Impressive to know that an artificial intelligence (AI) is able to predict death within one year just by analyzing cardiac tests. But it may be even more impressive that I don't know how he is capable of doing it. Brandon Fornwalt, director of the Artificial Intelligence laboratory of the medical company Geisinger, and his team will present the finding in Dallas (USA) on November 16.
Investigators commissioned an AI that examine 1.77 million electrocardiogram results (ECG) of almost 400,000 people to predict who had a higher risk of dying in the next year.
"AI is seeing things that humans do not see or think are normal"
They used two measurement patterns. For the first algorithm they only used raw data. In a later one, the electrocardiogram information was combined with the patient's age and sex.
To measure the performance of the program they used the metric unit called AUC, which distinguishes a model between two groups of people. In this study, the patients who died in a year and those who survived.
The result was that artificial intelligence accurately predicted the risk of death even in people that cardiologists consider to have a normal electrocardiogram. Three cardiologists who reviewed separately the electrocardiograms that seemed normal could not find the risk patterns that I can detect AI.
In statements to Newscientist, Fornwalt acknowledges his astonishment: "The model is seeing things that humans probably can't see, or at least we simply ignore and think they are normal. AI may have the potential to teach us things that we may have been misunderstanding for decades. "
The surprising thing, it has already been said, is that it is not clear which patterns AI uses. Therefore, some doctors are reluctant to use these algorithms. Christopher Haggerty, a contributor to Fornwalt, warns that it will be important to demonstrate in clinical studies that such an algorithm improves patient outcomes.