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

Area of Science:

• Ophthalmology
• Neurology
• Artificial Intelligence

Background:

• Dementia presents a significant global health challenge, necessitating early detection for effective intervention.
• Retinal fundus photography is a non-invasive imaging technique that offers insights into vascular and neurological health.

Purpose of the Study:

• To evaluate the efficacy of foundation model-based deep learning applied to retinal fundus photographs for dementia detection.
• To assess the capability of these models in predicting the future incidence of dementia.

Main Methods:

• Utilized health checkup data from over 36,000 Korean individuals, defining dementia by diagnosis and relevant medication.
• Assessed five vision foundation models with various fine-tuning strategies, focusing on the RETFound-MAE model.
• Employed quantitative saliency analyses to identify key image regions influencing model predictions.

Main Results:

• The RETFound-MAE model achieved an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.750 for dementia detection.
• The model demonstrated a C-index of 0.812 for predicting future dementia incidence.
• Model outputs were identified as independent risk factors for dementia (adjusted odds ratio: 1.155; adjusted hazard ratio: 1.045).

Conclusions:

• Foundation model-based analysis of retinal fundus images shows promise for scalable and interpretable dementia risk stratification.
• The study highlights biologically plausible regions, such as the optic disc, as important indicators.
• Further validation in diverse ethnic populations is warranted to confirm these findings.