Sber presents its AI-powered healthcare solutions at ICML

Jul 20, 2021

Sber AI Laboratory has presented two AI-powered solutions at the 38th International Conference on Machine Learning (ICML): a model that identifies pneumonia sources in lungs through X-rays with further prioritization of patients who must receive treatment and a model that evaluates the risks of severe symptoms in patients hospitalized with pneumonia, including when it is caused by COVID-19. Sber’s AI for Health neural network models have been designed with medical doctors. Now that the models have been adopted, doctors use them a lot in their work.

The model that identifies pneumonia sources on X-rays has reached the accuracy of 97.8% due to its continuous training and fine-tuning. Medical doctors in the city of Tomsk have uploaded 450 chest X-rays for the project. On top of that, the model has been trained using 350 X-rays of healthy patients, and approximately 50 X-rays – both with and without pathologies – have measured the quality of training. The cases when X-rays marked by human doctors and the model had non-empty overlapses were considered as the correctly labeled ones.

The risk assessment model to predict severe symptoms for pulmonology patients uses information documented in the electronic health record by the physician within the first 24 hours of hospitalization. Based on the data, the system assigns one of two high severity risk levels (ranging from 29% to 100%). Timely identification of high risk patients enables doctors to adjust treatment plans and reduce hospital mortality. The project was successfully implemented at Krasnoyarsk Regional Clinical Hospital in spring 2021. The proportion of high risk patients who were admitted to intensive care decreased by 14.3% and the proportion of cases of mortality dropped by 44.3%.

AI Lab developers have been designing medical solutions based on machine learning and computer vision for several years now. On the AI for Health track, Sber collaborates with medical institutions in both Moscow and other regions of Russia. We have seen how the value of such projects has grown in the last year and a half and how they simplify doctors' heavy workloads. Presenting our AI for Health projects at ICML is a big milestone in the recognition of Sber's developments by the international data researcher and data engineer community.

Alexander Vedyakhin

First Deputy Chairman of the Executive Board, Sberbank

ICML is a major conference that greatly influences research on machine learning and artificial intelligence. It is a globally renowned platform that brings together the world’s leading experts in the fields of DS/AI. The first ICML was held in 1980. In 2020, 11,000 professionals from 75 countries gathered for ICML.