DENVER -- A new study out of the University of California in San Francisco found the Apple Watch may be able to detect a heart condition that causes over 100,000 strokes every year.
Cardiogram, a heart-health app, and researchers from UCSF Cardiology Health teamed up to create the mRhythm study. It was focused on how effective the watch could be at tracking the most clinically common heart abnormality, atrial fibrillation (Afib).
If you’re like one out of every five Americans, you wear a Fit Bit or an Apple Watch. You can already track your heart rate, but now the technology might go one step further: preventing strokes and possibly saving lives.
The heart beats more than 100,000 times a day, adjusting to exercise, eating and sleeping. But for 3 million Americans, like 69-year-old Rick Barnes, there’s a problem.
The study is something that Barnes finds fascinating. The findings showed the wearable detected Afib, an irregular heartbeat, accurately 97 percent of the time. Luckily, he was to be able to recognize the signs.
“I first noticed it in the swelling of my ankles, and of course that was because I was in Afib and my heart wasn’t pumping blood throughout my body properly," said Barnes.
He used to do construction work and continues to be active. Monitoring his heart is important, especially as he gets older.
Not everyone can spot the signs of Afib. It's one of the toughest conditions to diagnose in a doctor's office, too. That’s where the mRhythm study’s finding could come into play in the future.
A spokesperson for the mRhythm study provided the following information to Denver 7:
Two-thirds of those are preventable with inexpensive anti-coagulants. While existing medical devices such as Holter Monitors or wireless patches can monitor heart rhythm for 24 hours to 4 weeks, recent studies show that even in a high-risk population, it takes 84 days of monitoring to find the first episode of atrial fibrillation. That means AF frequently goes undiagnosed. Since 1 in 5 Americans owns a fitness wearable such as a Fitbit or Apple Watch, if those wearables can be repurposed into continuous, long-term heart monitors, then we detect atrial fibrillation early, guide people to appropriate medical care, and thereby prevent strokes.
This could transform the work Dr. Larry Allen does at the University of Colorado Hospital, where Barnes gets treatment.
“If you’re wearing your watch and it tells you have a problem and you get into see me earlier when you wouldn’t have otherwise,” said Allen.
He told Denver7, 40 percent of his heart failure patients suffer from Afib.
“That’s exciting because I think then we are really preventing problems before they happen,” said Allen.
That can be the difference between life and death.
“To have a monitor to look at. I think could can save not only someone having a stroke but maybe even their life,” said Barnes.
The results of the study were presented in May, at Heart Rhythm Society, the top conference of cardiac electrophysiologists. 6,158 participants were recruited via Cardiogram for Apple Watch into the UCSF Health eHeart Study (health-eheartstudy.org). Roughly 200 participants with diagnosed paroxysmal atrial fibrillation were mailed a mobile electrocardiogram and asked to take at least one reading per day and whenever they felt symptoms. Engineers at Cardiogram then trained a deep neural network to identify atrial fibrillation from Apple Watch heart rate data.
This is the third major study of deep learning in medicine, following Google Brain's results on diabetic retinopathy in December and Stanford's study on skin cancer in January. One major contrast with previous studies of deep learning in medicine is that we invited both healthy and unhealthy participants to enroll, and designed a semi-supervised deep neural network capable of learning from both groups.
In the first (unsupervised) phase, the neural network was trained to predict heuristics of heart rate variability; in the second (supervised) phase, it was trained to predict the output of a mobile ECG on participants with diagnosed atrial fibrillation. By taking a semi-supervised approach, we can train an accurate neural network with much less hand-annotated data than would otherwise be required.