Published on: 2026-06-02
Source: Novosibirsk State University –
An important disclaimer is at the bottom of this article.
Employees of the Artificial Intelligence Center of Novosibirsk State University have patented a method that helps automatically identify the most important symptoms and indicators for predicting diseases in electronic medical records. The new method is already being used to “train” the decision support system “Doctor Pirogov” so that it can more accurately recognize symptoms and assess risks for patients.
The development relates to the field of medical informatics and the analysis of large arrays of medical data. Essentially, this is an algorithm that “scans through” thousands of de-identified patient records with the same diagnosis and looks for dozens and hundreds of fields of those features that are most commonly found in patients with this disease and most strongly influence the diagnosis.
—Then, using this method, we determine which symptoms are more informative for the given diseases and use it in the medical decision support system., — explained the head of projects of the Artificial Intelligence Center at NSU, head of the II laboratory and the largest genetic data center of ICIG SB RASVladimir Ivanisenko.
The technical method works as follows: first, a multitude of electronic medical records with the same diagnosis are brought to a uniform form and converted into a set of binary features (there is a symptom or anomaly in the analysis — one, no — zero). Then, these data are used to train a special type of neural network — the so-called Concrete Autoencoder, which uses the Gumbel-Softmax mechanism. Its feature is that it not only produces a final prediction but also selects specific features that made the greatest contribution to the result.To filter out random effects from random initialization, the model is run multiple times, and then by the frequency of “selection” of features, a stable set of prognostically significant symptoms and indicators is identified.
Such an approach increases the reliability of automated processing of medical data and, consequently, the accuracy of diagnosis when using decision support systems.
One of the key barriers to implementing AI in medicine, specialists call the difficulties with interpreting neural network outputs: the algorithm gives an answer, but the doctor does not understand why the system “decided” exactly that way. This reduces trust and makes using such solutions risky.
—Neural networks usually give some result that is not justified, not interpreted, or seen from how they obtained it. Otherwise, it is like with this type of autocode systems. They allow for indicating specific signs and symptoms of the patient that contribute the most to the correct diagnosis.— told Vladimir Ivanisenko.
Thus, the development of NGU solves two tasks at once: it remains within the framework of modern artificial intelligence methods, but at the same time makes the results more understandable for doctors. The doctor sees which specific symptoms and test deviations the system “highlighted” as the most significant for this particular disease, and can compare this with their own clinical experience.
The team has now completed the patenting of the method and is engaged in fine-tuning the system. The developers plan to present the results at the St. Petersburg International Economic Forum. According to Ivanisenko, “with the help of our system, we have indeed taught ‘Doctor Pirogov’ to better recognize symptoms.”
Although the development is used for configuring a specific artificial intelligence system of NGU, it is not tied only to “Doctor Pirogov.” It is about a fairly universal approach to working with electronic medical records, which can be integrated into various systems that support decision-making by a doctor.
A separate direction is the creation of risk models that assess the probability of disease development in a patient based on their data.
—This is, in particular, exactly what our system does: it determines the risks of diseases based on the presence of certain symptoms and indicators. For any system that tries to assess risks on the basis of symptoms, it is critical to understand which signs to pay attention to, which of them are the most significant,— noted Ivanisenko.
In the future, such methods can help make digital medicine more reliable and understandable: algorithms will not only provide predictions but also explain which clinical signs they are based on. This is important both for doctors making decisions and for patients who need to trust new technologies.
Please note; This information is raw content obtained directly from the information source. It is an exact report of what the source claims and does not necessarily reflect the position of MIL-OSI or its clients.