What IT companies should take into account when they implement AI in health care #Interview

How to accelerate development of the AI market in health care and what should be taken into account when AI-based solutions are implemented for visual data analysis in health care? Elena Sokolova, SberMedAI chief product officer, told Medvestnik about this.

AI-based services have already been used in Russian health care for some time. However, there are some problems which need solution to accelerate AI market development in health care. Today we will discuss the specifics of development and implementation of AI-based solutions to analyze visual data.

Firstly, there are barriers related to the quality of data used for training AI models, and to the fact that analysis of images by experts is costly. 

Creating high-quality and sought-after AI-based services for health care is possible only in close cooperation with the medical community. As part of the dialog with doctors, tasks are set, and them it is determined which tasks can be tackled using the existing technologies.

Model training is similar to the process when more experienced doctors train young colleagues. They show what  a particular disease looks like in images of a particular patient. They prompt young doctors and correct mistakes, and therefore share their expert knowledge.

To train a machine, one needs to collect a necessary number of examination results placed by medical experts. This is a hard and time-consuming process, which requires integrating expert knowledge of doctors and developers. SberMEdAI’s employees include radiologists, specialists in functional diagnostics, physicians, cardiologists, and each AI-based solution is supervised by a particular medical product owner.

Regional medical information and analytical centers have already accumulated large amounts of visual data. Besides, experimental regulatory frameworks for using such data are being rolled out in other regions based on the successful implementation in Moscow. Direct imteraction between startups and medical institutions is another solution.

To make it easier for developers to come onto the medtech market, reference datasets for training and testing AI-based solutions are created. For instance, in 2022, at mosmed.ai/datasets the Moscow Government provided access to 40 datasets of anonymized medical data, and Sber provided  MedBench, a free platform for solving AI-related tasks in health care. The platform website allows any person to download labeled datasets to build AI models.

Secondly, the fact that a domain is complex and that itis difficult to identify real needs of those who are involved in the process is a considerable barrier.

At this stage it turns out that an algorithm can be developed. However, it is very dfficult to implement it to ensure that it meets all user interests. ML specialists usually focus on product development rather than on technology application domains. It is important to communicate with a sufficient number of potential algorithm users and discuss issues, and find effective solutions for implementation.

There may be several scenarios for implementing the same model (depending on real needs of an end user): the first or second opinion, proiritizing examinations by case complexity, mass screening and “filtering” the normal, restrospective analysis or prospective study in real time. Therefore, the developer should not only offer an AI-based service, but also consider how to embed it in the existing business processes and the IT infrastructure of each region, and customize the product to ensure that it improves performance of a particular “link”. To achieve this, we should not forget about issues and performance metrics which exist in a particular region or health care institution.

As the technology is new, there is little experience in tackling standard tasks. Currently there is a lack of common standards in terms of infrastructure solutions. These issues can certainly be dealt with, but it is very important how proactive developers are and what appproaches they use. SberMedAI gives great attention to customer development when medical services are created.

Finally, doctors lack confidence in technologies.

Doctors are extremely conservative. This is related mainly to a high level of education, training and the domain itself, where errors can be fatal. Loyalty and confidence in new technologies are based on active and long-term popularization of AI-based technologies and support of the industry from the government. Here we see positive changes: in late January President Vladimir Putin issued an instruction for ensuring application of examination results obtained using AI in clinical guidelines and the mandatory health insurance system to accelerate development of AI-based technologies in health care.