Objective: Diabetes and hypertension pose significant health risks, especially when poorly managed. Retinal evaluation though fundus photography can provide non-invasive assessment of these diseases, yet prior studies focused on disease presence, overlooking control statuses. This study evaluated vision transformer (ViT)-based models for assessing the presence and control statuses of diabetes and hypertension from retinal images.
Methods: ViT-based models with ResNet-50 for patch projection were trained on images from the UK Biobank (n = 113,713) and Singapore Epidemiology of Eye Diseases study (n = 17,783), and externally validated on the Singapore Prospective Study Programme (n = 7,793) and the Beijing Eye Study (n = 6064). Model performance was evaluated using the area under the receiver operating characteristic curve (AUROC) for multiple tasks: detecting disease, identifying poorly controlled and well-controlled cases, distinguishing between poorly and well-controlled cases, and detecting pre-diabetes or pre-hypertension.
Results: The models demonstrated strong performance in detecting disease presence, with AUROC values of 0.820 for diabetes and 0.781 for hypertension in internal testing. External validation showed AUROCs ranging from 0.635 to 0.755 for diabetes, and 0.727 to 0.832 for hypertension. For identifying poorly controlled cases, the performance remained high with AUROCs of 0.871 (internal) and 0.655-0.851 (external) for diabetes, and 0.853 (internal) and 0.792-0.915 (external) for hypertension. Detection of well-controlled cases also yielded promising results for diabetes (0.802 [internal]; 0.675-0.838 [external]), and hypertension (0.740 [internal] and 0.675-0.807 [external]). In distinguishing between poorly and well-controlled disease, AUROCs were more modest with 0.630 (internal) and 0.512-0.547 (external) for diabetes, and 0.651 (internal) and 0.639-0.683 (external) for hypertension. For pre-disease detection, the models achieved AUROCs of 0.746 (internal) and 0.523-0.590 (external) for pre-diabetes, and 0.669 (internal) and 0.645-0.679 (external) for pre-hypertension.
Conclusion: ViT-based models show promise in classifying the presence and control statuses of diabetes and hypertension from retinal images. These findings support the potential of retinal imaging as a tool in primary care for opportunistic detection of diabetes and hypertension, risk stratification, and individualised treatment planning. Further validation in diverse clinical settings is warranted to confirm practical utility.
Supplementary information: The online version contains supplementary material available at 10.1007/s13167-025-00412-9.
Keywords: Deep learning; Diabetes; Health risk assessment; Hypertension; Improved individual outcomes; Innovative screening programs; Opportunistic screening; Predictive Preventive Personalized Medicine (PPPM / 3PM); Preventable diseases; Protection against health-to-disease transition; Retinal image; Risk stratification.
© The Author(s) 2025.