Smart Healthcare · global
Detecting Hypertension Through Wrist Signals: AI Smartwatch Screening Shows Promise but Still Misses About One-Third of Cases
A prospective study involving 196 participants found that a smartwatch system requiring no cuff-based calibration performed similarly to initial clinic screening in identifying hypertension. It could encourage more people unaware of their risk to visit a clinic, but it is not yet sufficient to confirm or rule out hypertension on its own.
Hypertension often has no obvious symptoms, and many people discover that their blood pressure has been persistently elevated only through an incidental measurement. If a watch worn daily could identify risk, screening could extend beyond the clinic. A new prospective clinical study provides preliminary evidence for this idea: AI analysis of wrist sensor signals performed similarly to conventional initial blood pressure screening in the clinic, but a substantial gap remains in reliably identifying all patients.
A research highlight published in *Nature Medicine* on November 19 describes a study published in *Nature* on November 5. The team, comprising Google researchers and clinical collaborators, used existing sensors in consumer smartwatches to explore how to identify potentially undiagnosed hypertension without first calibrating the system using a blood pressure cuff.
The method outputs a probability of hypertension, rather than systolic and diastolic blood pressure readings at a particular moment. Optical sensors record signals reflecting changes in blood volume in small vessels at the wrist with each pulse; an accelerometer helps identify periods of stillness and captures subtle vibrations caused by the heart ejecting blood. The AI aggregates signals collected over multiple days and incorporates the wearer's height to reduce the influence of interference such as brief movements on its assessment.
According to Google Research, initial model training used nearly 500,000 hours of sensor data from 2,109 people. For clinical validation, 196 participants were recruited separately at five sites in the United States; they had no known cardiovascular disease and were not taking blood pressure medication. Participants underwent seven days of smartwatch monitoring during daily life. The study then used 24-hour ambulatory blood pressure monitoring as the reference standard for determining hypertension, while also comparing the performance of clinic measurements.
In the initial screening setting, the smartwatch system had a sensitivity of 65.8%, compared with 55.3% for clinic measurements; both had a specificity of 90%. Sensitivity reflects the ability to identify people who truly have the condition, while specificity reflects the ability to correctly rule out those who do not. Although the smartwatch identified more cases numerically, the difference was not statistically significant, so these results cannot support a claim that it outperforms clinic screening.
Viewed another way, a sensitivity of 65.8% also means that the system still misses about one-third of hypertension cases. Not being flagged cannot be interpreted as having normal blood pressure, while a specificity of 90% also means that some people without hypertension may receive a risk alert. These results support its use as a tool to guide further testing, but are not yet sufficient to support a role in standalone diagnosis.
The study also analyzed 48 participants whose blood pressure was elevated at their initial clinic visit. Google Research reported that, in this smaller confirmatory testing subgroup, the smartwatch system had a sensitivity of 78.4%, a specificity of 90.9%, and a positive predictive value of 96.7%. However, these participants had already been selected through initial clinic screening, and their risk profile differed from that of general smartwatch users. The high positive predictive value in this subgroup cannot be applied directly to the general population.
The significance of this study lies in demonstrating the feasibility of looking for signs of sustained hypertension in everyday signals. Before widespread use, performance still needs to be confirmed in larger, more diverse populations, and it remains to be determined whether risk alerts lead to actual medical visits and diagnoses. The research team also emphasized that this remains an early feasibility study. Whether wrist signals can improve cardiovascular health will depend on whether alerts are followed by formal blood pressure assessment and care.