Prediction of Alzheimer's disease risk factors from retinal images via deep learning: Development and validation of biologically relevant morphological associations in the UK Biobank.
Level 3 - non-randomized controlled study
Cohort-based deep learning development and nested case-control comparison in the UK Biobank.
PubMed 42299837 · doi:10.1177/13872877261457650
What was done
Deep learning models were trained on 62,876 color fundus photographs from 44,501 participants in the UK Biobank to predict 12 systemic, metabolic, and lifestyle risk factors for Alzheimer's disease (AD). Six categorical factors (sex, smoking, sleeplessness, economic status, alcohol use, depression) and six continuous variables (age, age at completing education, body mass index, systolic blood pressure, diastolic blood pressure, HbA1c) were evaluated. Saliency mapping (CAM-Score) identified key retinal regions, and scores were compared between individuals with incident AD (mean 8.55 years before onset) and matched controls.
What was found
Deep learning models outperformed most morphometry-based machine learning models. Prediction accuracy varied across targets: AUROC ranged from 0.5654 to 0.9480 for categorical factors, and R² ranged from -0.0291 to 0.7620 for continuous factors. Saliency mapping consistently highlighted the optic nerve head and retinal vasculature. Several saliency-based scores differed significantly between incident AD cases and matched controls, though specific effect sizes and p-values for the case-control comparison were not reported in the abstract.
Why it matters
This study shows that non-invasive retinal fundus photography encodes structural and microvascular signatures linked to systemic Alzheimer's disease risk factors and preclinical disease vulnerability years before clinical onset.
Limits
Predictive accuracy was poor for several factors (R² below zero, AUROC near 0.57), and the abstract does not report individual metric breakdowns for each of the 12 factors. The UK Biobank cohort exhibits healthy-volunteer selection bias and limited ancestral diversity. Retinal markers primarily reflect systemic and vascular vulnerability rather than AD-specific neuropathology.
Cited by
- partial Artificial intelligence analysis of retinal photographs can predict the onset of Alzheimer's disease 5 to 7 years in advance.