Google Research detects signs of high blood pressure from smartwatch sensors without cuff calibration
Google Research described on 5 November an AI system that looks for signs of hypertension using only the optical pulse sensor and the accelerometer already found in consumer smartwatches. The work was published in Nature. Unlike existing cuffless wearables, which typically need regular calibration with an arm cuff and track changes from that baseline, the system does not use any cuff reading.
The team, led by Paolo Di Achille and Ming-Zher Poh, chose not to estimate exact systolic and diastolic values at a given moment, which they describe as highly sensitive to transient noise. Instead the model tries to identify a persistent hypertensive state over several days. It combines photoplethysmography, the optical measurement of blood volume at the wrist, with accelerometer data, which captures periods of stillness and the ballistocardiogram, the small vibrations of the wrist caused by each heartbeat.
A one-dimensional convolutional network was trained with self-supervised contrastive learning on 492,347 hours of unlabelled sensor data from 2,109 participants, learning to place overlapping 15-second windows from the same person close together. For each user the window embeddings are averaged over multiple days, combined with height as a proxy for arterial path length, and passed to a regularised logistic regression that outputs a probability of hypertension.
Evaluation took place in a prospective study at five US sites with 196 participants who had no known cardiovascular disease and were not taking blood pressure medication. The reference was 24-hour ambulatory blood pressure monitoring, the clinical standard. For initial screening the watch reached 65.8 percent sensitivity, 90.0 percent specificity and an 80.6 percent positive predictive value, compared with 55.3 percent sensitivity and a 77.8 percent positive predictive value for office measurements. In the 48 participants with elevated office readings, used to test confirmation, it reached 78.4 percent sensitivity, 90.9 percent specificity and a 96.7 percent positive predictive value, while multi-day home monitoring achieved 89.2 percent sensitivity and 81.8 percent specificity.
Google says the findings informed the blood pressure trends feature in the Google Health app.

What it means
The study's strength is the comparison against ambulatory monitoring rather than single cuff readings, and its modest sensitivity means the watch flags people for follow-up rather than ruling hypertension out. Its value is in reaching people who are never screened at all.