Machines Just Got Better than Humans at Predicting Death

A machine-learning (ML) technology has surpassed human capabilities when it comes to predicting the chances of death or heart attacks, according to research presented Monday at the International Conference on Nuclear Cardiology and Cardiac Ct in Lisbon, Portugal.

In testing, the algorithm, known as LogitBoost, analyzed 85 different variables from 950 patients, whom the researchers had followed for six years. It identified which of the participants had died or suffered heart attacks with an accuracy of more than 90 percent.

"These advances are far beyond what has been done in medicine, where we need to be cautious about how we evaluate risk and outcomes," said an author of the research, Luis Eduardo Juarez-Orozco of Finland's Turku PET Center and the University Medical Center Groningen in the Netherlands, in a statement. "We have the data, but we are not using it to its full potential yet."

ML is a form of artificial intelligence in which algorithms become better and better at predicting a given outcome without being explicitly programmed, usually through the intake of increasing amounts of data.

"ML refers to the family of mathematical algorithms characterized by their capacity to improve their performance in whatever task they are being assigned through iterative exposure to data," Juarez-Orozco told Newsweek. "This translates what we understand as "learning."

So, an ML algorithm is initiated and given a run through some data. It makes then a prediction, we check and tell it whether it succeeded or not and then it tries again, changing if it failed and improving if it succeeded (this is what is called supervised learning, although there are also other forms).

When doctors are determining the best course of action for treating a patient, they often use things called "risk scores." However, these are based on only a small number of variables and therefore can sometimes be inaccurate. Machine learning, on the other hand, can take into account many more variables, meaning it can predict outcomes with greater accuracy in many circumstances.

"Humans have a very hard time thinking further than three dimensions—a cube—or four dimensions—a cube through time," Juarez-Orozco explained. "The moment we jump into the fifth dimension we're lost. Our study shows that very high-dimensional patterns are more useful than single-dimensional patterns to predict outcomes in individuals, and for that we need machine learning."

Nine hundred and fifty patients with chest pain took part in the study, and the researchers collected data from them on several measures, which they inputted into the ML technology. This included information taken from CT scans about coronary health, medical records, sex, age, smoking status and other variables.

In the six years that the scientists followed the participants, 24 had heart attacks and 49 died from any cause. When given the data, LogitBoost was able to correctly predict these outcomes with a more than 90 percent accuracy by continually analyzing the data over and over again.

"Logitboost is part of a group of algorithms called ensemble boosting," Juarez-Orozco said. "In simple terms, the algorithm explores the entire structure of the data we give it and finds complex patterns that are useful to predict if a person is likely to have an infarction [death of tissue resulting from a failure of blood supply] or die."

The algorithm is like a toddler, it trains and gets better by exploring more cases so eventually it gets better and better at it, he said. "In particular, this algorithm trains focusing on difficult cases—by giving them more importance—to learn more and better."

He noted that doctors already collect a lot of information about patients—for example those with chest pain. Thus, machine learning can integrate these data and accurately predict individual risk.

"This should allow us to personalize treatment and ultimately lead to better outcomes for patients," Juarez-Orozco said.

Prior research in this field has mostly focused on ML algorithms making a diagnosis, and they have outperformed human operators and experts in this respect. The latest research differs in that it focuses on prognosis—or predicting the risk of an event—which is a substantially more difficult task.

"In fact, current models in international guidelines perform rather modestly in this prediction," Juarez-Orozco said. "So ML is indeed better than us for predicting these events, but our performance in this regard is also quite modest, so the 'human' benchmark might not be so impressive."

This article was updated to include additional comments from Juarez-Orozco.