Explainable foundation model for dementia screening and risk stratification using retinal fundus images

Dementia is a growing global health challenge, and early identification is essential for timely intervention. We evaluated whether foundation model-based deep learning of retinal fundus photographs can detect dementia and predict future dementia incidence. Using health checkup data from over 36,000 Korean individuals, dementia was defined by diagnosis with relevant medication. Five vision foundation models were assessed with multiple fine-tuning strategies. The best-performing model, RETFound-MAE with partial fine-tuning, achieved an AUROC of 0.750 for dementia detection and a C-index of 0.812 for future incidence prediction. Model outputs remained independent risk factors after adjustment (adjusted odds ratio, 1.155; adjusted hazard ratio, 1.045). Quantitative saliency analyses highlighted biologically plausible regions, primarily the optic disc and adjacent peripapillary areas. These findings suggest that foundation model-based analysis of routinely acquired retinal fundus images may provide a scalable and interpretable approach for opportunistic dementia risk stratification, warranting validation in ethnically diverse populations.