Google Pixel Watch Detects Heart Attacks

Later this month, Google is expected to roll out software on its Pixel Watch 3 in the United States that has the potential to correctly identify two-thirds of out-of-hospital cardiac arrests in people wearing the smartwatch. The feature uses AI to detect when the wearer no longer has a pulse, and it’s meant to combat the quiet killer of cardiac events that occur at home when people are alone and unable to call for help.

A team of Google researchers and scientists at the University of Washington recently published a study testing the software, with the aim of balancing the need for a low number of false positives—when 911 might be called but not needed—with the desire to identify a loss of pulse in as many cases as possible.

“You can make this more sensitive, but it just comes at a cost,” says Google research scientist Jake Sunshine who led the study. An algorithm that “excessively” calls 911, Sunshine says, “can’t exist in the world like that.”

The study was released by the journal Nature as an accelerated article preview on 26 February, the day after the FDA announced premarket medical device approval of the loss of pulse feature on the Pixel for the company Fitbit, which was acquired by Google’s parent company Alphabet in 2021.The feature was approved in Europe last year, and is expected to become available to people in the U.S. this month.

Training Pixel’s AI

Data from three cohorts of Pixel Watch wearers was used to train the model. The first cohort included 100 patients with an implanted cardiac defibrillator, which delivers small pulses of electricity to the heart when it detects irregular heartbeats. The patients wore a Pixel Watch when their heart temporarily stopped during a scheduled test of their defibrillator.

But that data was hard to get because the tests had to be done under medical supervision. “We can’t just take healthy volunteers and make their heart stop, and then send them on their way home, right?” Sunshine notes. So, the team turned to a second cohort of 99 participants who experienced a temporary loss of pulse when a tourniquet tightened on their arm. The two signals—or lack thereof—recorded on the wrist of defibrillator patients looked “indistinguishable” from the signals of people with a tourniquet wrapped around their arm. The third and largest cohort included nearly 1,000 Pixel Watch wearers living their daily lives.

The model was trained to identify the transition between a regular heart rhythm and loss of pulse. Part of the model processed the pulse signal to identify if the amplitude dropped and if the accelerometer detected any movement. Another part of the model used neural networks to quickly run through more than 500 signal features in order to confirm that a transition between pulse states took place.

But these signals can also occur when a wearer simply falls down or lies in an awkward position, Sunshine says. The model needed to proceed through additional checks before calling 911.

After identifying a possible loss of pulse, the watch turns on an infrared light that penetrates deeper into the skin than the standard green light thatis always on to detect pulse. The watch searches for a pulse as the green and infrared lights flood the wrist. At the same time, another algorithm checks that the pulse detected, if there is one, matches the regularity of a beating heart.

Finally, a “quite annoying” haptic buzz with an irregular pattern is turned on, Sunshine says. If the wearer is still motionless after 35 seconds of buzzing, then 911 is called. The goal is for classification to occur in around one minute.

The algorithms in Google’s Pixel Watch look for changes in pulse amplitude that might be a sign of cardiac arrest.
Google

Specificity Over Sensitivity

After training the algorithm in these controlled settings, Sunshine and his colleagues tested the feature on 355 Pixel wearers outside the lab, yielding one errant call to 911. The team also tested the model back in the lab on a new set of participants using the tourniquet technique to temporarily pause their pulse. There, the model correctly identified a loss of pulse in 67 percent of the more than 1,000 sessions of tourniquet-induced pulselessness conducted in the lab by 156 participants (21 of whom were professional stunt persons). This means that the algorithm did not catch around one third of cases where there was a loss of pulse.

The decision to maximize specificity over sensitivity is understandable to Mahsa Khalili, a postdoctoral researcher at the University of British Columbia, who studies out-of-hospital cardiac arrests and was not involved in this work. Similar to other models, this one will likely improve as more data comes in from the U.S.-based users who opt in to the feature, she says.

While many academic labs are limited by the number of participants willing to enroll in cardiac monitoring studies, Google is uniquely resourced and situated to reach many more end users, Khalili adds. The Google research team made the details about the participant numbers and system architecture were made available, but only published pseudo code of the model itself.

The opt-in feature is geared toward the general population. But like all medical devices, it is not for certain populations, such as people with severe cardiac disease, Sunshine says. He and the team expect to evaluate the real-world data as it comes in and disseminate the findings.