Researchers from Skoltech, the Z-union AI Technologies Consortium, Sber’s Practical Artificial Intelligence Center, and other research organizations have proposed an artificial intelligence-based approach for the preventive screening of 15 diseases based on fundus photography — medical images of the rear of the eye. The conditions whose risk the researchers were able to predict from these images include not only eye diseases such as glaucoma and cataracts, but also systemic ones: hypertension, lupus, AIDS, and more. Adopting approaches like this would enable patients to learn about a disease earlier, before symptoms become noticeable. The study was published in the journal Frontiers in Medicine.

“Invasive tests aren’t performed without a specific indication, but noninvasive screening can be done at scale,” explains the study’s lead author and Z-union AI Technologies Consortium CEO Robert Vasilev. “Suppose that a person visits an eye clinic with an eye pain complaint. They’ll be given a fundus photograph as a matter of course — it’s standard procedure. That same image carries information about the risk of a whole range of diseases, including ones that have nothing to do with the eyes or vision. And the sooner that patient learns they’re at risk for one of them, the better.”

The model the researchers built takes a fundus image as its input and outputs the probability of each of the 15 diseases being present. The prediction accuracy is given by a metric technically known as ROC AUC. While the score of 1.0 represents a perfect result, the model’s performance is 0.997, meaning that it is excellent at distinguishing healthy individuals from patients with diseases, including rare ones.

To train the model, the study’s authors compiled, annotated, and anonymized a unique dataset of more than 20,000 fundus images, collected from open sources and acquired from clinics. A distinguishing feature of the dataset is the representation of rare pathologies, making it a valuable resource for future research. The team formalized the medical task and enlisted medical experts to define the image-based markers for each target condition.

Study co-author Yuliya Sarana, a research scientist at Skoltech Biomedtech, noted: “We hope that our proposed noninvasive approach will find its way into medical practice and, combined with existing diagnostic methods, will make it possible to detect serious diseases at earlier stages. That, in turn, could help ensure timely treatment, extend healthy lifespan, and improve quality of life.”