Artificial intelligence can predict the risk of death in the short term, and researchers are confused about how it works

Artificial intelligence can predict the risk of an individual's short-term death (during the year) by examining the results of his or her heart tests, which sometimes may seem "normal" to doctors. Scientists currently do not know exactly how this AI works to achieve this.

Brandon Fornwalt, from health care provider Geisinger in Pennsylvania (US) and colleagues, asked artificial intelligence to examine some 1.77 million ECG results from nearly 400,000 people , in order to predict which would be at higher risk of death in the next year.

As a first step, you should know that an ECG records the electrical activity of the heart: it changes in case of heart disease, including before or after heart attacks, in people with atrial fibrillation (a disorder of rhythm cardiac) or other diseases.

The team created two versions of the AI. A first whose algorithm only received the raw ECG data (which reveals the electrical activity over time). And a second who received the ECG data combined with the age and sex of the patients.

The researchers then measured the performance of the AI ​​using a metric called AUC, which defines to what extent a model distinguishes two groups of people: in this case, the patients who died during the year and those who survived ... The AI ​​consistently scored above 0.85 (the perfect score being 1, and a score of 0.5 would not distinguish between the two groups). " The AUCs for the risk rating models currently used by physicians range from 0.65 to 0.8,  " explains Fornwalt.

For comparison, the researchers also created an algorithm based on ECG features currently measured by physicians, such as certain record regularities. " Anyway, the stress-based model has always been better than any model we can build from features we already measure from an ECG, " says Fornwalt.

AI has accurately predicted the risk of death, even among those considered by cardiologists to have a normal ECG result. The three cardiologists who examined the normal-looking ECGs separately were not able to detect the risk profiles identified by the AI.

This discovery suggests that the AI ​​identifies risks that doctors probably can not see, or at least they ignore and think normal,  " says Fornwalt. " Artificial intelligence can potentially teach us things that we may have misunderstood for decades, " he added.

At present, we still do not know which specific patterns are detected by the AI, which makes some doctors reluctant to use such algorithms. "  This research is based on historical data, and it will be important to demonstrate in clinical studies that such an algorithm improves outcomes for patients,  " says Christopher Haggerty, a Fornwalt collaborator.

Two studies on the performance of this new AI will be presented tomorrow, November 16, 2019, at the American Heart Association's Scientific Sessions.


Post a Comment

Previous Post Next Post