Imagine predicting exactly how and when a person will die. Such a deterministic reality seems a bit far away if possible, but researchers at USP have used machine learning to achieve a high level of prediction of deaths from respiratory disease.
With the help of artificial intelligence, scientists have been able to identify up to 88% of deaths from breathing problems. Within the group of people studied, the researchers also rated the lowest to highest risk of death from these diseases. The highest risk group (25% of the sample) had 100% of respiratory deaths.
Covid-19 was not part of the researchers’ analysis.
Deaths from cardiovascular problems, neoplasms, which together with breathing problems are the leading causes of death in Brazil, were also assessed. In these cases, however, the prediction did not have good results.
By now, you must have heard of machine learning at some point in the past few years. The idea is very summarized and general, to feed a program with a certain volume of information as a form of training. At this stage, the goal is for the application to be able to perceive patterns in this data that the human eye may miss.
Another dataset is then submitted to the program to try to identify patterns – and provide answers that humans cannot provide.
USP researchers have done just that with data from São Paulo residents aged 60 and over collected over the past two decades.
The researchers at Labdaps (Big Data and Predictive Analysis in Health Laboratory) at the USP’s School of Public Health used the same college’s Sabe Survey (Health, Wellbeing and Aging) database, which focused on people from the city São Paulo with 60 years or more.
In the Labdaps survey published this week in Age and Aging magazine, the causes of death of people in the five years following Sabe’s interviews, which began in 2000, were monitored with data collection in later years. For the study of machine learning algorithms, the scientists used the data collected by Sabe in 2006 and 2010, which included a total of 1,767 people.
However, the sabe was not made to indicate mortality. For this reason, the researchers had to compare this information with the death dates of the municipality of São Paulo.
After the intersection was completed, the machine learning algorithm for training was fed with 70% of the database. The other 30% was used for the predictive test.
According to the authors, this is one of the largest studies ever conducted into predicting death in large populations. “What exists in the literature is the application of machine learning to identify risks in certain population groups. For example, in people who already have heart problems, in people who have already been diagnosed with cancer,” says Carla Nascimento, PhD student in the area Public Health at USP, researchers at Labdaps, and one of the study’s authors.
The purpose of this type of technology is to influence the clinical behavior of health professionals.
“It opens up a number of ways to take preventive measures to prevent death from occurring,” says Alexandre Chiavegatto Filho, director of Labdaps. “It is not something, ‘you are going to die and you have nothing to do with it’. The great interest in knowing is to prevent it from happening.”
Although this is information that is theoretically only intended for the health team, patients may also have access to the data. And then comes the question: “Do people want to know?” Asks Chiavegatto Filho.
“Some people are afraid when they see the risk of an outcome. But I also don’t think it’s a fear that lasts long,” says Nascimento. “But I am not thinking of applying this to savings, but to professionals.”
The research team intends to move forward within this universe and is trying to understand how far this type of technology can actually change clinical behavior. To do this, the idea is to provide a randomized trial in which the program is made available to one group of doctors while another group is denied access to the technology. Afterwards, the patients of these specialists will be examined again.
The application will be made available to a network of 30 hospitals across Brazil through a platform called RandomIA, funded by Microsoft and Fapesq (Foundation to Support Research in Paraíba State).
However, before exposing the artificial intelligence algorithm, the Labdaps team tries to understand what kind of information doctors want to see in the application.
“Does he want something very simple. ‘This patient will die of cardiovascular disease in 5 years’. Does he want a probability? Or does he provide probabilities and degrees of uncertainty,” says Chiavegatto Filho. “Sometimes too many details can be very complex and the doctor can ignore the information. And with a few details, the doctor may not trust the result.”
However, it is worth noting that this type of technological application in the health sector is still in the study phase and not as close to being applied in routine medical practice.
The director of Labdaps believes that mechanisms of this type will be well received by the medical community when they are widely available. He compares the situation with the Waze app, which presents route options using a lot of collected data that the driver cannot necessarily easily access.
“It’s impressive how many difficult and complex decisions doctors make during their day,” says Chiavegatto Filho. “The doctor spends a lot of time gathering information from patients and recovering scattered information. Nothing that unifies all of this information and helps with decision-making. That brings artificial intelligence.”