Algorithms are already working in areas where there is a lot of digital data. But it's hard to implement them.
- Is there any
— With the advent of new capacities, self-learningalgorithms are spreading more and more, this is leading to a surge in all areas, including medicine. But medicine is a specific area, one of the most conservative industries. This is understandable - the life and health of people depend on it. Innovations here always go a long way before they prove themselves. Medicine attracts many technology companies, but in practice, testing is more difficult and some areas do not always work.
For AI to work, arrays are needed, which can beanalyze and learn. Therefore, we see many examples of developments from areas where the data is originally digital. From a CT scanner, MRI, ultrasound, data can only be seen on the screen. They have been accumulated for a long time, and now they are used for training and creating algorithms. Medical imaging is one of the first areas where AI began to spread, get recommendations and proven results. Moreover, algorithms that have already found application work here - for example, face recognition.
Or electronic medical records of patients.This is also structured digital information, it can be used to find medical diagnoses that were not initially detected. But each institution stores information differently, so it's not easy for researchers to get a dataset that they can work with in practice. However, research in this area is ongoing.
Where there is less digital data, AI is still difficultfind a use - for this there are not enough sets for training and tests. And in real practice, the accuracy of the algorithms will differ from what was obtained in the laboratory.
Dmitry Lisogor
- I understand that the sample is still small, but can we already say that diagnostics have become more effective in some of the areas?
— We are not talking about what the algorithm will putdiagnoses are more accurate than a person. It is now paying off in areas where there are many routine operations, such as mass screening studies in search of lung cancer. To identify a few people at risk, you need to look at a huge amount of data - AI will do it faster.
Or, for example, mammography is a study,which is recommended for women over 40 as a screening program against breast cancer. Doctors need to look at hundreds or thousands of images to identify a few people who are at risk. The algorithm sorts and draws the attention of specialists to those studies that require decision making and diagnosis. This is already being introduced into practice and greatly facilitates the work of doctors.
— That is, this is already a technology that works, although the sphere is very conservative?
— Recognition for technology is permissionregulatory bodies. In Russia, this is Roszdravnadzor, in the USA it is the FDA, and so on. But there was a big legal conflict: self-learning systems change with each patient, it turns out that we use different algorithms every day. A compromise was found - the algorithms are used in a "frozen" state. They are registered, the possibility of learning is turned off - the algorithms are used with the accuracy that was shown to the regulatory authorities.
The FDA has already issued dozens of registrationscertificates for software solutions based on AI, six such certificates were issued in Russia. Basically, algorithms work in the field of medical image analysis or medical record data.
— How does the medical records analysis system work?
— The main idea is to look not onlyon current diagnoses, but also retrospectively. When you come to the doctor, he has 12-15 minutes for an appointment, this is not enough to review the entire history - what the patient came with 3-10 years ago and what specialists he contacted. It is necessary to bring together the entire array of data in order, for example, not to miss the development of chronic diseases. This system allows you to dig deeper, and the doctor looks at the clues and makes the final decision.
To collect data, you need to ask the permission of each patient. But that could change
– There are also patients in this interaction – have you conducted research on their attitudes towards AI? Are they willing to trust data and potential treatments to technology?
— Philips conducts an annual study“Future Health Index” – where we ask questions to a group of doctors, clinic leaders or patients. In 2018 and 2019, research addressed the topic of patient attitudes towards AI. Patients have caution, they express concerns about the fact that their data will be used in algorithms.
Interestingly, professionals in the fieldhealth care does not share this concern. The latest study, Future Health Index 2021, examined the opinions of industry leaders - clinic leaders, chief physicians and heads of departments. Respondents in Russia positively assess the prospects for the introduction of telemedicine and AI tools into practice. Thus, 53% of respondents named telemedicine as one of the digital medical technologies in which they would like to invest the most resources today. And in three years, 85% of those surveyed would like to prioritize investment in AI.
The fear of patients is associated with ignorance - theythey think that the machine will make diagnoses, but it will not work to talk to the doctor. It’s not like that—computer programs have been involved in decision making before, they just haven’t been talked about so loudly. For example, in the Moscow region, a technology for making an appointment with a doctor was introduced. The patient communicates by phone with a bot that tries to classify which specialist and with what degree of urgency he should be referred. According to statistics, 80% of patients made appointments with doctors in this way, and no one asked questions about AI. Therefore, it is rather a psychological barrier, given that the final decision will always remain with the doctor.
- Did the patients express concerns about the safety of data?
- AI brings not so many risks in terms ofcybersecurity. Many data that is used to train or test algorithms becomes anonymized. If the patient's name is removed from the lung scan, the researchers will lose contact with real patients. For some algorithms, of course, you need to look at whether the diagnosis was confirmed or not, but scientists have strict protocols for data security.
Collect and accumulate data without permissionpatients are not yet available. Although at the state level, possible changes in legislation are being discussed in order to create sets of anonymized data without mandatory permission. There are no options not to use AI while developing medicine further. Therefore, it is necessary to remove the legislative barriers that hinder their development. For example, making anonymized data available to researchers without the permission of patients.
Does it work differently now? For training and testing, you need a large amount of data - do you ask permission from each patient?
- There is no other way, this is the only one allowedmethod. The lack of datasets is one of the biggest problems and obstacles to the development of AI. This is not only specific to Russia, but a global problem. Some authoritarian countries do this forcibly - in China they can be used without permission. But Russia is careful about the personal data of patients.
