Dr. AI-hurts: the Russian neural network has surpassed the accuracy of Microsoft's AI diagnostics

The Russian AI health assistant based on the GigaChat model has surpassed a similar Microsoft development in diagnostic accuracy. Such results were obtained during tests in which both neural networks needed to identify the disease based on real-world case data. The domestic development achieved 93% accuracy, and the American — 85%. Today, dozens of AI-based medical systems are being created and used in our country, which are becoming more and more advanced. According to experts, their use makes it possible to reduce the cost of diagnosis and minimize the number of tests.

Medical diagnostics using AI

Created by the AI Institute's specialists, the AI health assistant based on the GigaChat neural network model demonstrated 93% accuracy of general medical diagnostics during the experiment. Earlier, a similar service from Microsoft in a similar study showed a result of 85%, the creators of the Russian product reported. The AI needed to determine pathology based on real-world case data from the reputable New England Journal of Medicine.

The Russian AI worked blindly, starting with basic patient data— gender, age, and symptoms. Next, the model consistently requested tests, imaging, and consultations to arrive at a diagnosis. On average, three actions were enough for the car to diagnose: receiving information from the patient — a comment or a question from the doctor — the patient's response.

All clinical cases were marked with three types of complexity. Russian-developed artificial intelligence has successfully recognized such rare pathologies as Whipple's disease, aceruloplasminemia and methemoglobinemia caused by the rasburicase protein.

— Today, multi-agent systems are able to confidently search for rare, masked pathologies beyond the typical algorithms of the admission department. The AI assistant showed that it was able to rebuild a hypothesis in time, rejecting the most likely but false path, and did it faster than an expert observer with 15 years of experience in medicine expected," Ivan Oseledets, Director General of the AIRI Institute, told Izvestia.

In concert with sounds: AI will customize children's hearing implants in minutes

How a neural network will help speed up the adaptation of young patients with hearing loss

The experiment is of a pilot nature and is currently ongoing. AIRI researchers plan to expand the sample to include cases from other medical journals. However, it is already clear that the technology can be used not only for practical medicine, but also for training doctors, offering them realistic simulations of complex cases.

— We see how the multi-agent architecture accelerates and improves the diagnosis process. It is especially important that the system demonstrates flexibility: it reviews hypotheses, requests additional data, and even responds to the emotional presentation of clinical scenarios. In the future, this opens the door to consultations where AI will become a reliable physician's assistant," said Sergey Zhdanov, Director of Sberbank's Center for the Health Industry.

The AI assistant is undergoing pilot testing in the application of the Sberbank healthcare company. At the moment, it has already been used more than 160 thousand times.

AI assistant from the AI Institute is one of many artificial intelligence-based systems that are being created and operated in Russia. There are dozens of them now. The capabilities of neural network technologies are used to analyze medical images and medical records. Some services are patient-oriented and help everyone who wants to maintain their personal health. Others are intended for professionals, for example, to improve laboratory diagnostics. There are universal programs for combating pathologies of various organs and body systems, as well as specialized solutions for dentistry, ophthalmology, neurodegenerative and other diseases.

How AI is used for medical purposes in Russia

According to Evgeny Popov, head of the healthcare department at the Yandex Cloud Technology Center for Society, the most common area of AI application in medicine is computer vision in radiation diagnostics. This area is the most standardized in terms of data and processes, they are well digitized. The company has implemented a project in this area to identify fetal pathologies during pregnancy together with the Kulakov National Research Medical Center.

— Scripts based on large language models (LLM) are also becoming in demand. For example, our center has developed a service based on YandexGPT that speeds up the processing of documents during clinical trials, including to improve the quality of diagnosis and develop new treatment methods. The solution is already being applied at the N. N. Petrov National Research Medical Center of Oncology," said Evgeny Popov.

As the press service of MWS AI, a company that creates software products based on neural networks, told Izvestia, the most promising option for Russian developers is to create AI assistants for solving applied problems in medicine. Another area is speech analytics. With the help of AI, you can record conversations between a doctor and a patient, automatically decrypt them and fill out the necessary documents. This will reduce the bureaucratic burden on doctors.

Russia has a chance to become a world leader in the field of medical image recognition, as our country is traditionally strong in the field of computer vision, where neural networks analyze videos and images, the company added.

According to practitioners, AI assistants can be a good help in their work, provided that several conditions are met.

— Any AI is only as good as the correct information is embedded in it and how correctly it was asked a question. Such systems make it possible not to forget or review anything, so that the doctor makes an informed decision in a situation where there is a lot of documentation and other data," said Andrey Prodeus, chief pediatric allergist and immunologist at the Ministry of Health of the Moscow Region.

As Maxim Kolyasnikov, senior lecturer at the Department of Future Technologies at the Moscow Institute of Physics and Technology, noted, the main advantage of using modern AI systems in diagnostics is the ability to predict research costs and minimize them by eliminating excess ones at earlier stages. In the end, we are getting closer to the ideal, where an accurate diagnosis and proper treatment will be provided in the shortest possible time and with minimal cost, he stressed.

At the same time, one should not rush to make serious generalizations based on a comparison of the two studies, concluded Stanislav Stragnov, head of the Laboratory for the Analysis of public health indicators and digitalization of healthcare at MIPT.