- Статьи
- Science and technology
- Mind in wards: Scientists have developed a system for using AI in intensive care units across the country
Mind in wards: Scientists have developed a system for using AI in intensive care units across the country
Russian scientists are creating the first public data set for intensive care units, designed to train artificial intelligence algorithms. It will become the foundation for domestic AI solutions in intensive care and will allow for more accurate identification of patients in need of emergency care. A key element of the development was "clinical phenotypes", a classification that will replace limited and not always accurate disease codes. This will allow artificial intelligence to be trained on the real conditions of patients, rather than on documentation data, and will significantly improve the quality of care for patients in critical condition. For example, the development of sepsis can be predicted six hours before it occurs, doctors told Izvestia.
AI training for intensive care units
The National Medical Research Center (NMIC) in the field of "anesthesiology and intensive care (for adults)" at Sechenov University became the basis for the development of the first national structured data set of an intensive care patient. It will make it possible to more effectively train domestic AI solutions for intensive care and intensive care units (ICU) throughout the country.
According to doctors, the main problem of modern AI systems in intensive care is dependence on incomplete or missing ICD-10 codes (International Classification of Diseases tenth revision). In the NICU, less than 30% of cases of critical conditions receive a code — this is a global practice, the university told Izvestia.
The new dataset solves this problem by introducing "clinical phenotypes" — algorithmic identification of pathophysiological conditions based on objective indicators: vital functions, laboratory data and dynamics of the patient's condition. More than 80 such phenotypes were identified in the data set based on 5.3 thousand cases, including sepsis, acute respiratory distress syndrome, acute renal failure and other critical conditions.
— In the process of creating the dataset, developers had to overcome the barrier of inaccuracy of data based on ICD codes. To solve this problem, the concept of clinical phenotypes was introduced. These are algorithms that identify pathophysiological conditions based on objective criteria such as vital signs, laboratory test results, and patient condition dynamics. Thanks to this, AI models are trained not on documentation artifacts, but on real clinical conditions of patients," Natalia Zhivitsa, an analyst at the Department of Educational Program Analysis and Scientific Research at Sechenov University, told Izvestia.
In her opinion, this radically changes the paradigm of working with data in intensive care and opens up new horizons for creating solutions that actually work in the NICU.
The project is being implemented jointly with the industrial partner QuattroLab, the creator of the intensive care and anesthesiological information System (RAIS). The effectiveness of the proposed approach is confirmed by the results of the development of a machine learning model capable of predicting the development of sepsis six hours before the appearance of clinical signs. This result was made possible by using clinical phenotypes and appropriate algorithms that allow the model to be trained on well-structured data. In particular, this eliminates the so—called data leakage from the future (data leakage) at the training stage and ensures the correct formation of samples - with a clear separation of clinical cases by the presence or absence of sepsis.
"Sepsis is one of the most dangerous conditions in intensive care with a high mortality rate," Irina Larina, head of the RAIS analytics department, told Izvestia. — Without a qualitatively new approach to data implemented in RICORD, the creation of similar solutions remains extremely difficult.
A step towards the medicine of the future
The creation of the first such dataset is not just a scientific project, but a strategic investment in technological sovereignty and the quality of critical care medicine of the future, said Andrei Yavorovsky, Director of NMIC.
"We are laying the foundation on which the future of artificial intelligence in critical medicine in Russia will be built,— he stressed.
AI is already in demand in anesthesiology and intensive care, and the appearance of a public dataset opens up new opportunities for practical use, said Inna Trukhanova, Head of the Department of Anesthesiology, Intensive Care and Emergency Medicine at the Samara State Medical University, market expert at NTI Helsnet.
— First of all, this is an opportunity for an earlier response to critical incidents intraoperatively, when there are no obvious manifestations either clinically or on monitors. In addition, AI is able to improve the accuracy and speed of assessing the patient's condition in intensive care through accurate and fast integral scoring. In anesthesiology, technologies can be used to select individual anesthesia regimens, taking into account concomitant diseases, test results and medications taken," she explained.
Intensive care clinical data is one of the most technically complex types of data for training neural network models, said the head of the AI department. Cloud.ru Dmitry Yudin. The high frequency of measurements, heterogeneous sources and a significant number of omissions create serious requirements for the quality and structure of the training sample. In this sense, the approach with clinical phenotypes — when the patient's condition is described through objective parameters rather than formal classification codes — methodologically corresponds to the specifics of the task.
— The development of domestic domain datasets is a necessary condition for the emergence of applied AI solutions in medicine. Stable models cannot be built without a high-quality and representative database. For tasks of this class, the volume of the dataset is of fundamental importance: 5.3 thousand cases is a meaningful starting point, but scaling the database remains a prerequisite, especially for rare critical conditions," the expert said.
In modern conditions, any verified data is valuable for the development of medical AI solutions, said Nikita Nikolaev, marketing director and co-founder of the company that creates medical decision support systems for Celsus, a participant in the NTI Helsnet market. Such a set will be useful to developers at least for validation, that is, checking the quality of the model metrics. In addition, it can also be used for AI training, despite its relatively small volume.
Переведено сервисом «Яндекс Переводчик»
