At the onset of a heart attack, people often have a few telltale symptoms: pain in the chest or arm, heartburn, shortness of breath. But only one of these is immediately obvious to the outside observer, who may hear a gasping sound known as agonal breathing.
Wouldn’t it be great, then, if the devices we own could listen for this critical indicator and take action? That’s exactly what researchers at the University of Washington have done with a tool for smart speakers such as Google Home and Amazon Echo that can detect abnormal breathing and call for help.
Heart attacks strike every 40 seconds in the U.S. and affect more than 700,000 people annually, according to the CDC. The researchers found that around half of people experiencing a heart attack made the signature gasping sounds while on a call with 911.
“The good news is that if you can detect these agonal breathing sounds, you can almost double or triple someone’s chance of survival if you’re going to give them immediate CPR,” said Shyam Gollakota, a professor at the University of Washington and one of the authors on the study.
Following the proof-of-concept study, the next step for the team is to develop the skill through Sound Life Sciences, a Seattle startup that is already working on another application that uses a smartphone’s microphone and speaker to spot opioid overdoses. Gollakota also serves as Sound’s CEO.
The Apple Watch last year debuted a feature that lets users run an ECG right from their wrist. The information can indicate problems that could lead to a heart attack, but the device doesn’t monitor your heart around the clock. Heart rate monitors used in hospitals are too clunky for everyday wear, though AI-driven wearable versions are becoming available.
When it comes to monitoring health, smart speakers have two key advantages over smartphones and wearables: they’re always listening and connected to power, making them more reliable.
To obtain the audio for their algorithm, the researchers relied on real 911 calls to emergency medical services in the Seattle area. They extracted a total of 236 agonal breaths and re-captured those through smartphones and smart speakers with other household sounds. The researchers then used machine learning techniques to grow the number of heart attack-indicating breaths to more than 7,000.
For the control group, they used 83 hours of audio data from sleep studies to derive 7,305 sound bytes of normal sleep sounds.
The machine learning algorithm was then put to the test. The tool correctly identified the telltale gasping breath 97 percent of the time. The ordinary sleeping sounds were misclassified 0.14 percent of the time.
Gollakota worked with fellow UW researchers and professors Justin Chan, Thomas Rea and Jacob Sunshine — who is also on the Sound Science team — for the study.
Gollakota, a leader in the field of wireless networking who previously co-founded Jeeva Wireless, thinks the skill would be useful “any place where you don’t want people to be wearing devices,” which could include hospital wards and nursing homes. Once abnormal breathing is detected, the app could alert someone who is in a position to administer CPR quickly.