A new approach to analyzing electrocardiograms–a ubiquitous test of the heart’s electrical function–could predict who is most likely to die after a heart attack. Researchers at MIT found that measuring how much the shape of the electrical waveform varies from beat to beat identifies high-risk patients better than existing risk factors. If the findings hold up in further clinical trials, the technology could be used to figure out which heart attack patients need the most aggressive treatment.
Scientists hope the same approach will eventually help them predict when a healthy person is likely to suffer a cardiac problem. They are working with Texas Instruments to integrate the software into the new generation of wearable heart monitors.
The research also shows how computational analysis can glean useful information from the reams of medical data routinely collected and ignored. “It’s a very novel approach,” says Jean-Philippe Couderc, biomedical engineer at the University of Rochester, who was not involved in the project. “It’s a unique way of looking at how the electrocardiogram varies on a beat to beat basis.”
Electrocardiograms record the heart’s electrical activity through sensors placed on the chest. Cardiologists can then spot abnormal heart rhythms by visually inspecting the resulting waveform for major features linked to the function of the top and bottom chambers of the heart, as well as the heart’s ability to “reset” itself between beats. While some simple algorithms exist to analyze this data, they are notoriously inaccurate. “Cardiologists routinely ignore them,” says Collin Stultz, a professor at MIT, as well as a practicing cardiologist, who is involved in the project.
To determine if more subtle features within electrocardiogram data could provide useful clinical information, Stultz, John Guttag, also at MIT, and Zeeshan Syed, now at the University of Michigan, started with a large data set of 24-hour electrocardiogram recordings collected at Brigham and Women’s Hospital in Boston as part of a clinical trial for a new drug. Employing a number of computational techniques, including signal processing, data mining, and machine learning, the researchers developed a way to analyze how the shape of the electrical waveform varies, a measure they dubbed morphological variability. At the heart of the approach is a method called dynamic time warping, used in speech recognition and more recently in genome analysis, which allows researchers to align and compare individual beats. “We compute the differences for every pair of beats,” says Stultz. “If there is lots of variability, that patient is in bad shape.”
The team then applied the algorithm they had developed to a second set of electrocardiogram recordings and found that patients with the highest morphological variability were six to eight times more likely to die after a heart attack than those with low variability. “We found that it consistently works as well or better than commonly accepted risk metrics that physicians use,” says Stultz, including diabetes, age and smoking status, as well as cardiac ultrasound and various blood tests.