Home » MedTech Features » AI Finds Hidden Atrial Fibrillation in Subtle Beats of the Heart

AI Finds Hidden Atrial Fibrillation in Subtle Beats of the Heart

 

Atrial fibrillation, the fleeting arrhythmia of the heart, is hard to catch. Elusive and intermittent, it’s often overlooked during a seconds-long electrocardiogram (ECG, or EKG). But it is also a harbinger of potentially deadly health episodes like stroke and heart failure.

But could artificial intelligence (AI) look at the tiniest nuances of systole and diastole—and pick out the irregularities that go missed by regular observation?

The machine can learn the intricacies of the human heart—potentially leading to better outcomes, as Mayo Clinic researchers recently reported in The Lancet. But limitations remain, as the authors, and outside experts, relayed to MedTech Boston.

“An EKG will always show the heart’s electrical activity at the time of the test, but this is like looking at the ocean now and being able to tell that there were big waves yesterday,” said Paul Friedman, M.D., one of the authors, the chair of the Department of Cardiovascular Medicine at the Mayo Clinic.

The AI was trained from the Mayo Clinic Digital Data Vault, using more than 450,000 EKGs from 126,526 patients. All the data involved normal sinus rhythm taken from 12-lead EKGs lasting the standard 10 seconds, according to the paper.

The convolutional neural network used the Keras Framework with a Tensorflow backend made by Google, programmed with Python.

The machine learned to pick out the nuances of the heart rhythms of those diagnosed with AF, they explained.

The learning network was then tested on 36,280 patients. Of that group, just greater than 3,000 (roughly 8.4%) had been diagnosed with AF.

When the AI’s results were cross-referenced against those of the previous diagnoses, they found an overall accuracy of 83.4 percent.

“Our data indicate that a simple, inexpensive, noninvasive, 10-second test—the AI-enhanced standard ECG— may permit identification of patients with under-detected AF,” conclude the authors.

The clinical implications, if the technology is brought to bear, could be important, the doctors wrote. For instance, if AF is found, anticoagulation treatments after embolic strokes of undermined mechanism (ESUS) can be a lifesaver, the authors contended.

“Although it would require further study, it is possible that this algorithm could identify a high-risk subset of patients with ESUs who could benefit from empiric anticoagulation,” they concluded.

AF is nothing less than a “global pandemic,” according to an accompanying Lancetcommentaryauthored by Jeroen Hendriks, Ph.D., of the Centre for Heart Rhythm Disorders at Royal Adelaide Hospital in South Africa.

The AI model to find AF could be “groundbreaking,” wrote Hendriks.

“Rather than finding the needle in the haystack by prolonged monitoring, authors basically suggest that AI will be able to judge by looking at the haystack if it has a needle hidden in it,” he wrote. “Combining ECG algorithms with age, gender, clinical features and biomarkers may further improve identification of AF patients.

“Additionally, linking these variables with genetic markers, AR-enabled algorithms and smart monitoring by means of wearable to diagnose AF and quantify AF burden, promises a safer and more efficient prevention of AF-related complications,” Hendriks added.

But AI’s relationship to cardiology, and the rhythm of the human heart is complex currently, said Alisa Niksch, currently the chief medical officer of biotech company Genetesis. Niksch is not involved in the Mayo Clinic’s work—but she said their ongoing research into “A-fib,” as well as the early warning signs of heart failure, and prolongation of the QT interval is interesting.

“There is a wider union of cardiac monitoring and AI – which is smart,” said Niksch, formerly the director of Pediatric Electrophysiology and the Pediatric Cardiopulmonary Exercise Lab at Tufts University Medical Center..

However, there are limiting factors – especially in the way that you don’t want the machines to over-diagnose patients who may be perfectly healthy. Rigorous validation and testing are needed on diverse populations to further the technology, she said.

“All those things have to be tested—it’s an arduous process,” Niksch told MedTech Boston. “The neural networks learn so quickly even one error can be amplified to an unknown degree in a matter of days.”

Similar posts

Leave a Reply

Your email address will not be published. Required fields are marked *

Follow us!

Send this to a friend