Imagine a smart watch that surreptitiously scans your body for telltale signs of disease. Over time, it detects a quiver in your heartbeat – the tell-tale pattern of a common heart condition. An alert prompts you to seek out further testing for atrial fibrillation, an irregular heartbeat which otherwise might have lurked silently until it caused a stroke.
That’s the tantalizing intersection of wearable consumer technology and medicine that lies in the future. But according to a new study that married the cutting-edge of artificial intelligence with the Apple Watch’s sensor data, we’re not there yet.
Researchers from the University of California, San Francisco, set out to test the viability of using the Apple Watch to detect signs of atrial fibrillation, which is a major cause of stroke. In the scenario that was closest to a real-world use of the technology, people that tested positive for atrial fibrillation had only an 8 percent probability of actually carrying the diagnosis. The results were, according to an accompanying editorial in JAMA Cardiology, “humbling.”
“It really just doesn’t perform,” said Eric Topol, a cardiologist at the Scripps Research Institute who was not involved in the study. “This doesn’t pass muster for use in detection of atrial fibrillation.”
That doesn’t mean the idea isn’t an exciting one, or that limitations of the study can’t be overcome. The study was seen as a proof-of-concept that screening tools could be taken out of hospitals and deployed in people’s everyday lives, not as a failure. Already, the UCSF researchers are working to address the limitations and continue the work.
And the space is growing. AliveCor, a health-tech company, has developed a mobile electrocardiogram and watch band for the Apple Watch that allow people to actively monitor their heart rates. Apple in the fall announced it was launching a 500,000-person clinical trial with Stanford University to test whether the Apple Watch could be a way to detect irregular heart rhythms and flag signs of atrial fibrillation.
“We have to be very careful about false positives and causing distress when it’s really not needed or adding to health care costs, for example, because of unnecessary testing — which is why I do agree more refinement is needed,” said Gregory Marcus, a cardiologist at UCSF who led the work. “But it’ll be coming. It’s going to get better, and it’s going to be coming soon. This is the first heads-up: your smart watches have the capability of doing this, so it’s coming and it’s theoretically possible.”
At a time when there is almost boundless excitement around the potential for consumer technologies to make medicine better, the study shows that entrepreneurs and doctors won’t just be able to send sensor-laden devices out into the world to transform medicine. They will need to figure out the appropriate uses, improve and tune the technology so it won’t give false reassurance or cause unnecessary alarm – and figure out how to connect data to useful interventions.
The researchers started with data harvested from the Apple Watch, from healthy people and those with atrial fibrillation. A neural network, essentially a computer system modeled on the human brain, was set loose on the data, to learn the difference between the two groups. That allowed the researchers to develop algorithms to predict who had atrial fibrillation and who did not based on the Apple Watch sensor data.
The good news: when they put Apple Watches on a very select population of 51 atrial fibrillation patients who weren’t moving around and were about to undergo a procedure to shock their irregular heartbeats brought back into rhythm, the algorithm worked wonderfully well. The bad news: When researchers tried to test their algorithm on data from about 1600 people in less pristine conditions, as they wore Apple Watches in their daily lives, its ability to correctly predict atrial fibrillation plummeted.