Sberbank: System for modeling and predicting diagnoses

Main article: Data mining Data mining

2023: Creation of a system for modeling and predicting diagnoses

Sberbank's Laboratory for Artificial Intelligence (AI) has developed a medical solution that simulates the health of patients according to their electronic medical records. Alexander Vedyakhin, First Deputy Chairman of the Management Board of Sberbank, spoke about this on November 22, 2023 .

The bank's specialists have adapted the Transformers neural network architecture to work with patient diagnosis sequences. This approach made it possible to obtain a formal representation of medical profiles of patients in the form of embeddings (embedding, numerical vector of features), useful for many scientific and practical problems.

This solution opens up opportunities for modeling tracks of further development of patient conditions. This property of the model is tested in the task of predicting the next diagnosis and confirmed by high metrics in an experiment on the public medical record datacet MIMIC-III (open database with information about patients admitted to intensive care units of a large medical center).

{{quote 'author = said Alexander Vedyakhin, First Deputy Chairman of the Management Board of Sberbank. | This is not the first breakthrough MedTech solution of Sberbank and our partner companies. In Russia, AI services developed by Sberbank are increasingly used: computer vision for decoding medical images (CT, X-ray, mammography), automatic filling out of medical documentation (from voice to text), diagnosis of diseases and others. We are also exploring the possibilities of prognostic models that work with medical data. As practice shows, such solutions can be in demand not only in medicine, but also in related areas. This is another contribution to the preservation of human lives and the development of health care,}}

The Sberbank model has the potential for application in insurance scoring. With this model, it will be possible to obtain a significant increase in the accuracy of assessing insurance risks relative to traditional methods, and these are new opportunities for personalizing tariffs and reducing costs. As of November 2023, the model works with historical data, the next stage will be to train it to work online with data received from customers to make a decision at the time of appeal.

Also in the course of the study, an H2D method (Harbinger Disease Discovery) for searching for harbingers of diseases was proposed, which allows automatically generating hypotheses about the relationships between diseases. So, thanks to him, it was possible to find a strong dependence between a group of psychological disorders and breast cancer in women. The validity of this hypothesis is confirmed in related scientific research. As a result, scientists have formed a new set of hypotheses about the harbingers of the five most common types of cancer. The H2D method will help the medical and scientific community in finding new areas of medical research.