Brandon Fornwalt Geisinger in Pennsylvania, US and colleagues tasked with AI inspecting 1.77 million electrocardiograms (ECG) results in almost 400,000 humans to predict a greater threat of dying inside the next year.
That AI can now predict a person’s probabilities of death inside a year via searching at coronary heart test results – even when they look ordinary to doctors. How it does so is a mystery.
The ECG records the electrical activity of the heart. Its pattern changes in cardiac conditions including heart attacks and atrial fibrillation.
The two variations of the AI: in one, the algorithm was once only given the raw ECG data, which measures the voltage over time. In the other, it was fed ECG data in mixture with patient age and sex.
The AI’s overall performance using a metric regarded as AUC, which measures how nicely a model distinguishes between two groups of human beings – in this case, patients who died inside a year and those who survived. The AI persistently scored above 0.85, the place the best score is 1 and a rating of 0.5 suggests no big difference between the two groups.
AUCs for risk scoring models currently used by doctors range between 0.65 and 0.8, says Fornwalt. Researchers also created an algorithm based on ECG features that doctors currently measure, such as certain patterns from the recordings.
“No matter what, the voltage-based model was always better than any model you could build out of things that we already measure from an ECG,” says Fornwalt.
“That finding suggests that the model is seeing things that humans probably can’t see, or at least that we just ignore and think are normal,” says Fornwalt.
AI accurately predicted the risk of death even in people deemed by cardiologists to have a normal ECG. Three cardiologists who separately reviewed normal-looking ECGs weren’t able to pick up the risk patterns that the AI detected.
“AI can potentially teach us things that we’ve been maybe misinterpreting for decades,” he says.
It’s still unclear what patterns the AI is picking up, which makes some physicians reluctant to use such algorithms.
This research was based on historical data, and it will be important to demonstrate in clinical studies that such an algorithm improves patient outcomes, says collaborator Christopher Haggerty.
The research will be presented at the American Heart Association’s Scientific Sessions in Dallas on November 16.