- How much data are you talking about? How much data does an algorithm need to learn?
- Depends on the algorithm, but we are talking about hundreds,if not thousands of studies. Let's say we get a thousand studies to train the algorithm, but we need the same number of studies for testing. On them, the trained algorithm confirms or does not confirm its accuracy. And only then can we talk about the application in practice. And there he will face tens of thousands of studies, and his accuracy will vary.
- You mentioned the conservatism of healthcare. Where is the initiative to introduce AI systems in medicine coming from?
— It cannot be said that there are two polar camps.But like many endeavors, they are driven by entrepreneurs and tech companies who make a product and find uses for it. The main drivers are the creator companies, and of different sizes - they can be start-ups and large technology companies such as IBM, Google, Microsoft, Philips. We also create algorithms and help them find their way into clinical practice.
Doctors are not against technology. But the main initiators on the part of medicine are not the clinicians themselves, but health care administrators, they must think about how to improve the efficiency of the system.
A vivid example is at hand:experiment on the introduction of AI for the analysis of medical images in the Moscow healthcare system. This is a big program that the city started last year, and now the second phase of the study is underway. 21 services participate in 13 directions, including Philips, they have already conducted more than 2 million studies using algorithms. The results are available to physicians who work with patients and decide whether to take the recommendations into account or not.
Algorithms can work in any hospital. But they are not a replacement for doctors.
- Do you need to purchase equipment and infrastructure for implementation in AI? Or is it not very costly for institutions?
— Many algorithms are provided without localinfrastructure from the side of the medical institution - they are initially based in the cloud, only access is required. There is another problem: medical images are already stored somewhere, and we need to solve the issue of integration with them. But everything else is processed at external capacities so as not to create additional computing clusters in institutions.
How will technology affect hospitals? Will the staff be reduced? Will medical education change? Will their approach to diagnosis change? What is your prediction for this?
— Technology is developing so fast that I calledIt would be a vision, not a forecast. In the future, doctors who do not use AI will be replaced by doctors who use AI. No one is saying that algorithms will replace human specialists. The medical specialty is one of the most difficult to master, it requires a lot of training and practice - it is because of this that in many areas there are not enough specialists. The demand for medicine and its availability is growing, the population is aging, there are more diseases, and the detection rate is improving. But the number of doctors is not growing at the same rate.
AI will increase the efficiency of their work, soit is impossible to talk about reducing their number. Technology will help to focus on those situations where you need to make a decision and make a diagnosis, on communication with the patient. This is very important: without dialogue it is very difficult to make a diagnosis, prescribe treatment, check its progress and quality. Technology should free up time for diagnosis and treatment to be more human.
- Will AI replace the contact between the patient and the doctor? And in general, how important is it in the future, where research will be based on objective data and analysis of data arrays?
- The contact between the doctor and the patient is important, because the twoThere are no people with the same diagnoses, there will always be nuances. If we are talking about a cold, it is not so critical. But if it is an oncological disease or a disease of the cardiovascular system, which can have long-term consequences not only for the patient, but also for his environment, communication from a specialist cannot be replaced. He must correctly prepare everyone for the further path - complex treatment or surgery.
- Will AI affect the salaries of doctors and their role in medicine in general?
– I don’t think there is a risk in this – the amount of workWill not change. For example, in a lung exam today, thousands of fluorograms need to be looked at to identify suspected tuberculosis. The doctor will only be able to identify several patients who have suspicions in a week. This will delay the diagnosis, the start of treatment, sometimes it can be critical. In the future, the result will be the same, but we will achieve it faster, because the algorithm will weed out images that do not deserve attention. But the diagnosis will remain with the doctor, his work will not become less responsible or qualified.
— During the pandemic, a huge amount of data has appeared on which you can test or train algorithms. Have you been able to use this data?
— Doctors who looked for blackouts related toCOVID-19 may have missed masses associated with tuberculosis or lung cancer. Therefore, AI is used for secondary research, detection of concomitant diseases. We offer a service that allows you to avoid the integration of dozens of algorithms - it is enough to connect our platform once, and through it the customer gets access to the algorithms of different developers in the marketplace format. With it, you can take one picture of the lungs, and AI will look for various abnormalities - tuberculosis or cancer.
— There are hundreds of publications about AI in medicine,From the outside, it is almost impossible to assess which areas are promising and which are too overvalued. For example, they talk a lot about radiology – is there really going to be a big breakthrough there? What other areas are worth following?
- Radiology seems promising only because oforiginal digitalization. The data is already digital, so it is easier to create and implement algorithms. Probably, a very big transformation will take place in this area, many start-ups will appear, technological giants will come. For example, Microsoft recently bought a voice analysis and medical image recognition company.
The use of AI in the related field seems promising.area — the search for new drugs. The algorithm allows you to simulate the body's response to certain components. There have already been news about molecules that have avoided the traditional way of working in the lab, without using a large number of experimental animals. Of course, this molecule still goes through all the stages of clinical trials, but this significantly shortens the development cycle. Perhaps it is with the help of AI that they will find cures for diseases that are not yet treatable.
Read more
See what Saturn looks like from the Moon. Photo taken by NASA orbiter
Hubble completes a journey through the outer solar system: what he saw there
The fusion reactor KSTAR sets a record: it holds the plasma longer than ever