.png)
Much has been said about artificial intelligence in health and its potential in various fields. For example, the 2023 edition of Future Health Index, a report that raises some perspectives on the future of health around the world, found that 83% of the leaders interviewed plan to invest in AI over the next three years, highlighting two priorities: the application of AI as a clinical decision support tool and the use of technology to bring greater operational and clinical efficiency.
However, at the same time that these tools are seen as part of a major trend - which has been increasingly reflected, especially with the popularization of ChatGPT - the understanding of what artificial intelligence really is, how it works and what has actually been implemented today in health is still very diffuse.
To help understand all these issues and this imprecise scenario of the potential and use of artificial intelligence in health, we have the help of Alexandre Chiavegatto Filho in an exclusive interview. He is a teacher specializing in Machine learning in health at the University of São Paulo (USP), has a post-doctorate from Harvard University and is a columnist for Estadão on this topic.
After all, what is the definition of artificial intelligence?
In a very simple way, Alexandre Chiavegatto defines artificial intelligence as “the ability of machines to make intelligent decisions”. In this sense, the great challenge would be the definition of what intelligence is, a philosophical discussion that he answers with the meaning that makes the most sense: “it would be the ability to make the best possible decision based on the available information”. In other words, at the end of the day, artificial intelligence is a matter of data processing analysis, which shows the centrality and importance of their role.
Still in relation to the definition and operation of artificial intelligence, it has become common to hear the terms Machine learning and Deep Learning, but without a precise explanation of what they mean. Chiavegatto explains that currently the area of artificial intelligence is dominated by Machine learning, but that was not always the case.
“Before, we had machines making intelligent decisions based on pre-established rules by humans. This is what we call “classical artificial intelligence” today. But today AI is almost synonymous with machine learning, which works as follows: instead of entering the rules for the computer to make an intelligent decision, we guide the learning of them by that computer based on the available data and examples”.
As for what is called Deep Learning, he explains that it is just one of the algorithms that is used today for learning rules, and is therefore a subfield within the area of Machine learning.
For each dose of “hype”, a skeptical one is recommended
It has become increasingly common to access the news and see headlines, both international and in Brazil, that point to artificial intelligence in health as a “gigantic market” or a “great revolution”. These calls bring an air of justifiable optimism, since these technologies have great potential in the sector, but the truth is that we are just beginning, as Chiavegatto points out.
“Today we are in the prehistory of artificial intelligence. Today it practically does not exist in the most important area of all, which is health. Then we heard people say 'wow, we're living a Hype where they're putting AI in everything, 'but the truth is that artificial intelligence hasn't even started to enter 99% of areas,” she adds.
In addition, the professor points out that today we can call practically any data analysis artificial intelligence, due to the fact that it learned from data. Without a doubt, this contributes to this perception that “artificial intelligence is everywhere”. But, as he points out, there are many low-quality analyses, given that today there are still not many people who understand the area thoroughly and are able to assess what is being done. In this sense, he recommends: “a first tip is to be very skeptical about what people present as results”.
Health has not yet experienced its “ChatGPT moment”
Chiavegatto says that the application of artificial intelligence in health is still very early, mainly because it is a very “consequential” area. In other words, if an algorithm presents an erroneous probability of prognosis, this can have very profound consequences in patients' lives. In this sense, much greater care must be taken when inserting this type of tool into clinical practice.
“We still don't have a major AI milestone in health. We have several findings that indicate the possibilities of use in the sector, but there has not yet been any major 'ChatGPT moment', in the sense of changing the direction of the area. We are waiting for the area to mature and are trying to solve all the technical problems to put this into practice”.
However, he reports that laboratories have discovered that the algorithms that work in the health area are the same as those that work with various everyday applications, such as Instagram, Waze, Netflix, and ChatGPT. It seems that these same algorithms are going to transform health. “We just need more care and patience,” adds Chiavegatto.
The complexity and challenges in the application of artificial intelligence in health
The professor also talks about some problems raised at the Big Data and Predictive Health Analysis Laboratory of the School of Public Health of USP. Since Brazil is a country with continental dimensions and healthcare is an extremely complex area, does an algorithm that learns from patient data in São Paulo work the same way in the interior of another state in a completely different Brazilian region? Chiavegatto says no.
“If we compare São Paulo with, for example, a city in the interior of Pará, it is clear that there are other types of patients, there is a different training of health professionals, a different availability of exams. In this sense, our studies have shown that quality drops considerably because of the enormous differences within Brazil. So one challenge that we are working on is precisely how can we transfer this knowledge to different regions of the country? We call this transfer learning.”
In addition, another challenge is the issue of continuous learning. Alexandre Chiavegatto explains that, as the algorithm predicts things, it relearns from the results, which is very important in health, where changes are expected in terms of protocols and processes. There is a need to readapt the algorithm to these new realities.
