Researchers have found that artificial intelligence (AI) can analyze retinal photographs and identify patterns associated with several established Alzheimer's disease risk factors—years before dementia develops.
The findings build on a growing body of research suggesting the retina may serve as a noninvasive window into brain health.
Although this technology is not capable of diagnosing Alzheimer's, investigators believe its utilization in routine eye exams could eventually help identify people who may benefit from earlier evaluation or preventive interventions.
Explain why the retina is being studied for this neurodegenerative disease.
The retina is considered an extension of the central nervous system, and it shares many structural and vascular features with the brain.
Because of these similarities, researchers have spent years investigating whether retinal changes mirror neurological diseases.
… for instance?
Previous studies have linked retinal abnormalities to Alzheimer's and mild cognitive impairment, while systematic reviews have concluded that retinal imaging shows promise as a source of future biomarkers.
- And other research has demonstrated that AI can detect active Alzheimer's from retinal photographs and predict cardiovascular risk factors—including age, smoking status, and blood pressure—from the same images.
These results prompted researchers to ask a different question:
Instead of detecting Alzheimer's itself, could retinal photographs reveal the risk factors that contribute to developing it?
How did researchers study retinal photographs?
Investigators trained deep learning models using 62,876 color fundus photographs from 44,501 participants in the UK Biobank.
Rather than diagnosing Alzheimer's, the AI was trained to predict 12 established Alzheimer's-related risk factors.
The categorical variables included:
- Sex
- Smoking
- Sleeplessness
- Socioeconomic status
- Alcohol use
- Depression
The continuous measures included:
- Age
- Age at completing education
- Body mass index (BMI)
- Systolic blood pressure
- Diastolic blood pressure
- Hemoglobin A1c (HbA1c)
Researchers also compared retinal findings between participants who later developed Alzheimer's—an average of 8.55 years before diagnosis—and matched individuals who did not develop dementia.
Tell me about these participants.
The study included 44,501 unique participants from the UK Biobank, contributing a total of 62,876 color fundus photographs.
Researchers also performed a separate comparison involving participants who later developed Alzheimer’s and matched controls.
For those incident Alzheimer’s cases, the retinal images were captured an average of 8.55 years before diagnosis, allowing investigators to examine whether retinal patterns differed before clinical onset.
And what did the AI detect?
The deep learning models successfully predicted multiple Alzheimer's risk factors directly from retinal photographs, although prediction accuracy varied across the different variables.
Overall: The AI generally outperformed traditional machine-learning models based solely on retinal measurements.
- The strongest performance was observed for biological characteristics such as age and sex, while prediction of lifestyle-related factors—including smoking, alcohol use, depression, and sleep problems—was more modest.
Go on …
To better understand how the models reached their predictions, researchers generated saliency maps that highlighted the retinal structures receiving the greatest attention from the AI.
The models consistently focused on the optic nerve head and retinal blood vessels—two structures that previous studies have also associated with systemic vascular and neurologic health.
Why do these findings matter?
Many Alzheimer's risk factors are documented through medical records or patient questionnaires.
Lifestyle information, on the otherhand—such as smoking, alcohol consumption, and sleep quality—is often incomplete or self-reported.
Based on the outcomes of this study: The researchers suggest retinal photographs may offer a more objective picture by capturing the cumulative effects of vascular, metabolic, and lifestyle exposures over many years.
- In other words: Retinal imaging may function less as a questionnaire and more as a biological record of long-term health.
Importantly, several AI-derived retinal scores differed significantly between participants who later developed Alzheimer's and matched controls, suggesting that retinal changes related to disease vulnerability may appear well before cognitive symptoms emerge.
Any limitations to keep in mind?
Definitely a few worth noting …
All participants were enrolled through the UK Biobank, meaning the findings may not generalize to populations with different demographic or health characteristics.
The research also identified associations rather than proving that retinal changes cause Alzheimer's disease.
Finally, although the AI identified biologically meaningful retinal patterns, it was designed to predict Alzheimer's risk factors—not diagnose the disease or estimate an individual's future risk with clinical certainty.
And what did the experts have to say?
The authors concluded that:
- Retinal photographs encode structural information reflecting multiple Alzheimer's disease risk factors
- Deep learning can uncover these patterns more effectively than conventional retinal measurements alone
Lead investigator Ruogu Fang, PhD, noted that retinal morphology may provide measurable indicators of neurovascular integrity and serve as an integrated biological sensor of cumulative Alzheimer's risk rather than simply replacing patient questionnaires.
The authors emphasized, however, that retinal imaging remains a research tool and requires additional validation before it could be incorporated into routine clinical care.
So what comes next?
The investigators have publicly released their analysis code to encourage independent validation and future research.
They also noted that because retinal photography is already widely performed during routine eye care, future studies could determine whether existing retinal images might help identify patients who would benefit from additional neurological evaluation or earlier preventive strategies.