Since 2020, artificial intelligence (AI) has been one of the most talked about topics in the world. In health, applying AI-based technologies opened up a world of possibilities, especially with regard to ways to reduce operating costs and obtain greater assertiveness in treatments.
However, enthusiasm does not always generate positive results in practice. Name specific cases of debacle of AI in health can be difficult, since several factors influence the outcome. Even so, a classic example is the Watson Health from IBM. Put on sale in 2021, it can be said that the technology failed mainly due to the attempt to apply it in contexts that are too wide - that is, in several institutions, each with different ways of operating and varied audience profiles.
Despite this, the future of AI in Health continues to show promise. To understand the current moment of this technology, it is worth considering some recent data:
- In 2023, the Future Health Index, a report that raises perspectives on the future of health around the world, indicated that 83% of the leaders interviewed plan to invest in AI in the next three years;
- Research by ANAHP (National Association of Private Hospitals), conducted in the same year in partnership with ABSS (Brazilian Association of Health Startups), shows that 62.5% of the private hospitals already use some artificial intelligence technology. In addition, the research also indicated that investments in this type of resource are expected to increase by up to 12% this year;
- The healthcare AI market is estimated to have an annual growth rate of 36.4% from 2024 to 2030, according to a report by Grand View Research.
But after all, how mature and secure are these technologies today? What are the clearest paths to the failure or success of a technology?
AI in healthcare in 2024
Defining how mature a type of technology is can be a challenge. However, to stipulate the answer to this question, it is possible to resort to AI application segmentations.
“Generally speaking, we are not at an early stage. Artificial intelligence has been around for over 50 years, but the maturity level of these technologies depends a lot on the problem you're dealing with. There is no artificial intelligence that will solve all problems,” explains Daniella Castro, researcher, co-founder and CTO of Huna, a startup that develops solutions to optimize the diagnosis of cancer and chronic diseases.
Considering a broader scenario, AI applications in the marketing and social media market, for example, “are already very mature,” says the expert. However, “there is still a good way to go in Health”, says Daniella.
In health, the expert indicates that the main applications of artificial intelligence technologies have been in solutions for hospital management and clinical assistance. In this sense, the doctor and director of Applied Innovation and Artificial Intelligence at Dasa, Felipe Kitamura, says that, in health, AI is still in its 'infancy'.
This is because “what is technologically possible is already very different from what we find in reality. We have seen several practical implementations of Artificial Intelligence over the past few years, but even with this advance and several practical uses, it is still a long way from what can be done”, explains Kitamura.
Even so, it is already possible to think of elements that can influence the success or failure of a technology. To do this, we must consider the way in which the technology was built, its purpose And your context. “Due to the characteristics of each business, it may be more or less difficult to prove the value of a particular solution,” says Felipe.
Failures and Bottlenecks of AI in Health
In 2021, IBM's Watson Health went on sale and was one of the first technologies to be seen as a failure by the market. The reasons that led to this were several, but the main one was the early ambition to want to carry out several different tasks, serving many different audiences.
Other elements that also contribute to an AI technology failing are:
- Poor quality data collection;
- Data generated with biases;
- Inadequate data processing;
- High cost technology;
- Stability of the context in which it will be applied;
- If the technology was created to solve a problem, or if it is a solution in search of one;
- Lack of training of professionals to use technology.
However, there are also four other technical issues that directly influence the path of an AI technology. According to Kitamura, they are:
1. Generalization capacity
To develop an algorithm, it is necessary to use a database as a basis. And the results are likely to be positive, but by How long? Em What contexts will optimal performance persist?
“If it uses data from other institutions, there is no guarantee that it will work. When the model works on data from another institution, the algorithm generalized. When it doesn't work, we say it hasn't generalized. And that's not always binary. The generalization may be partial. For example, training the data from my institution, I obtained an algorithm with 90% accuracy, but when we take it to another institution, it can keep 90%, it can fall to 88% or even less”, exemplifies Felipe.
2. Data drift and the issue of accuracy
Accuracy is the word used to describe the encounter of “exactness” with “precision”. While exactness refers to how close a set of measurements is to an actual value, precision is about how close the measurements are to each other.
In this sense, the accuracy of an AI technology in health can be influenced by different factors. Among them: changes in the profile of the audience served, replacement of equipment during exams - for example, changing the brand of a CT scanner and the images coming out with a different quality than the previous machine - among others.
Therefore, even if the accuracy is good initially, when put to work every day it can lose accuracy over time and “not because the algorithm changed, but because the data changes over time in every institution, it's natural”. In English, this situation is called “data drift”, representing changes in an algorithm's database.
3. Fairness
“Fairness” is the English word that means Justice. Within the health technology market, this concept refers to respect and transparency with patients.
This means that if a technology has not been properly trained to be used with a specific audience, it should not be used with them. For example, sometimes an algorithm is structured with data from adults but not from children. Therefore, the tool cannot be applied to children, because it is biased and can be misleading. In other words, the same is true for some diseases, ethnicities, and demographic profiles.