There is also another sensitive point, which concerns the identification and correction of possible artificial intelligence biases due to the data available for its learning. “There is, for example, a real risk that the algorithm will recommend better decisions for rich people than for the poorest population, since today there is a reality that we collect more data from richer patients,” he explains.
The dynamism of artificial intelligence in the health regulatory environment and the idea of the “scapegoat”
A frequent issue in the midst of this debate concerns the health regulatory environment and how it affects the development of artificial intelligence in the sector. It is evident that health is an extremely regulated area, as it should be. In this regard, Chiavegatto points out that there is a very big challenge both for the FDA in the United States and for Anvisa in Brazil.
“It's difficult to regulate because they are, in a way, health devices that will change over time. How do we regulate something that is making a decision today in one way but that maybe 3 months from now will make a different decision or help in another way? That's because the algorithms are relearning as they receive more data. But Anvisa has been debating this for a long time and the FDA has already regulated several AI devices,” she says.
In addition to the regulatory issue, questions are also being asked about the possibility of holding artificial intelligence responsible after a decision taken, a topic that still generates confusion. After all, could artificial intelligence be blamed for a wrong clinical choice and be placed as a scapegoat? Chiavegatto explains that no, since the algorithms will not make health decisions.
“Nothing in health is 100%. I cannot say that there is a 100% chance that a particular patient will have cancer in 5 years, for example. It is always a probability that it will be passed on to the health professional to use together with other tools in their daily lives. The difference is that the Machine learning will unify all tests and patient health data into the same result, providing a probability. It is a guideline for the doctor to make the best possible decision”.
Possible applications of AI in health: expanding access and more time for the practice of medicine
When thinking about the potential of artificial intelligence in health, there are many expectations. In this sense, the USP professor raises two central applications that the tool can have in health, namely the aid of AI in clinical decisions, which must occur more in the long term, and the reduction of bureaucracies, which must be incorporated more quickly by the sector as a whole.
For him, the greatest potential of artificial intelligence lies in its ability to help medical professionals make the best clinical decisions regarding the diagnosis and prognosis of patients. Chiavegatto points out that this application of AI expands access to quality health in all Brazilian regions, even those that do not have medical specialists.
“There are many Brazilian cities that have only one doctor and he has to play the role of cardiologist, pulmonologist, etc., because there is no specialist to refer the patient. In the coming years, artificial intelligence will help a lot in this. Physicians will have access to the same quality of diagnosis and clinical decision. This will radically change health care in Brazil. But that's also the hardest part and it's going to take the most time.”
On the other hand, the application of AI that should be done more quickly concerns the reduction of the doctor's time spent on red tape. Chiavegatto points out that one of the main complaints - both from doctors and patients - in outpatient care is that the professional spends a lot of time entering the information in the electronic record. In this way, the teacher shares that there are already algorithms that can listen to the conversation between the doctor and the patient and are able to fill in the medical record automatically.
“Another point is that the doctor will not waste so much time reviewing the medical record. There will be an algorithm drawing your attention to the information in the medical record that is relevant to the symptoms reported by the patient, providing a summary of what matters for that clinical condition. Other points are filling out reports, reimbursements, consultation reports and surgeries. There will be more time left for the doctor to actually practice medicine,” she adds.
The great, but little explored, Brazilian potential in the field of artificial intelligence in health
Regarding the Brazilian potential in the area, Chiavegatto assesses that “Brazil, in the health sector, has the potential to be the most advanced country in the world in terms of artificial intelligence”, which is mainly due to the fact that most of the services in the country take place within the same system, which is the SUS, which collects unified data.
“There are other countries, even those that are more developed, that have much worse health information systems than Brazil. They have difficulty collecting unified data on deaths, births, hospitalizations, etc. Brazil, on the other hand, collects a lot of data and 75% of the population uses the same system exclusively, which is the SUS.”
Although the country has the potential to lead this transformation, the USP professor comments that is not the case. “Some of our representatives have a greater focus on regulating the area of artificial intelligence, unlike the vast majority of other countries whose leaders are focused on encouraging and promoting the area of artificial intelligence,” Chiavegatto adds.
On the other hand, on the side of medical education, he points out that he has seen a growing interest among medical students in relation to artificial intelligence. Chiavegatto says that, although many outsiders believe that doctors would be averse to these tools, he is no longer asked by these professionals if “AI will replace the doctor”. In fact, he reports that he has seen many students seeking courses, starting to program and developing algorithms.
“The residency at USP, for example, already has a mandatory digital health discipline and undergraduate students are introducing several subjects related to this topic, something that will be inevitable for the future of doctors and will enhance their impact. This has attracted a lot of students”, he concludes.
