The Art of making decisions
About AI in medicine ― Corresponding Member of the Russian Academy of Sciences Valeria Gribova
Yanina Khuzhina, "Scientific Russia"
What qualities of natural intelligence should artificial intelligence have in medicine? How is the knowledge base replenished and what problems prevent the introduction of AI into medical practice? This is our conversation with Valeria Viktorovna Gribova, Corresponding Member of the Russian Academy of Sciences, Doctor of Technical Sciences, Deputy Director for Scientific Work of the Institute of Automation and Control Processes of the Far Eastern Branch of the Russian Academy of Sciences.
― Why is artificial intelligence needed in medicine in the first place?
― We consider it primarily as a support system for medical decision―making - not replacing the doctor, but helping him to make the right choice. The fact is that for any doctor to make decisions, it is necessary to take into account a huge number of factors: the patient's anamnesis and his complaints, possible contraindications to treatment, compatibility of medications, etc. It is objectively difficult for the human brain to cope with such a volume of information, and mistakes are inevitable here.
Artificial intelligence in medicine ― this is not a substitute for a doctor, but his assistant, if you want, an intelligent assistant.
― A person has such abilities as knowledge integration, reflection, learning, etc. Which of them are primarily necessary for such an assistant?
― Any AI system simulates the abilities of natural intelligence: basically the very abilities that you have listed, including the ability to reason by analogy, draw conclusions based on knowledge, the ability to learn, argue and explain their decisions, etc. And, of course, when we talk about practical medicine, all these abilities of natural intelligence the artificial intelligence system should also be inherent in order to help the doctor most fully and well when making decisions.
― Is it possible to create a system that is intuitive to any doctor, or do you need to undergo special training to work with AI?
― In clinical practice, there are already medical information systems where doctors enter all information about patients. If AI-based medical decision support systems are integrated with such medical information systems, then, in fact, nothing special will be required from the doctor. The process may look something like this: the doctor enters information into the system in his standard mode, which then gives tips and advice; guided by them, the doctor makes a decision. At the same time, it is very important that such systems can justify this or that decision, because the responsibility for it in any case lies with the doctor.
The system should explain in detail why this or that diagnosis was made, this or that treatment was chosen.
― Then for such a system of medical decision-making, some kind of uniform terminology, uniform descriptive standards should be approved?
― Yes, of course. Terminology forms the basis for an unambiguous interpretation of all observations and prescriptions, therefore medical reference books should certainly be created. There are various medical terminology reference books abroad that doctors use in their professional activities. Unfortunately, there are no such full-fledged reference books in Russia, but the process of their development for various branches of medicine has already begun. We expect that eventually such full-fledged reference books will be actively used in medical practice. Indeed, everything starts with terminology, and then, using it, we form knowledge bases and data that will be unambiguously interpreted by the entire medical community. It's only a matter of time.
― Should the AI programs that are planned to be implemented in our medical system be self-learning? How is the knowledge base replenished in such a program?
― All AI systems can be divided into two large classes. Systems using big data, or so-called data sets, have already been introduced into medical practice and are used to analyze medical images, thanks to which it is possible to establish what is in front of us: covid, tumor, etc. If it comes to clinical data, that is, complaints, anamnesis, then it is still very difficult to learn from the available data. The data I am talking about are medical histories, and they really are of very poor quality, and therefore the result of such training will be quite weak. There are other approaches to creating a knowledge-based AI system. The first is when edits to the system are made by experts with great authority and unique knowledge.
By introducing their knowledge into the system, the expert actually replicates them in this way, and this is very cool. By the way, if earlier doctors were not so willing to scale their knowledge, today the situation has changed and specialists are increasingly sharing their experience.
Often, doctors themselves come to us with a proposal to contribute their knowledge to the system. Another way to replenish the knowledge base is clinical recommendations, when we can create formalized knowledge bases based on text analysis, but even in this case, an expert level is required who could refine this knowledge. The third option for replenishing the knowledge base is the formation of knowledge based on training choices: the so-called inductive generalization of data. When we inductively generalize data based on a training sample, we can see some new interesting and previously unknown dependencies. Therefore, when we talk about systems based on knowledge bases, we actually use the three approaches listed above to create them, the key of which, of course, is expert formation when maintaining the knowledge base.
― One of the main functions of AI in medicine ― this is the construction of hypotheses. Please tell us more about this.
- yes. When a patient comes to the doctor with several complaints, the question often arises: what section of medicine do these complaints belong to? Do they lie in the competence of an endocrinologist, oncologist or other specialist? After analyzing these complaints, collecting information about the patient, the artificial intelligence system will tell you what else you need to ask the patient, what additional laboratory or instrumental studies to conduct in order to make or, conversely, refute the diagnosis that arose from a huge variety of hypotheses in fewer steps. The less information we have about the patient, the more hypotheses are obtained at the output, and the doctor's task is to first build a set of hypotheses with the help of AI, and then select the most suitable ones from them. At every intellectual step of the doctor's work, the AI system should help him make the right decisions and make fewer mistakes.
― Let's talk about orphan diseases. The odyssey of patients with rare diseases to their true diagnosis is a complex winding road that takes at least six to seven years from the beginning of complaints. Is there any way artificial intelligence can be useful in solving this problem?
― We are engaged in the diagnosis of orphan diseases together with colleagues from Moscow under the guidance of Professor B.A. Kobrinsky. You are quite right to say that their diagnosis is very difficult, because doctors often simply do not have enough experience with such diseases due to their rarity. It turns out that there is practically nothing to study here ― there are simply no training samples. Therefore, we use knowledge from the medical literature and reasoning by analogy (by precedents): when there is no reliable knowledge, but there are some similar cases for which such a diagnosis has been proven. We can find a similar medical history, a similar case, in order to somehow help the doctor make the right decision.
― Valeria Viktorovna, tell us finally about how you came to the profession?
― I can't say that since childhood I dreamed of doing artificial intelligence. I wanted to program, I really liked it. When I graduated from the Leningrad Polytechnic Institute with a degree in applied mathematics and came to work at the Institute of Automation and Control Processes of the Far Eastern Branch of the Russian Academy of Sciences, I was assigned to a team that dealt with AI. At that time, I knew very little about this direction, but in the process I became very interested in it. It literally captured me. I realized how interesting and difficult it is, that there is a future behind it. In those years, this direction was not popular at all, but today, perhaps, only the lazy have not heard about AI. I am glad that the time has finally come when everyone understands the importance and relevance of artificial intelligence. Doing it, you have to dive into different fields of knowledge, be it medicine, robotics, agriculture, manufacturing, etc. ― all this is very interesting. This is how you delve into, try to understand each subject area. And when you see how the AI system comes to life, showing very non-trivial results, it gives strength and energy to move on.
Portal "Eternal youth" http://vechnayamolodost.ru