Heart Attacks are a growing concern for many people. They can affect almost anyone with little warning. However, there may be an algorithm that can predict heart attacks up to four hours before doctors can.
Cardiac or respiratory arrest are often predicted using a scorecard called the Modified Early Warning Score. This scorecard roughly calculates how severe a patient’s circumstances are by looking at things like heart rate, blood pressure and temperature. If a patient appears to be high risk, doctors can lower the rates of cardiac arrest as well as shorten hospital visits.
At Carnegie Mellon University in Pittsburgh, Sriram Somanchi and his colleagues decided to see if they could get a computer to make predictions of when a heart attack is inevitable. “We had to understand what happens in Code Blue patients before they enter Code Blue,” Somanchi stated.
Researchers used the data from 133,000 patients and instructed a machine-learning algorithm to look at 72 parameters in each patient’s history. It looked at vital signs, platelet count, age and blood glucose. The patients’visited North Shore University Health System between 2006 and 2011. A ‘Code Blue’was called 815 times.
Using all of this collected data, the machine could predict, sometimes up to 4 hours before it happened, when a patient would go into cardiac arrest. Although not perfect, the machine was accurate 2/3 of the time while patients’scorecards only flagged 30 percent of the cardiac arrests.
Although it seems like a good idea, do most hospitals look at 72 parameters when taking data from patients? ”When we look at it from a statistical point of view, a small model is better,”says Peter Donnan at the University of Dundee in the UK. Many hospitals do not collect as much data from patients whereas the scorecard only uses a few parameters to make predictions.
Somanchi says that the algorithm announces a false positive 20 percent of the time so in hopes of improving it, they plan on training the machine using data from a variety of other hospitals. They will present their findings at the Knowledge Discovery and Data Mining Conference from the 24-27 August in New York City.