October 16, 2021

This AI can predict when you’re going to die, but scientists don’t understand how

By examining a patient’s heart tests, and even when they look normal, artificial intelligence is able to predict the risk of a person dying within a year. But how AI does it remains a mystery.

And you, would you like to know the date of your death ? Researchers at Geisinger, a healthcare provider in Pennsylvania, United States, were able to predict which patients were at risk of dying during the year. To do this, they trained artificial intelligence (AI) to detect signs of potential heart problems in the future, such as heart attacks or atrial fibrillation .

The machine looked at the results of 1.77 million electrocardiograms (ECGs), or heart activity recordings from nearly 400,000 patients. Two versions of AI have been developed. One analyzed the raw data, the other also received the age and gender of the participants. The findings of the study, relayed by New Scientist , were presented at the American Heart Association conference in Dallas (United States) on November 16, 2019.

Humans don’t see

The AI predictions were then sifted through, using what is called the “AUC”. This metric measures the performance of a model that distinguishes between two groups of people. In this case, it was on the one hand the deceased persons , on the other hand those who survived. A score of 0.5 does not indicate any difference between the groups. A score of 1 is perfect. However, technology has systematically obtained a score higher than 0.85 . The doctors, them, had a score which varied between 0.65 and 0.8, assure the authors of the research.

According to them, the AI ​​model would therefore work better than existing methods to detect potential deaths . He would even have identified heart problems in patients already studied by cardiologists. Each in turn, three doctors have indeed examined ECGs of “normal” appearance, and were not able to detect the profiles at risk , as did the machine.

It may interest you :  Europe makes policy proposal for regulating artificial intelligence

“ This finding suggests that the model is seeing things that humans probably can’t see, or at least that we ignore and think are normal ,” says New Scientist Brandon Fornwalt, principal investigator of the study. Artificial intelligence can teach us things that we may have misinterpreted for decades. ”

Operation is still unclear

However, scientists still do not know what patterns are detected by AI , and thus find it difficult to explain how it works . It is this lack of knowledge that worries health professionals, who are reluctant to make decisions based on a misunderstood algorithm.

This study is not the first attempt to predict death. Last year, Google researchers created a predictive model using electronic health records to predict how long a patient was in a hospital, when they left, and when they died. AIs have also been developed to diagnose heart disease and lung cancer . And sometimes with more precision than human doctors.