heterogeneous dataset, highlighting ongoing challenges with real-world generalizability.
A key innovation lies in the use of an ensemble approach, combining predictions from multiple models to improve stability and provide a measure of confidence.
Importantly for clinicians, the study links retinal age to systemic disease. Patients with diabetes, cardiac disease, or a history of stroke demonstrated significantly higher retinal age gaps – suggesting their retinas appear “older” than expected. These findings reinforce the notion that retinal imaging can act as a window into systemic health, particularly cardiometabolic status.
The model also highlighted anatomical regions driving predictions. As demonstrated in the findings, attention was consistently focused on the optic disc, macula, and major vascular arcades – structures long associated with systemic vascular health.
From a practical perspective, the appeal of this approach lies in its simplicity. Fundus photography is already widely used in screening programs, and integrating an AI-based “retinal age” output could add value without requiring additional tests. Patients with a high retinal age gap then might be flagged for further cardiovascular or metabolic evaluation.
However, limitations remain. The study population was predominantly Asian, and performance dropped in more diverse datasets. Image quality and acquisition variability also influenced accuracy – an important consideration for real-world deployment.
Nonetheless, this work adds to growing evidence that the retina can serve as a non-invasive biomarker of systemic aging. For eye care professionals, it raises an intriguing possibility: that routine retinal imaging could evolve from a purely ophthalmic tool into a broader platform for preventive medicine.
