Predicting Alzheimer's Disease Risk from Retinal Photos with Deep Learning

This post reviews Prediction of Alzheimer's Disease Risk Factors from Retinal Images via Deep Learning

• Authors: Seowung Leem, Yunchao Yang, Adam J. Woods, Ruogu Fang
• arXiv: 2605.00665 | 2026-04-30

The eye as a window to the brain

The retina, the light-sensitive tissue at the back of the eye, is effectively an extension of the brain. It shares the brain's embryonic origin, has similar vascular structures, is affected by the same metabolic processes, and can be observed directly and non-invasively. Changes in the brain may therefore leave traces in the retina.

Why Alzheimer's disease screening needs alternatives

Alzheimer's disease (AD) is the most common neurodegenerative disease. Traditional diagnosis relies on:

• Cerebrospinal fluid tests (invasive)
• PET scans (expensive, involve radiation)
• Cognitive tests (only detect late-stage decline)

Potential retinal markers of AD include vascular changes (AD is associated with microvascular pathology), ganglion cell loss (thinning of retinal nerve fibers), metabolic markers, and inflammatory markers.

What the paper does

Using 62,876 fundus photographs from UK Biobank, the authors trained deep learning models to predict 12 AD-related risk factors, covering:

• Cardiovascular risk (hypertension, cholesterol)
• Metabolic risk (diabetes, BMI)
• Lifestyle (smoking, alcohol, exercise)
• Genetic factors (APOE genotype)
• Inflammatory markers

Key findings

1. Retinal images genuinely carry AD risk signals. Models predict multiple risk factors in ways that are not random correlations but biologically grounded. 2. Morphological associations are interpretable. Examples: vascular narrowing correlates with hypertension; nerve fiber layer thinning correlates with cognitive decline. 3. Predictions have clinical value, enabling identification of high-risk individuals for early intervention and prevention.

Not a black box: biological relevance

The paper emphasizes biological relevance alongside accuracy:

• Gradient heatmaps show which retinal regions the model attends to
• Attention regions are checked against known retina-brain associations
• Predictions are validated on independent datasets for reproducibility

This turns prediction from a "black box" into evidence-based inference. As the post puts it, echoing Feynman: good medical AI doesn't just predict — it gives reasons that are consistent with known biology.

Takeaways for medical AI builders

1. Does your prediction have a biological basis? 2. Is the model explainable — which regions/features does it attend to? 3. Has it been validated on independent data? 4. Can "easy-to-acquire" data (fundus photos) infer "hard-to-acquire" information (brain health)?

The core lesson: the highest standard for medical AI is not just accurate prediction, but accurate and explainable prediction. When AI can read brain health from a retinal photo, it becomes not just a prediction tool but an engine of scientific discovery — uncovering hidden connections between organs. The eye truly leads to the brain, and AI is letting us read that channel for the first time.