Researchers from Seoul National University have developed an artificial intelligence model that analyzes retinal images to identify people who may currently have dementia and predict their future risk of developing the condition. The research involved teams from Seoul National University Hospital, SNU College of Medicine and university spin-off XAIMED, using more than 100,000 retinal fundus images collected from over 36,000 people during routine health examinations.
The researchers developed separate deep learning models for detecting existing dementia and estimating future risk. More than 14,000 retinal images from 10,448 individuals were used for dementia detection, while nearly 65,000 images from 25,874 individuals were analyzed for future risk prediction, with participants followed for a median of 5.5 years. The study compared five vision foundation models and multiple fine-tuning approaches to identify the strongest-performing configuration.
Published in npj Digital Medicine, the study found that the RETFound-MAE model with partial fine-tuning performed best, achieving an AUROC of 0.75 for identifying existing dementia and a C-index of 0.81 for predicting future dementia. It also outperformed the conventional CAIDE Dementia Risk Score in the study population. Combining the retinal AI model with individual CAIDE risk factors further improved dementia detection performance.
The researchers also analyzed which parts of the retina influenced the model's predictions, including blood vessels and the optic disc, to improve explainability. However, they emphasized that the technology is not intended to independently diagnose dementia. At the selected threshold, the model had a sensitivity of 62.5% and positive predictive value of 36.3%, meaning individuals identified as high risk would still require cognitive assessment and other confirmatory testing.
The potential advantage is that retinal photography is relatively inexpensive, non-invasive and already widely performed during routine health checks and ophthalmology appointments. This could allow AI analysis to operate as an opportunistic screening tool, identifying people who may benefit from further dementia evaluation without requiring dedicated neurological testing. The researchers said external validation across different populations, institutions and retinal cameras, along with health economic studies, will be required before clinical deployment.
Click here to read the original news story.