Researchers at Mayo Clinic and Arizona State University are studying whether artificial intelligence (AI) can detect retinal patterns linked to Alzheimer’s disease before patients experience symptoms. The approach might eventually provide cost-effective, early diagnosis.
“Our vision is that patients’ annual eye scans with an optometrist could be coupled with an AI model that generates a clear and individualized risk profile for neurodegeneration and cerebrovascular disease,” says Oana M. Dumitrascu, M.D., a vascular neurologist at Mayo Clinic in Phoenix/Scottsdale, Arizona. “The eye can be a window to the brain.”
The retina’s vascular parameters correlate with brain vascular health as well as systemic vascular health. It’s also known that the retina can accumulate amyloid proteins and exhibit inflammatory changes.
“We have shown that specific analysis of the retinal vasculature can effectively distinguish patients with presymptomatic Alzheimer’s disease from controls — perhaps because those abnormal protein deposits in the brain or retina are affecting the health of the vasculature as well,” Dr. Dumitrascu says. “The advantage of the retina is that those proteins might be visualized noninvasively and repeatedly, without radiation.”
Routine eye care examinations use color fundus photography, which can reveal retinal vascular abnormalities. But those abnormalities might not be specific to Alzheimer’s disease. Fundus photography also can’t quantify abnormal protein accumulation.
“That’s why we rely on AI. It allows us to analyze millions of tiny image changes simultaneously,” says Yalin Wang, Ph.D., a professor in the School of Computing and Augmented Intelligence at Arizona State University in Tempe, Arizona. “We have developed an AI system to identify Alzheimer’s disease at an early stage and in an inexpensive, scalable way — inspired by the vision of seeing the brain through the eyes.”
Seeking scalability
Almost 7 million Americans are living with Alzheimer’s disease. That number is expected to double in coming decades. Diagnosis typically occurs after symptom onset, when existing therapies tend not to be effective. Earlier diagnosis of Alzheimer’s disease would allow patients to receive treatment and implement lifestyle changes that might slow the disease’s progress.
The first AI-based algorithm for disease screening approved by the Food and Drug Administration was retinal imaging coupled with AI to screen for diabetic retinopathy. Dr. Dumitrascu and colleagues set out to investigate whether AI-assisted retinal imaging can screen for neurodegenerative disorders.
In an exploratory trial, the researchers found that increased supratemporal retinal amyloid-beta deposits in the distal perivenular regions can differentiate between normal and impaired cognition and are inversely associated with hippocampal neurodegeneration. That study was published in Acta Neuropathologica Communications.
“But identifying abnormal proteins in the retina involved a complicated ocular imaging system. We realized that this approach may not be scalable anytime soon,” Dr. Dumitrascu says. The researchers turned to AI, publishing their findings in Mayo Clinic Proceedings: Digital Health.
“With the help of Dr. Wang and his team, we now see that an AI model applied to the much simpler nonmydriatic color fundus photography can accurately distinguish patients with Alzheimer’s disease from non-Alzheimer’s population,” Dr. Dumitrascu says. “The next challenge is translating this work into a preclinical Alzheimer’s population.”
To learn more, the researchers are testing their system among patients at Mayo Clinic as well as at rural and mobile clinics in Arizona. The aim is to obtain data, with patient permission, from demographically diverse populations.
Dr. Wang notes that 42% of retinal images the researchers obtain in a rural setting are of too poor quality to provide information. “But with our AI tools, we can rescue roughly 57.7% of those unusable images and turn them into usable images,” he says. “Our model can find subtle Alzheimer-related changes in the retina in low-quality images collected in a real-world setting.”
Rural screening also includes blood collection, low-field brain MRI, cognitive testing and simple genetic testing. “We are using the participants’ retinal images to provide a quantification of their vascular health as well as to correlate the retinal imaging findings with more-established biomarkers of early neurodegeneration,” Dr. Dumitrascu says. The researchers also are using data from electronic health records and agentic AI to help build the model.
Early diagnosis of Alzheimer’s disease would help not only to guide patient care but also to screen patients for clinical trials. “We need to find better ways to conduct clinical trials in neurodegeneration,” Dr. Dumitrascu says. “That means accurately identifying patients early on — differentiating Alzheimer’s disease from other pathologies — and also having an accurate modality to monitor patients during the treatment being studied. One of the most important applications of this model will not necessarily be to tell patients that in two decades they might develop Alzheimer’s disease but to enroll patients in preventive clinical trials.”
The researchers hope the model can ultimately provide cost-effective information for patients. “Ideally, patients would have this type of scan in a familiar setting,” Dr. Dumitrascu says. “The results would be standardized, reliable and connected to the appropriate follow-up care.”
For more information
School of Computing and Artificial Intelligence. Arizona State University.
Dumitrascu OM, et al. Retinal peri-arteriolar versus peri-venular amyloidosis, hippocampal atrophy, and cognitive impairment: Exploratory trial. Acta Neuropathologica Communications. 2024;12:109.
Dumitrascu OM, et al. Color fundus photography and deep learning applications in Alzheimer disease. Mayo Clinic Proceedings: Digital Health. 2024;2:548.
Refer a patient to Mayo Clinic.
SOURCES: