Google Smartwatch Model Flags Hypertension Without Cuff Calibration, Nature Study Shows

The model, called AI-PPG-ACC-HTN, does not give a blood pressure reading. It reads optical pulse signals (photoplethysmography) and accelerometer data collected over several days, then estimates the probability that the wearer has sustained hypertension. Google says the aim is opportunistic screening: getting people who do not know they have the condition to see a clinician and have it measured properly.

The potential reach is large. Hypertension affects more than one billion people worldwide, and more than half of adults who have it do not know, according to Google. The company says the gap is especially wide among adults aged 18 to 44, who rarely have their blood pressure checked. Wearables that estimate blood pressure have usually needed regular calibration against a cuff, which limits their use as passive screening tools.

The researchers first trained the model without diagnostic labels, using 492,347 hours of unlabelled sensor data from 2,109 consenting participants. During this self-supervised stage the network learned to tell which segments of signal came from the same person, picking up each wearer's typical pulse and heartbeat patterns. It then averages those patterns over several days to produce a risk estimate.

The team tested the system in a prospective study of 196 people at five US sites. Participants had no known cardiovascular disease and were not taking blood pressure medication. Results were checked against 24-hour ambulatory blood pressure monitoring, which Google describes as "the clinical gold standard for hypertension diagnosis."

At the initial screening stage, the watch correctly identified 65.8% of participants with hypertension. A standard office measurement caught 55.3%. The watch correctly ruled out 90.0% of people without the condition. When it flagged someone, it was right 80.6% of the time, against 77.8% for the office reading.

The researchers also looked at 48 participants whose first office reading was high. In that group the watch's sensitivity was 78.4% and its specificity 90.9%. A repeat office measurement scored 67.6% and 63.6%. Multi-day home cuff monitoring caught the most cases, with sensitivity of 89.2%, though it was less specific at 81.8%.

Paolo Di Achille and Ming-Zher Poh of Google Research led the work. Their co-authors came from Google Research, Beth Israel Deaconess Medical Center and Harvard Medical School.

Accounts differ on how closely the research ties to a Google product. Google's blog says the work "helped inform" the blood pressure trends feature in the Google Health app, part of its Health Guardian tools for Pixel and Fitbit devices. The same post says the methods are exploratory and do not describe any existing or upcoming commercial feature. The Decoder reported that the findings had been folded into the app feature. Google says the research is not meant to replace arm cuffs or provide a diagnosis.

The validation group was small, and it left out people on blood pressure medication or with known heart disease. Both limit what the results show about wider use. The researchers call the study early feasibility work, and larger trials in more varied populations would be needed before the method could be relied on as a screening tool.