“I need to be fair to the patients. In the ideal world, algorithm models would be biased and would work perfectly for everyone, but we know that in the real world this doesn't always happen. And when that doesn't happen, it's necessary to be transparent: I knew that my model doesn't work for children, so I can't use it for children. I will only use it on adults, because I know that if I use it as a child I will be harming them, because they will make much more mistakes with this audience”, exemplifies Kitamura.
4. The relationship between trust and clinical outcome
Finally, how much a professional trusts artificial intelligence is essential for everything to occur correctly. However, it's not about trusting blindly, it's necessary to have balance.
As Kitamura explains, “the fact that an algorithm works very well and gets it right a lot doesn't mean that, when I put it to work in the hospital or clinic, the doctor will start to get it right just because he has the algorithm available”. That's because, two dangerous scenarios can happen.
The first is the case of overconfidence. “If the doctor starts to blindly trust the algorithm and not his own opinion, the final accuracy that the patient will have will ultimately be that of the algorithm, because he is simply copying and pasting the algorithm's response”, exemplifies the doctor.
On the other hand, excessive mistrust can also cause harm. For example, in cases where the doctor is afraid of the technology and does not want to use it, the outcome will be based solely on the doctor's accuracy. In that case, there's no way to know if the algorithm could have gotten it right or not.
“The best way to use the algorithm is for the doctor to have his own judgment based on the study he had. Then the doctor must be a good doctor, he has to know well to know the cases in which the algorithm is clearly wrong. If he doesn't trust and disdain the algorithm's result, he trusts his own opinion. Eventually it will happen that the doctor made the wrong judgment, due to tiredness, fatigue, or any other reason. Then, when the algorithm presents a result that differs from what he thought, it is an opportunity for him to rethink and review the case, so he can improve its accuracy. So that's another major challenge of human-machine interaction. How to ensure that the doctor knows how to do this? That he will in fact know how to use the platform?” added Felipe Kitamura.
Success stories
On the other hand, the most niche innovations are having a good time. In Brazil, the Brazilian Society for Health Informatics (SBIS) is one of the organizations that help to think about what works or doesn't. According to Grace Sasso, vice-president of SBIS, “what is working right at the moment is mainly precision medicine and predictive”, says the expert. “Another point is the optimization of efficiency and resource governance. The more you use, the better the operations will be, whether in institutional areas, in hospital areas, in the area of primary health care, or in clinics,” adds Grace.
In this sense, it is possible to rely on two Cases of successes that occurred at Dasa.
1. Algorithm for accelerating magnetic resonance imaging
In one of its artificial intelligence projects, Dasa applied an algorithm to 80 machines that accelerated the time taken to generate the magnetic resonance images. The image ends up coming out with a bit of noise, but the algorithm reconstitutes it until it regains quality.
The final results of this application were: patients now spend less time in the discomfort of the machine and the number of exams taken per day has been increased, which is beneficial for patients and for the companies involved. “This is great for the health sector as a whole because the provider can reduce the cost and can perform more exams without increasing expenses,” says Felipe.
Another positive point is that, by optimizing existing equipment, the absence of increased expenses reduces the need to pass on costs to patients.
2. Technology to expand diagnoses
Another project that is doing well at Dasa is a technology developed to help detect diseases as early as possible.
“Sometimes the patient goes for an exam and doesn't return the appointment. It may be because the pain he was experiencing improved, but there is a relevant percentage of cases in which we found something on the same exam that was not being suspected. For example, the person had a stomachache and went for a CT scan of the abdomen. Regarding the stomach pain itself, she had nothing. It's just that we accidentally discovered a liver tumor. This happens with a certain frequency and can you imagine if that person is not going to seek the exam? Sometimes she had a tumor that was curable and six months later it's no longer curable,” says Felipe Kitamura.
Thus, Dasa's natural language processing model was born, which today reads more than 8,000 daily reports without taking the radiologist out of office.
“The radiologist is still there, organizing in the same way. He looks at the image, writes the report and releases the result. When the report is ready, these algorithms read all these reports and when they identify a disease that requires a next step, be it a biopsy or some other test, the algorithm detects these cases and then the examining doctor calls the patients' doctors to report these findings,” explains Kitamura.
The decision to take the next step in the investigation ends up being made by the doctor and his patient. Even so, Dasa has already been able to demonstrate that this technology reduces the patient's time to start the next phase from 17 to 7 days. In addition, this measure generates a good financial return, since late treatment is often more expensive for the health system.
The future
Finally, the AI specialist and vice-president of SBIS (Brazilian Society for Health Informatics), Grace Sasso, says that she sees as a trend for 2024 the expansion of: persuasive technologies, models that help optimize precision Medical and tools of prediction.
For the market, she reinforces that one of the challenges to be faced from now on is in relation to accessibility and equity in health.
“We need to train people to improve the development of technologies and at the same time look more at the opinion of those who make use of information. We see many applications where information is recorded, but you don't know where that data goes, there's no return of it, there's no storage security. For example, where is this data stored? Who is using it? Why are they using it? We need to be increasingly careful with these aspects. We are going along a line of 'super intelligence, 'but society is not prepared for that. Humans need to be preserved, people need to be listened to, and then we have a dichotomy and I don't think we can go back,” concludes Grace Sasso.
