Google Study Finds Smartwatch AI Can Flag High Blood Pressure Without a Cuff

In a prospective trial at five US sites, the system identified people with hypertension with accuracy comparable to standard cuff-based screening, the researchers said. The work, which Google calls AI-PPG-ACC-HTN, was checked against 24-hour ambulatory blood pressure monitoring, the clinical reference standard.

The stakes are large. Hypertension affects more than one billion people worldwide, and more than half of adults who have it do not know, the researchers wrote, with the gap widest among people aged 18 to 44. Because the condition rarely causes symptoms, it is usually caught only through incidental screening. Commercial cuffless wearables mostly track changes against a baseline cuff reading, which limits their use as stand-alone screening tools.

How the System Works

Rather than trying to estimate exact systolic and diastolic readings at a given moment, which the researchers said are easily thrown off by noise, the system tries to identify whether a person is in a persistent hypertensive state by looking at signals gathered over several days.

It uses two sensors common on modern watches: the optical photoplethysmography sensor, which reads changes in blood volume at the wrist, and the accelerometer, which can pick up the tiny vibrations caused by the heart ejecting blood. A convolutional neural network was first trained without medical labels on 492,347 hours of sensor data from 2,109 consenting participants, learning to recognise 15-second waveform segments from the same person. For each user, those features are averaged across days of wear, combined with the person's height as a stand-in for arterial length, and fed into a classifier that produces a hypertension probability score.

The Clinical Results

The study enrolled 196 people with no known cardiovascular disease who were not taking blood pressure medication. As an initial screening tool, the watch achieved sensitivity of 65.8%, specificity of 90.0% and a positive predictive value of 80.6%. Office readings in the same group had sensitivity of 55.3% and a positive predictive value of 77.8%.

The researchers also tested the system as a confirmatory tool in 48 participants whose initial office readings were high. There it reached sensitivity of 78.4%, specificity of 90.9% and a positive predictive value of 96.7%, against 67.6%, 63.6% and 86.2% for repeat office measurements. Multi-day home cuff monitoring still detected more cases, with sensitivity of 89.2%, though its specificity of 81.8% and positive predictive value of 94.3% were lower than the watch's.

Google said the findings helped inform the blood pressure trends feature in Health Guardian, part of the Google Health app. The researchers were careful to frame the work as early feasibility research rather than a diagnostic product. They said smartwatch screening is not meant to replace arm cuffs or to diagnose hypertension on its own, but could serve as a form of "opportunistic screening" that prompts people at risk to seek a formal assessment from a clinician.

The blog post did not name the smartwatch model used in the study.