Smart Speakers Could Save Thousands of Cardiac Arrest Victims Each Year
Lane Long posted on July 12, 2019 |
Image Credit: Sarah McQuate, University of Washington
Image Credit: Sarah McQuate, University of Washington

University of Washington researchers are looking to leverage smart speaker technology to “hear” when people’s hearts stop beating. The growing proportion of homes equipped with such devices, they hope, will provide a frictionless early detection system to many future victims of cardiac arrest. Developed by applying machine learning tools to a large dataset of real cardiac arrest patients, the new skill will enable speakers to accurately detect when an event is occurring and alert the necessary parties.

Agonal Breathing

Agonal breathing, a sign frequently exhibited by those experiencing cardiac arrest, produces a distinctive set of noises that can be detected by speaker-enabled platforms like Amazon’s Alexa and Google’s Home. “It’s sort of a guttural gasping noise, and its uniqueness makes it a good audio biomarker to use to identify if someone is experiencing a cardiac arrest,” said co-author Dr. Jacob Sunshine. Around half of all cardiac arrest patients will take agonal breaths, so passive detection of this indicator has good potential to save lives. Given that cardiac arrest can occur at any time, including when people are often alone, the UW team sees having a system “on-call” to notify first responders 24/7 as a high impact solution.

Development Process

Alexa may soon be able to “hear” if its owners’ hearts stop beating. Image credit: Health24.
Alexa may soon be able to “hear” if its owners’ hearts stop beating. Image credit: Health24.

The researchers used a process centered around real audio data gathered from cardiac arrest victims to “teach” their system to pick out agonal breathing. They chose to use 911 calls placed to Seattle-area first responders as an initial dataset for a simple reason. Bystanders placing these calls are often asked by the operator to place the phone next to the patient precisely because the sounds of agonal breathing are so distinctive. The clear, purposeful recording of agonal breathing events were fertile soil for the development of the algorithm.

The team used 162 different calls for help over an 8-year period as a building block. They then split apart agonal breathing incidents into 2.5-second clips to broaden the number of audio samples. Next, they played back the recordings through smart hardware that has grown ubiquitous in the home, like the iPhone 5S and Alexa. Finally, with the help of various machine learning mechanisms, they broke down their clips still further into over 7,300 positive agonal breathing recordings to complete the “positive dataset.”

To teach the skill to pick out actual agonal breathing events versus normal breathing patterns that are often obfuscated by things like background noise or snoring, they built a “negative dataset” comprised of 83 hours of normal sleep with all the attendant sounds. They fine-tuned their algorithm by juxtaposing the two audio sets against one another and noting which noises truly indicated cardiac arrest versus more normal sounds. Introducing still further permutations to the algorithm by including common background noises and manipulating distances helped them dial in the detection mechanism to the point that it functions as a reasonable proof-of-concept.

Results, Further Testing, And Roll-Out

The current iteration of the technology is able to detect, on average, agonal breathing events 97% of the time as long as those events occurred within 20 feet of the device. While that’s a promising start, the UW scientists know more refinement of the algorithm is necessary. The false positive rate is still a too-high .22%, which could undermine the tool’s effectiveness by desensitizing loved ones to a notification.  To correct this,they plan to access more 911 calls in places beyond the Seattle metro area to improve their tool’s accuracy across longer distances, more background noise, and different sound profiles from person-to-person. By expanding the dataset and applying their machine learning techniques to more real audio samples, they hope to be able to roll out a commercially viable tool within a few years.


Recommended For You

Recommended For You