You know that movie Countdown that came out a while back, the one where there’s an app that all too accurately predicts the day of your imminent death? Well this AI is kind of like that, but without the horror soundtrack, and with a far more altruistic purpose.
A new study found that when artificial intelligence looked at the results of past patients’ heart test results, it could do a pretty darn good job of predicting a person’s death within a year. And it could do so, mysteriously, better than doctors.
The study, led by Dr. Brandon K. Fornwalt and Dr. Chris Haggerty of healthcare provider Geisinger, looked at 1.77 million electrocardiogram (ECG) results from upwards of 400,000 people in hopes of pinpointing those with a higher risk of dying within a calendar year.
“This study is based on actual mortality data collected in our regional health system over the last 20+ years. We trained artificial intelligence algorithms / models to predict 1-year mortality by analyzing data from electrocardiograms (electrical measurements of heart activity),” Fornwalt tells SYFY WIRE.
To perform the study, the team fed one version of the AI an algorithm wih raw ECG data, and another version with that same data along with patient age and sex. Then they used a metric that measures the distinction between groups of people, in this case, those who died within a year, and those who lived.
Where 1 is a perfect score and .5 indicates no difference between the groups, the AI time and time again score higher than .85. According to Forrnwalt, that’s better than the risk scoring models currently displayed by doctors, which range between 0.65 and 0.8.
What’s even more interesting perhaps, is that the AI was able to predict death in patients that had ECGs which appeared completely normal at first glance. Indeed, three separate cardiologists examining such ECGs were unable detect risk patterns uncovered by the AI — a result that still stumps Fornwalt and his team.
“To be honest, right now we do not know and are actively working on exploring why the AI models had good performance to predict 1-year mortality even in electrocardiograms interpreted as normal by physicians,” says Fornwalt. “We suspect there are subtle features or patterns that the model is identifying which physicians typically believe to be within the range of normal variation. Future work will need to explore this in more detail using the latest techniques for exploring what the neural network is actually using to make the prediction.”
A clinical study is necessary to more accurately tweak the algorithm, but this tech could potentially help doctors look for heart disease warnings where they hadn’t been looking before.
The results of the study will be presented later this week at the American Heart Association’s Scientific Sessions in Dallas.
(via New Scientist)