AI Assesses the Risk of Hypertension and Other Systemic Diseases From Eye Fundus Images

Researchers from Skoltech, the Z-union AI Consortium, Sber AI Lab and other scientific institutions have developed an artificial intelligence-based approach for preventive screening of 15 diseases using retinal fundus images.

The conditions that the researchers were able to predict from eye images included not only eye diseases such as glaucoma and cataracts, but also systemic conditions, including hypertension, systemic lupus erythematosus and AIDS.

The researchers say that implementing such technologies could help identify health risks at an earlier stage, including in people who have not yet developed noticeable symptoms. The study was published in Frontiers in Medicine.

“Invasive tests are not performed without specific indications, whereas non-invasive screening can be carried out on a large scale,” explained Robert Vasiliev, the study’s first author and director of the Z-union AI Consortium, as reported by Nauka Rossiya.

“For example, a person may visit an ophthalmology clinic because of eye pain. A fundus image will routinely be taken as part of the examination. That same image contains information about the risk of developing a range of diseases, including conditions that have nothing to do with the eyes. The earlier a patient learns that they are at increased risk of one of these diseases, the better.”

The AI model takes a fundus image as input and estimates the probability of each of the 15 target diseases. Its overall performance, measured by the ROC AUC, reached 0.997, including for rare conditions. On this scale, a score of 1.0 represents a perfect result.

The researchers therefore say the model demonstrates a very high ability to distinguish healthy individuals from people with disease based on retinal images.

To train the system, the research team assembled, annotated and anonymized a unique dataset containing more than 20,000 fundus images collected from public sources and obtained from medical clinics.

One of the dataset’s distinguishing features was its inclusion of rare diseases, which the researchers say could make it a valuable resource for future studies. With the help of medical experts, the team also formalized the clinical task and identified imaging features associated with the target conditions.

Yulia Sarana, a research scientist at Skoltech’s Center for Bio and Medical Technologies and a co-author of the study, said the team hopes the approach will eventually be used in clinical practice.

“We hope that our proposed non-invasive approach will find applications in medical practice and, together with existing diagnostic methods, help identify serious diseases at earlier stages,” Sarana said. “This, in turn, could contribute to starting treatment in a timely manner, extending the period of healthy life and improving its quality.”

Follow NEWS.am Medicine on Facebook and Twitter