A smartwatch that can't read blood pressure spots hypertension anyway

In 2010, a Harvard-MIT graduate student named Ming-Zher Poh taught an ordinary webcam to read a pulse from a face. Every heartbeat pushes a little extra blood into the vessels under the skin and changes how the skin reflects light. His software picked up the difference. He told Harvard’s student paper it was “a proof of concept” and listed what he wanted to measure next: breathing, blood oxygen and blood pressure.

Sixteen years later, on November 5, 2026, Poh and his colleagues at Google Research published a paper in Nature about a smartwatch. For about a week, 196 people wore a Google watch, and an algorithm flagged 65.8% of those with high blood pressure. A standard reading in a doctor’s office caught 55.3%. Yet the watch never measured blood pressure at all. It has no idea what your numbers are. It looks at the shape of your pulse, the faint kick each heartbeat gives your wrist, and your height.

The man who set out in 2010 to teach machines to measure blood pressure ended up finding hypertension by skipping the measurement altogether.

A disease that doesn’t hurt

By the World Health Organization’s estimate, 1.4 billion people aged 30 to 79 had hypertension in 2024, about one in three. Roughly 600 million of them don’t know it. The gap is widest among the young: the study’s authors note that awareness is lowest in adults aged 18 to 44.

The reason is simple. It doesn’t hurt. High blood pressure can go years without a single symptom, so the only way to find it is to strap on a cuff. Usually that’s a doctor at a checkup, once a year or less. A reading in the office is a snapshot. Some people’s pressure spikes from nerves the moment they sit down across from a doctor; others look fine in the clinic and run high at work and at night. The snapshot misses the other 525,599 minutes of the year.

Wearables have tried to fill that gap for years and kept hitting the same wall: calibration. A cuffless watch could track how your pressure rose and fell from a baseline, but someone had to reset that baseline every so often with a regular cuff. Without it, the watch had no way to say “this person has hypertension.”

A pulse, a recoil and your height

Google came at the problem sideways and stopped chasing numbers. The model, called AI-PPG-ACC-HTN, doesn’t estimate pressure in millimeters of mercury. It answers a different question: how likely is it that this person’s blood pressure is persistently high?

The data come from two sensors found in nearly every smartwatch. An optical sensor shines light into the skin of the wrist and records the pulse wave: how the vessels swell with each beat. The accelerometer, the watch’s motion sensor, picks up a ballistocardiogram. Each time the heart pumps blood into the aorta, the whole body recoils slightly, a bit like a rifle kicking back after a shot, and the wrist gets a barely perceptible nudge. The signal shows up only when you sit or lie still, and the watch finds those moments by itself. The third input doesn’t come from a sensor at all. It’s your height, a rough stand-in for how far the pulse wave has to travel through your arteries.

The model learned without any medical labels. The researchers fed it 492,347 hours of recordings from 2,109 volunteers, chopped into 15-second slices, and left it to work out how one pulse differs from another. Think of someone who listens to a foreign language for thousands of hours without a dictionary until they can hear when a sentence sounds wrong. A simple formula sits on top: it turns several days of recordings, plus height, into a final probability.

Afterward, the authors checked what the model had picked up on. In people at higher risk, the pulse wave was slightly narrower and its leading edge rose more slowly. A narrower wave may point to stiffer arteries. In a stiff tube, a pressure wave travels faster and bounces back sooner. That’s why the length of the path, and with it your height, matters to the model. The body’s recoil looked different too: certain peaks in the recording were deeper and wider. No doctor would see that by eye. Statistics over thousands of hours can.

A week on the wrist versus a minute in the clinic

The trial ran at five clinics in the United States, in partnership with Beth Israel Deaconess Medical Center and Harvard Medical School. It enrolled 196 people with no known heart disease who weren’t taking blood pressure pills. For roughly a week, they wore the watch and went about their lives. The benchmark was 24-hour ambulatory monitoring: a cuff stays on the upper arm for a full day and inflates itself at regular intervals, at work, at home and during sleep. Doctors treat it as the most reliable way to make the diagnosis.

A test like this answers two separate questions, much like a smoke detector: how many real fires does it notice, and how often does it shriek over burnt toast? The watch noticed 65.8% of people with hypertension; the first office reading, 55.3%. Both raised a false alarm for about one healthy person in ten. When the watch did go off, the alarm was real 80.6% of the time, though that share also depends on how many participants had hypertension to begin with. One important caveat: the gap between the watch and the doctor isn’t statistically significant. With 196 people, a ten-point gap could be chance. The honest conclusion is more modest: the watch did no worse than the clinic.

It’s a familiar scene. The doctor gets a high reading and suggests you come back another day, in case it was just nerves. The study tested exactly that. For 48 participants, the first office reading came back high, but 24-hour monitoring confirmed hypertension in 37 of them, while the other 11 actually had normal pressure. So what best separates real hypertension from white-coat jitters: a second visit, a week with the watch, or a home cuff?

The sharpest eye here belonged to the humble home cuff, used several days in a row. The watch raised the fewest false alarms: just one. A second trip to the doctor lost to both.

Apple already sells it, Google is hedging

The idea isn’t new. On September 11, 2025, the US Food and Drug Administration cleared hypertension notifications on the Apple Watch. These don’t measure pressure either: the optical sensor gathers data over 30 days, and if the signs persist, the watch suggests checking your blood pressure. In the study Apple submitted to the FDA, the watch went off for 41.2% of people with hypertension and stayed quiet for 92.3% of those without it. Apple expected more than a million people with undiagnosed hypertension to get a notification in the first year.

The Apple and Google numbers can’t be compared head to head. Apple’s benchmark was a home cuff, not 24-hour monitoring, and the participants and time frames were different too. But the ballpark is clear: a watch on the wrist catches somewhere between two in five and two in three cases and misses the rest.

Google already has a feature of its own. In August 2026 the company unveiled the Pixel Watch 5 and a suite of tools called Health Guardian. Among them are cuffless blood pressure trends on the Pixel Watch 3, 4 and 5 and the Fitbit Air, with one data point at the start of each month. The blog post on the new paper says the research helped inform that feature. It also calls the methods exploratory, with no bearing on current or future products. Both claims sit in the same post.

What the watch can and can’t do

Google’s study left out people with heart disease and those already on blood pressure medication. That makes sense: the goal was to find people who don’t know they’re hypertensive. Still, the paper can’t tell you how the model behaves in people with heart conditions or in anyone already being treated. And the company that built the algorithm also tested it: Google employees are among the authors, alongside Harvard physicians.

What the study does offer is new. A watch with no cuff calibration at all went up against the strictest benchmark and an ordinary office visit, and it didn’t lose. For a disease that 600 million people don’t know they have, that is a real step.

So what does this mean if you wear a watch? A silent watch doesn’t mean your blood pressure is normal. Apple’s FDA documents say so outright, and Google’s watch missed nearly one in three people with hypertension in the study. The most sensitive tool in this research was an ordinary blood pressure cuff, used at home for several days. So the watch’s real job is to get people to that cuff.

The watch from Poh’s work still doesn’t measure blood pressure. But it does what he described to the student paper back in 2010: it lets people “collect long term data for days or months.”