What do the Apple Watch and Nokia Pulse Ox have in common? They’ve both
got pulse oximeter sensors that measure heart rate using
photoplethysmography (PPG), the expansion and contraction of capillaries
based on changes in blood volume. They’re accurate to a degree, but
require a fair amount of electricity because they’re light-based — they
emit a signal onto the skin that reflects back to a photodiode.
One battery-saving alternative might be accelerometers, a sensor
commonly found in smartphones, smartwatches, and activity trackers that
measures non-gravitational acceleration. In a paper published on the
preprint server Arxiv.org, researchers at Philips Health and the
University of Bristol describe a machine learning algorithm that can
predict heart rate almost exclusively from the sensors, boosting the
battery life of the wearable to which they’re attached.
“Consumer PPG sensors typically consume up to 5000 times the power than
the accelerometer used in wearables, which is an impediment to the long
battery life desired in wearable technology,” the researchers wrote. “As
accelerometers are widespread and exist in any device which would
likely also contain a heart rate sensor, we are interested in
considering the feasibility of acceleration as a means of predicting
heart rate.”
They tapped data from test subjects participating in the EurValve
project, a multiyear clinical study of patients who have undergone heart
valve replacement surgery. Each sports a wearable with an accelerometer
(with a three-week battery life) and a Philips Health monitor with a
pulse oximeter (with a four-day battery life), and had a custom-designed
compute unit — the Smart Home in a Box (SHiB) — installed in their home
that receives and processes data from both wearable devices.
The researchers trained two machine learning models. The first was a
baseline: a regression model that relied exclusively on data from the
accelerometer, aligned it with wearers’ heart rates, and attempted to
predict future heart rates. The second model, which could run on the
SHiB units, took an “active learning” approach that allowed it to pull
data from either health monitor, depending on the situation.
“This approach will predict heart rate from the streaming accelerometer
data in an online fashion and be able to request the measurement of true
heart rate via PPG when required,” the team wrote.
They employed a few clever tricks to cut down on energy use. The second
model learned to assume that particular acceleration patterns, like
walking or jogging, indicated that heart rate is likely to increase, and
intelligently decided whether to measure heart rate using the
accelerometer data or pulse oximeter data.
“Typically, in active learning problems it is possible to query the
label … of samples, specifically for data samples for which the label
will be particularly useful,” the team wrote. “This is however not
feasible in our setting, where acceleration data is arriving constantly
and we wish to consistently produce a heart rate estimate, and it is not
possible to retrospectively measure the heart rate.”
The researchers evaluated the active learning model on three patients,
each with four weeks’ worth of data collected two months apart. The mean
absolute error (MAE, or the distance between two continuous variables)
was between just 2.5 and 5 heartbeats per minute, and the energy savings
were significant. In one example when the heart rate sensor was queried
20.25 percent of the time, MAE was 2.89.
That’s good news for fitness fanatics and smartwatch fans alike.
https://www.geezgo.com/sps/30613
Comments
Post a Comment