Summary: A new study using artificial intelligence to analyze routine retinal photographs has revealed an affordable, non-invasive way to detect major Alzheimer’s disease risk factors decades before symptoms emerge. By training deep learning models on retinal images from over 40,000 participants in a UK-based database, researchers identified specific retinal regions—especially the retinal blood vessels and optic nerve head—that correlate with biological and lifestyle risk markers linked to Alzheimer’s vulnerability.
The AI models accurately predicted traits such as biological sex, blood pressure, smoking status, alcohol use, and chronic sleep problems. Because the retina is a direct extension of the central nervous system, routine retinal images can act as an integrated biological sensor of accumulated neurovascular damage. This offers a practical window for early lifestyle and medical interventions long before irreversible brain changes occur.
Key Facts
- The ocular window: Retinal structure—particularly the retinal vasculature and optic nerve head—provides measurable indicators of neurovascular health and potential Alzheimer’s vulnerability.
- Large dataset: The model was developed and validated using retinal images from more than 40,000 people drawn from a major UK biobank.
- Objective risk mapping: AI can detect lifestyle and biological risks (high blood pressure, smoking, alcohol use, insomnia) that are often underreported or missing from medical records.
- Early intervention potential: Alzheimer’s pathologies evolve over decades; low-cost retinal screening could flag at-risk individuals long before clinical decline.
- Accessible and economical: Retinal photographs are routinely taken during eye exams, diabetes screenings, and glaucoma checks—making this approach far more scalable than MRI or PET scans.
Source: University of Florida
Often called “the window to the soul,” the eye may also reveal the brain’s health.
A large study of tens of thousands of patients shows that inexpensive, commonly captured retinal photographs can predict many risk factors associated with future Alzheimer’s disease. These findings suggest that routine eye imaging could become a practical screening tool for identifying long-term dementia risk.
“Alzheimer’s disease develops over decades, yet most diagnostic tools focus on late-stage pathology when intervention is limited,” said Ruogu Fang, Ph.D., professor of biomedical engineering at the University of Florida and the study’s lead author. “By examining retinal biomarkers, we can identify individuals at elevated risk, recommend targeted testing, and promote lifestyle changes that may reduce that risk.”
Fang and collaborators, including UF’s Adam Woods, Ph.D., and Meta researcher Yunchao Yang, Ph.D., published their results in the Journal of Alzheimer’s Disease.
Retinal photographs are common in clinical practice—patients with diabetes, glaucoma, or cataracts typically have multiple images on record, and many routine eye exams also capture fundus photos. That ubiquity makes retinal screening a cost-effective complement to more expensive imaging techniques like MRI or PET.
Using deep learning to analyze more than 62,000 color fundus photographs from 44,501 unique participants in the UK Biobank, the research team trained models to predict a set of Alzheimer’s-related risk factors. They identified retinal regions most important to predictions and linked those regions to known anatomical and vascular features, including the optic nerve head and large retinal vessels.
“AI enables detection of subtle retinal differences across thousands of subjects that were previously unnoticed,” said Seowung Leem, a doctoral student at UF and first author of the paper. “Those subtle changes may serve as early indicators of future disease risk.”
The models predicted both categorical traits (sex, smoking, sleeplessness, socioeconomic status, alcohol use, depression) and continuous measures (age, education completion age, BMI, systolic and diastolic blood pressure, HbA1c). Performance varied by trait but exceeded many traditional morphometry-based approaches. Importantly, saliency maps repeatedly highlighted biologically plausible areas—the optic nerve and retinal vasculature—as the most informative regions.
Unlike medical records, which often rely on incomplete or biased self-reporting for lifestyle factors, retinal imaging captures the cumulative physiological effects of those behaviors. The photographic evidence reflects years of neurovascular change and therefore can provide an objective indicator of individual risk not visible in chart notes alone.
“Retinal morphology may offer measurable signals of neurovascular integrity that relate to Alzheimer’s vulnerability,” Fang said. “Retinal imaging acts less like a questionnaire and more like an integrated sensor of cumulative risk.”
Previous work by the group showed retinal photos can identify active Alzheimer’s cases. This study extends that work by linking retinal patterns to risk factors that precede clinical disease by years or decades, suggesting a role for early prevention strategies—aggressive cardiovascular care, sleep optimization, targeted lifestyle programs, cognitive training, or early therapeutic approaches—when the brain is still resilient.
Funding: Supported in part by the National Science Foundation.
Key Questions Answered:
A: The retina is developmentally and anatomically an extension of the central nervous system. Its blood vessels and nerve fibers mirror the brain’s microvasculature and neural pathways. Systemic issues—vascular disease, chronic inflammation, or sleep disturbances—leave traces in retinal structure, making the eye a visible proxy for brain health.
A: Medical charts often depend on self-reported behaviors, which can be inaccurate. The AI reads the physical, cumulative impact of those behaviors on ocular tissue over decades, turning a simple photo into an objective measure of neurovascular damage rather than a subjective report.
A: Early identification enables proactive prevention. Options include aggressive cardiovascular risk management, structured cognitive training, targeted sleep therapy to support brain clearance mechanisms, lifestyle modifications (smoking cessation, alcohol moderation, exercise), and consideration of early therapeutic strategies when appropriate. These measures are most effective when implemented long before extensive neural loss occurs.
Editorial Notes:
- Edited by a Neuroscience News editor.
- Journal paper reviewed in full.
- Additional context provided by staff.
About this AI and Alzheimer’s research news
Author: Eric Hamilton
Source: University of Florida
Contact: Eric Hamilton – University of Florida
Image: Image credit: Neuroscience News
Original Research: Closed access. “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” by Ruogu Fang et al., Journal of Alzheimer’s Disease. DOI: 10.1177/13872877261457650
Abstract
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
Background
Systemic metabolic and lifestyle factors are linked to Alzheimer’s disease through epidemiology and biomarker studies. Whether color fundus photography (CFP) contains structural retinal signatures that correspond to these Alzheimer’s-related risk domains has been unclear.
Objective
To test whether deep learning can predict a set of Alzheimer’s-related risk factors from CFP and to map the retinal structures that drive those predictions, assessing whether CFP reflects pathways to Alzheimer’s vulnerability.
Methods
Using 62,876 CFPs from 44,501 participants in the UK Biobank, models were trained to predict 12 factors linked to Alzheimer’s pathology or incidence: six categorical (sex, smoking, sleeplessness, economic status, alcohol use, depression) and six continuous (age, age at education completion, BMI, systolic and diastolic blood pressure, HbA1c). Performance, model saliency, and saliency-derived scores were evaluated and compared with retinal morphometry. Scores were also compared between participants who later developed Alzheimer’s and matched controls.
Results
Model performance varied by trait, with categorical AUROC values ranging from about 0.57 to 0.95 and continuous R² values from slightly below zero to 0.76, often outperforming morphometry-based approaches. Saliency analyses consistently highlighted biologically meaningful regions—chiefly the optic nerve head and retinal vasculature—and aligned with measurable morphometric differences. Several saliency-derived scores differed significantly between participants who later developed Alzheimer’s and matched controls, suggesting overlap between retinal correlates of risk factors and early disease-related changes.
Conclusions
Color fundus photography encodes retinal signatures associated with Alzheimer’s risk factors. While not diagnostic on its own, deep learning-derived retinal representations can reveal biologically meaningful structural changes linked to vulnerability and may help identify individuals who could benefit from early preventive strategies.