A research group from Seoul National University has created an artificial intelligence tool capable of identifying early indicators of dementia in retinal scans and estimating an individual’s likelihood of developing the disease later. The training dataset was compiled from individuals who underwent routine checkups at Seoul National University Hospital between 2004 and 2016, with the team including researchers from SNU Hospital, SNU College of Medicine, and the university’s spinoff company XAIMED.
The team developed two specialized models. The first was designed to recognize active dementia cases, using 14,000 images from 10,448 participants. The second model forecasted future dementia risk by analyzing 65,000 images from 25,874 individuals over a median follow-up of 5.5 years. Their findings were published in npj Digital Medicine, where five AI approaches were evaluated based on accuracy, clinical utility, and transparency. The researchers employed saliency maps to highlight retinal regions influencing predictions, while anatomical segmentation quantified the AI’s focus on structures like blood vessels and the optic disc.
The RETFound-MAE model, refined through partial fine-tuning, delivered the best performance in both applications. For detecting existing dementia, it achieved an AUROC of 0.75, surpassing the CAIDE Dementia Risk Score, which registered 0.62 on the same measure. In risk prediction, RETFound-MAE’s C-index of 0.81 also exceeded CAIDE’s 0.69. Decision-curve analysis further confirmed the AI’s greater net benefit across varying risk thresholds, outperforming both the CAIDE score and conventional screening strategies.
When integrated with CAIDE’s risk factors, even without education-level data, the AI’s ability to distinguish dementia cases improved marginally, raising the AUROC from 0.75 to 0.77. The researchers clarified that the tool would serve as a screening aid rather than a definitive diagnostic instrument. At its selected threshold, it identified 62.5% of true dementia cases but produced a 36.3% positive predictive value, indicating that high-risk alerts would require additional verification. The model’s sensitivity suggests it could effectively flag potential cases for further evaluation, though its positive predictive value shows the need for complementary assessments to confirm results.
Existing dementia screening often depends on expensive, invasive procedures such as MRI or cerebrospinal fluid tests, which restrict broad adoption. Retinal imaging, already standard in eye clinics, provides a more affordable alternative. The SNU team observed that prior studies had concentrated on symptomatic Alzheimer’s or mild cognitive impairment, rather than predictive risk assessment using routine scans. Unlike earlier research, which often relied on specialized imaging or cross-sectional data, this model leverages standard fundus photographs collected during regular eye examinations, expanding its potential for population-level screening.
