Medical Technology · global
Helping Watches Detect Hypertension Earlier: Seven-Day AI Screening Clinically Validated
A study of 196 people shows that smartwatches can identify signs of hypertension from everyday pulse and movement signals. But about one-third of people with hypertension may still be missed. The technology’s value lies in bringing potential risks into the clinic.
Hypertension often has no obvious symptoms, and many people discover something is wrong only when their blood pressure is measured. Could an early alert from a watch worn every day shorten this period of unawareness? A study published in *Nature* on November 5 presents clinical validation: artificial intelligence can identify some people who may have hypertension from signals accumulated by consumer smartwatches, providing a gateway to further testing.
The system, called AI-PPG-ACC-HTN, combines photoplethysmography (PPG) and accelerometer data. The former uses reflected light to track changes in blood volume at the wrist with each pulse, while the latter provides signals of movement and subtle mechanical vibrations. The model aggregates recordings over multiple days to identify persistent signs of hypertension. It requires no initial calibration with a blood pressure cuff and does not output moment-by-moment systolic and diastolic blood pressure readings.
According to the Google Research team, the model first underwent self-supervised pretraining using 492,347 hours of sensor data provided by 2,109 consenting participants to learn features of physiological waveforms. It then aggregated signals from multiple days to make individual-level classifications. These data were used to build the model; its screening performance was tested in a separate prospective, multicenter clinical study.
The validation study enrolled 196 people at five sites in the United States. None had known cardiovascular disease or were taking blood pressure medication. Using 24-hour ambulatory blood pressure monitoring as the reference standard, the study compared its findings with the analysis of seven days of watch recordings: the system had a sensitivity of 65.8%, a specificity of 90.0%, and a positive predictive value of 80.6%. In other words, among these participants, about eight in ten people who received a positive screening result met the reference test’s criteria for hypertension.
At the same specificity, initial in-clinic blood pressure screening had a sensitivity of 55.3%. Although the watch system’s figure was higher, the findings support performance comparable to in-clinic screening and do not justify claiming superiority over current methods. A sensitivity of 65.8% also means that, for every 100 people classified as having hypertension by the reference test, about 34 may not be identified by the system. Receiving no alert is therefore insufficient to rule out hypertension.
This evidence also has limits to its applicability. A study of 196 people is not yet sufficient to establish how the system performs in people who already have cardiovascular disease, are taking medication, or belong to other clinical populations. Positive predictive value also changes with the prevalence of hypertension in the population using the system, so the study’s 80.6% cannot be directly treated as the probability that any watch user has hypertension after receiving an alert. Moving from detecting abnormalities to improving health outcomes also requires knowing whether alerts prompt people to seek medical care and whether appropriate management follows.
Google says the findings have helped develop the blood pressure trends feature in Health Guardian within the Google Health app. This is the research team’s account of the product application and does not mean the product has received equivalent validation in every usage scenario. The clinical role proposed in the paper remains clear: the watch provides clues to possible abnormalities, and positive results should be followed by clinical evaluation and confirmation with cuff-based blood pressure measurements. Whether it can expand opportunities for early detection depends on whether this follow-up care pathway can be put into place.