Google’s smartwatch AI flags hypertension without cuff calibration in small study

A week of smartwatch signals can flag sustained high blood pressure without first calibrating against an arm cuff, according to Google Research’s study published November 5 in Nature. In 196 participants, the experimental AI detected 65.8% of hypertension cases, compared with 55.3% for initial office screening. It produces a risk score, not a blood-pressure reading.

The researchers are tackling a detection gap: more than half of adults with hypertension do not know they have it. The condition rarely causes obvious symptoms, so discovery often depends on an incidental screening. Google’s response is to look for persistent physiological patterns across days of ordinary watch wear, rather than trying to reconstruct a single blood-pressure measurement.

Building a risk score from ordinary wrist signals

The system, called AI-PPG-ACC-HTN, combines two sensor streams. Photoplethysmography, or PPG, uses reflected light to track changes in blood volume at the wrist. An accelerometer identifies still periods and captures tiny wrist vibrations caused by the heart ejecting blood into the aorta. Those mechanical signals add another view of cardiovascular activity alongside the optical pulse waveform.

First, researchers trained a neural network on 492,347 hours of sensor data from 2,109 consenting participants. This stage used self-supervised learning: the model learned patterns from unlabeled recordings rather than requiring medical labels for every segment. It learned to place overlapping 15-second optical and motion recordings from the same person close together in its mathematical representation.

The next stage averages those learned signal features across multiple days, reducing the influence of temporary disturbances such as movement or temperature changes. It combines the result with the wearer’s height, used as a proxy for the distance pulse waves travel through arteries. A statistical classifier then calculates a continuous probability of hypertension, rather than systolic and diastolic pressure values.

A week on the wrist, checked against clinical monitoring

The prospective validation enrolled 196 people across five U.S. sites. Participants had no known cardiovascular disease and were not taking blood-pressure medication. Researchers judged the smartwatch’s seven days of real-world monitoring against 24-hour ambulatory blood-pressure monitoring, the clinical reference used to establish hypertension status. The Nature paper compared the watch with both initial screening and subsequent confirmation workflows.

The higher detection rate is a point estimate, not a precise measure of the advantage. The study’s 95% confidence interval for smartwatch sensitivity ran from 54.0% to 76.3%; the office-screening interval was 43.4% to 66.7%. Among participants flagged by the watch, 80.6% had hypertension according to the reference test, compared with 77.8% of those flagged at the initial office screening.

Researchers then examined the 48 participants whose initial office reading was elevated. In this smaller confirmatory group, the AI achieved 78.4% sensitivity and 90.9% specificity. Its positive predictive value—the share of positive results that were correct—was 96.7%. Return office measurements produced lower point estimates: 67.6% sensitivity, 63.6% specificity and 86.2% positive predictive value.

Multi-day home cuff monitoring offered a different balance in that confirmatory analysis. It had higher sensitivity, at 89.2%, but lower specificity, at 81.8%, than the watch. Its positive predictive value was 94.3%. The results therefore do not show the AI leading every measure or removing the role of established cuff-based testing.

Clues in the waveforms, not a finished diagnostic

To investigate what the model might be recognizing, researchers compared average waveforms from people with the highest and lowest predicted risks. Higher-risk patterns included narrower optical pulses, which Google says may suggest stiffer arteries that expand less. They also showed differences in the wrist vibrations associated with heartbeats. These observations offer physiological clues to the score; the researchers describe them as an effort to understand what the model may attend to.

Google calls the work early feasibility research and says the methods do not indicate an existing or upcoming commercial feature or device capability. The intended role is narrower: passively identify people who may need clinical assessment and cuff-based confirmation. It is not a standalone diagnosis or a replacement for clinical arm cuffs.