Science & Enterprise logo
Science for business people. Enterprise for scientists.

Algorithm Devised for Fast Heart Rhythm Screening

Heartbeat graphic
(Maxpixel.net)

5 Aug. 2019. A machine-learning algorithm used with results from a standard electrocardiogram, is quickly able to detect irregular heart rhythms, a process that often takes much longer. A team from the Mayo Clinic in Rochester, Minnesota describes the algorithm and its findings in the 1 August issue of the journal The Lancet (paid subscription required).

Researchers led by cardiac electrophysiologist Paul Friedman are seeking a quick, inexpensive, point-of-care diagnostic method to detect atrial fibrillation, a disorder where the atria, or upper chambers of the heart, beat irregularly instead of the normal, smooth regular beats that move blood effectively through the blood stream. Because of these irregular heart rhythms, blood can pool in the atria and form clots. If a clot should break off and flow to the brain, it can cause a stroke. American Heart Association estimates as many as 20 percent of people who have a stroke also have atrial fibrillation.

People with atrial fibrillation often do not display symptoms, thus most cases are detected in visits to doctors offices or clinics. An electrocardiogram or ECG, a standard heart health test measures electrical signals with up to 12 electrodes or leads attached to the chest and torso. An ECG test does not provide enough data to draw conclusions, however, and if irregular electrical signals are detected, more tests are needed. Those tests, such as an echocardiogram that uses ultrasound or a CT scan, provide detailed real-time images of the heart, but require more time and expense.

Friedman and colleagues developed an algorithm that harnesses machine learning to evaluate an individual’s ECG results to screen for atrial fibrillation at the point of care. The researchers inspected nearly 181,000 medical records from Mayo Clinic patients, from December 1993 through July 2017 for their analysis, providing some 650,000 ECGs.

The team assessed these data with a convolutional neural network that combines features of image analysis and machine learning. In a convolutional neural network, an algorithm dissects an image by layers to understand the features in the image. Different aspects of each layer discovered and analyzed by the algorithm are translated into data that the algorithm then uses to train its understanding of the problem being solved, with that understanding enhanced and refined as more images and data are encountered.

The researchers collected some 455,000 ECGs from about 126,500 patients to train the algorithm, with more than 64,000 ECGs from 18,000 patients to validate its calculations. The team then tested the algorithm on 131,000 ECGs from more than 36,000 patients. The results show an overall sensitivity, or accurate detection atrial fibrillation or flutter of 79 percent, with a specificity or accurate detection of no heart rhythm problems of 80 percent. When the analysis includes ECGs of patients over the first month of testing, the sensitivity percentage rises to 82 percent and specificity increases to 83 percent. The team concludes that the data show algorithm-enabled ECGs offer a point-of-care tool to detect atrial fibrillation.

Many of the same researchers constructed a similar algorithm to detect left ventricle dysfunction, another largely asymptomatic heart condition, which they reported in January 2019 in the journal Nature Medicine. The team tested that algorithm with the records of nearly 53,000 Mayo Clinic patients, and reported sensitivity and specificity percentages each of 86 percent. In addition, patients without left ventricle dysfunction but still positive scores on the algorithm were at 4 times higher risk to develop the condition than those with negative scores.

More from Science & Enterprise:

*     *     *