Smartwatch high blood pressure detection: Google's study

On November 5, Google Research published a Nature paper with doctors from Beth Israel Deaconess Medical Center and Harvard Medical School. It describes an AI model that reads the pulse sensor and motion sensor on a regular smartwatch and flags likely high blood pressure, with no arm cuff needed to calibrate it. The result Google and Nature’s own summary lead with is a comparison: 65.8% against 55.3%.

That second figure is what a single blood pressure reading at an office visit catches. Google’s post calls office measurement the primary initial screening gateway. My read: look closely and this is less a story about a watch beating a clinic than about how little one office reading catches.

What the watch actually does

It doesn’t produce a systolic or diastolic number. According to Google Research’s write-up, the system estimates whether someone has persistent hypertension by averaging what the sensors record over several days. Participants wore a Pixel Watch 3 for seven days, sleep included, according to the study’s preprint.

The model was first trained without medical labels on 492,347 hours of sensor data from 2,109 people, Google says. Then each person’s averaged signal is combined with their height and fed to a regularized logistic regression that outputs a probability. Google says existing cuffless wearables need regular cuff readings to calibrate, so they only track changes from a baseline. This one skips that step.

The one number: about eight people

The test group was 196 adults at five US sites, none with known cardiovascular disease and none on blood pressure drugs. The reference was a 24-hour ambulatory monitor, the clinical gold standard, according to Google. Here is what the Nature abstract reports for the first scenario, initial screening.

Measure	Smartwatch AI	One office reading
Sensitivity (cases caught)	65.8%	55.3%
Specificity (healthy people cleared)	90.0%	90.0%
Positive predictive value	80.6%	77.8%

My arithmetic, not a figure from the paper: those percentages only fit together if about 76 of the 196 people had hypertension and 120 didn’t. On that base the office reading caught about 42 of the 76 and the watch about 50. I checked it against all six percentages and it matches each one, but treat it as a back-of-envelope count and look at the paper’s own tables.

So the watch found roughly eight more people than the office check did. Medical Xpress, summarizing Nature Publishing Group’s release, says the difference was not statistically significant, and the 95% confidence intervals in the abstract (54.0% to 76.3% for the watch, 43.4% to 66.7% for the office reading) overlap heavily. By simple arithmetic, a 65.8% sensitivity also means the watch missed about 34 of every 100 people who had the condition.

My read: a thin margin.

The follow-up scenario

Google also tested a second use: confirming high blood pressure in the 48 people whose first office reading was elevated. Guidelines call for out-of-office tracking at that point.

Method (48 people)	Sensitivity	Specificity	Positive predictive value
Smartwatch AI	78.4%	90.9%	96.7%
Repeat office reading	67.6%	63.6%	86.2%
Multi-day home monitoring	89.2%	81.8%	94.3%

Home monitoring caught more cases than the watch, 89.2% against 78.4%, which Google’s own post shows. My read: forty-eight people is a small group. The preprint puts the 95% interval on home monitoring’s specificity at 48.2% to 97.7%, so these percentages sit on small counts.

Not a product, says Google

The blog post calls the work “early feasibility research” and says its methods “are not indicative of any existing or upcoming commercial product feature or device capability.” It also says the findings helped inform the blood pressure trends Health Guardian feature in the Google Health app, and doesn’t explain how far that link goes. The preprint adds that participants were told the system was investigational and not a validated medical tool.

For a related look at where wearables are heading, see our post on the Aina smart ring.

What I make of it

My read: the useful claim here is modest and I think it holds up. A watch that says “worth getting a cuff check” costs the wearer almost nothing, and the paper frames it that way, as a prompt toward a clinic and not a diagnosis.

My read: the worrying version is the opposite message. With sensitivity at 65.8%, a quiet watch is not a clean bill of health, and anyone who treats silence as an all-clear would be misled about one case in three.

My read: the evidence is also narrow. It’s one watch model, 196 people, and a study written up by Google’s own researchers, so I’d want an independent team to repeat it before reading much into the eight-person gap. If you’re worried about your blood pressure, a cuff reading or a doctor still answers the question.

Sources

Note on sources: I read the Nature abstract through PubMed, not the full paper, so details such as the Pixel Watch 3 wear protocol and the training data come from the Google post and the earlier preprint, which may differ from the published version.