Seoul National University Hospital develops AI to predict dementia via retinal photos

A research team led by Seoul National University Hospital has developed an artificial intelligence model capable of screening for current dementia and predicting future risk using only retinal fundus photographs. The study, published in ‘npj Digital Medicine’, utilized over 100,000 retinal images from 36,322 health examination participants collected between 2004 and 2016.

The AI model, which utilized the ‘RETFound-MAE’ foundation model with partial fine-tuning, demonstrated superior performance compared to the existing clinical dementia risk assessment tool, CAIDE. The AI achieved an AUROC of 0.750 for current dementia screening and a C-index of 0.812 for predicting future risk.

Notably, the researchers implemented ‘explainable AI’ technology, using heatmaps to visualize the specific retinal areas influencing the AI’s decisions. The analysis showed the model focused heavily on the optic disc and its surrounding areas, which aligns with known biological signals associated with dementia. This technology offers a potential low-cost, non-invasive tool for early dementia intervention using data from routine health checkups.