Scientists have linked changes in the retina to the risk of Alzheimer's disease

A typical retinal photograph may contain information about a person's age, vascular condition, and lifestyle—characteristics associated with the likelihood of developing Alzheimer's disease. Scientists expect that in the future, analyzing such images using artificial intelligence will help identify people who need an in-depth examination long before cognitive impairment occurs. About what changes the algorithm found in the eyes of the study participants, how reliable the method turned out to be, and why it cannot yet be considered a diagnostic test, see the Izvestia article.

What the retina can tell you about brain health

The retina is a layer of nervous tissue on the inner surface of the eye. Her condition depends not only on ophthalmological diseases, but also on the functioning of the vascular, nervous and metabolic systems of the body. Therefore, scientists consider the fundus as a relatively accessible source of information about the processes that can occur in the brain.

Unlike complex neuroimaging techniques, color photography of the fundus is done quickly and does not require intervention in the body. Such images are already widely used during the examination of patients with diabetes, glaucoma and other diseases. The researchers suggest that the same images can be further analyzed to assess systemic risk factors.

How scientists trained artificial intelligence

The authors analyzed 62,876 color photographs of the fundus obtained from 44,501 UK Biobank participants. This database contains medical, biological, and behavioral data from more than 500,000 UK residents. The initial information of the participants was collected between the ages of 40 and 69.

According to the section describing the research methodology, the scientists first checked the image quality and excluded images that a special algorithm found unsuitable for analysis. The photographs of the right eye were mirrored so that the position of the anatomical structures corresponded to the images of the left eye.

After processing the data, the participants were distributed among the samples for the development and verification of the model in a ratio of 80 to 20. This separation was carried out at the human level so that the images of one participant could not simultaneously appear in the training and verification samples.

Artificial intelligence was trained to recognize 12 characteristics based on the state of the retina, which the authors associated with the risk of developing Alzheimer's disease. These included gender, age, smoking, alcohol consumption, sleep problems, recurrent depression status, income level, age of completion of full-time education, body mass index, systolic and diastolic blood pressure, and glycated hemoglobin HbA1c levels. The latter indicator allows you to estimate the average blood glucose over a long period of time.

The participants provided information about smoking, alcohol consumption, sleep problems, education, income, and mental state on their own. Blood pressure and body mass index were measured during the examination, and HbA1c and genetic sex were determined based on laboratory data.

Medical records do not always contain complete information about the patient's lifestyle, the authors noted. In addition, people may not accurately estimate the frequency of alcohol consumption, sleep duration, or other habits. In the future, a retinal image can become an additional objective source of data, since the state of blood vessels and nervous structures can reflect the combined effect of various factors on the body.

Vision is under attack, sleep is on the verge: what diseases await zoomers

Constant smartphone use can lead to early development of cataracts.

Which parts of the eye did the AI pay attention to?

To understand the logic of the model, the scientists applied significance maps. They show which areas of the image most strongly influence the output of the algorithm. The most important were the arteries and veins of the retina, the disc of the optic nerve and the excavation located in it. These structures can change with age, with high blood pressure, metabolic disorders, and prolonged exposure to bad habits.

When determining the gender, the model mainly analyzed the shape of the optic disc and the excavation. When recognizing a history of smoking, she also paid more attention to the area of the optic nerve outlet. Certain features of the arteries and veins, including their width, density, and complexity of the vascular network, were associated with alcohol consumption, sleep disorders, and other indicators.

In addition, the algorithm evaluated the curvature of blood vessels, the ratio of the diameter of arteries and veins, and the density of their branching. However, he recognized the complex characteristics of tortuosity worse than the general shape of the optic disc and the width of the vessels.

Some of the patterns found cannot be considered specific to Alzheimer's disease. Retinal vessels change under a large number of conditions, from hypertension and diabetes to natural aging. Therefore, the algorithm can record the total accumulated vascular and metabolic damage, rather than the early manifestations of a specific neurodegenerative disease.

What was discovered a few years before the diagnosis

The authors devoted a separate part of the work to participants who were diagnosed with Alzheimer's disease after receiving retinal photographs. As stated in the full text of the study, 52 people were included in this group. Patients with vascular and frontotemporal dementia were excluded from it in order to leave only cases of Alzheimer's disease without concomitant types of dementia. For comparison, the researchers selected 52 more participants without the disease, comparable in basic demographic and medical characteristics.

The scans were taken an average of 8.55 years before Alzheimer's disease was detected. The interval between retinal photography and diagnosis ranged from 2.38 to 11.41 years. Some indicators of the significance of retinal regions differed between future patients and the control group. According to the authors, this may indicate an intersection of changes associated with known risk factors and processes accompanying the preclinical stage of the disease.

Life by Touch: how blind people find their way in the world

And who helps them adapt to the new realities?

However, the researchers themselves warn that this analysis was retrospective and exploratory in nature. It does not prove that the model is capable of predicting Alzheimer's disease in advance. The algorithm was trained to recognize risk factors, rather than determine who would later develop the disease. The authors explicitly point out that the result obtained cannot be interpreted as evidence of the predictive ability of the model in relation to future cases of the disease.

In addition, only 52 future patients and 52 people from the control group participated in the additional analysis. This sample size is insufficient for the development and clinical validation of a full-fledged prognostic test.

How the method can be used in the future

The main potential advantage of the technology lies not in making a diagnosis, but in pre-selecting patients. If further research confirms its reliability, the images already obtained during a routine ophthalmological examination can be further analyzed using artificial intelligence.

The algorithm could detect combinations of signs that require a doctor's attention. After that, the patient could be offered blood pressure and glucose measurements, cognitive testing, consultation with a neurologist, or studies of specific biomarkers. This approach may be useful for mass screening, since fundus photography is more accessible and cheaper than MRI and positron emission tomography.

Scientists have previously tried to recognize the disease by the structure of retinal vessels. In a paper published in Scientific Reports in 2021, a machine learning model distinguished between images of patients with Alzheimer's disease and a control group with an average accuracy of 82.44%. Small vessels turned out to be the most informative then.

However, such results so far relate to research models, and not to diagnostic procedures available to patients. The accuracy obtained from a limited sample cannot be automatically transferred to the general population or to the operation of the system in a conventional clinic.

The new work develops this direction: instead of trying to detect an existing disease, its authors focused on characteristics that may be associated with the vulnerability of the body years before the onset of symptoms. This approach potentially allows us to shift attention from late diagnosis to early assessment of risk factors.

To apply the technology in clinics, larger longitudinal studies will be required, testing on independent and ethnically diverse samples, as well as comparing the results with confirmed biomarkers of Alzheimer's disease. Until that time, the condition of the retina can be considered only as one of the possible sources of additional information about the health of blood vessels and the nervous system, but not as a way to independently determine the likelihood of a future disease.