Smartwatch Signals Detected Hypertension With 65.8% Sensitivity

A consumer smartwatch collected pulse and motion signals over several days and let an AI system flag people who might have hypertension without first calibrating against an arm cuff. In a prospective five-site study, the system detected 65.8% of hypertension cases while correctly clearing 90.0% of participants who did not have the condition.

Key Number: 65.8% sensitivity. The smartwatch identified about two-thirds of participants with hypertension, compared with 55.3% for the initial office measurement, but the difference was not statistically significant.

Looking for a pattern, not a single pressure reading

The system does not estimate a moment-by-moment systolic or diastolic pressure. Instead, it looks for a persistent physiological signature across several days. Optical photoplethysmography measures changes in blood volume at the wrist, while the accelerometer captures movement and subtle mechanical vibrations produced as the heart ejects blood.

A neural network converted overlapping 15-second waveform segments into numerical representations. The researchers averaged those representations over time, combined them with participant height and used a classifier to produce a hypertension probability. Height acts as a proxy for the path travelled by the pulse through the arterial system.

Pretraining used 492,347 hours of unlabeled sensor data from 2,109 consenting participants. The final prospective validation enrolled 196 people across five US sites who had no known cardiovascular disease and were not taking blood-pressure medication.

Comparison with 24-hour monitoring

Researchers evaluated both the smartwatch and office blood-pressure readings against 24-hour ambulatory blood-pressure monitoring, which repeatedly measures pressure during normal activity and sleep. The smartwatch reached 65.8% sensitivity, 90.0% specificity and an 80.6% positive predictive value during seven days of monitoring.

The first office measurement reached 55.3% sensitivity, the same 90.0% specificity and a 77.8% positive predictive value. The point estimates favored the watch, but the Nature press summary states that the difference in sensitivity was not statistically significant. The evidence therefore supports comparable screening performance in this sample, not superiority.

Among 48 participants whose first office reading was elevated, the watch was also tested as a follow-up screen. That smaller analysis produced stronger point estimates, including 78.4% sensitivity and 90.9% specificity, but its size makes the result less stable and unsuitable for broad conclusions.

A notification should lead to a cuff

The intended role is opportunistic screening. A device already worn for other reasons could notice a multi-day pattern and suggest that its user obtain a validated measurement. It is not supposed to declare a diagnosis, adjust medication or reassure someone that their pressure is normal.

That distinction matters because a sensitivity of 65.8% means many people with hypertension would receive no signal. Earlier analysis of commercial smartwatch notifications has warned that missed alerts can create false reassurance if users treat silence as evidence of normal blood pressure.

The same separation between automated triage and diagnosis applies elsewhere in medical AI. NewTqnia recently covered AI-assisted HPV screening used to refine who needs closer cervical assessment. In both cases, the useful output is a reason for further testing, not a final clinical answer.

What should be tested next?

A larger independent study needs to include people across more ages, health conditions and medication histories, then measure what happens after an alert. The important endpoint is not simply whether the watch produces a score, but whether its prompts lead to accurate confirmation without causing false reassurance, unnecessary anxiety or avoidable clinical visits.