AI News
All stories
Reported

Google AI detects signs of high blood pressure using smartwatch sensors

THE DECODER reports that a Google AI system detects signs of high blood pressure from smartwatch sensors.

Research · · 3 min read

Google Research has built an AI system that estimates whether a person has high blood pressure using only the sensors already inside a consumer smartwatch. No arm cuff is needed for calibration. The work was published in Nature on November 5, 2026, and Google says its findings have been folded into the blood pressure trends feature in the Google Health app. The paper's own framing is more cautious. The authors call the work early feasibility research and say the methods are exploratory and do not indicate any existing or upcoming commercial feature.

The system uses two sensors. The first is photoplethysmography, an optical sensor that shines light into the wrist and measures how much is reflected back, which tracks the changing blood volume with each pulse. The second is an accelerometer, a motion sensor that, when the wearer is still, picks up the faint wrist vibrations caused by blood being ejected from the heart into the aorta. This signal is called the ballistocardiogram. The researchers first trained a one-dimensional convolutional neural network with self-supervised contrastive learning on 492,347 hours of unlabeled sensor data from 2,109 participants. The model learned to place overlapping 15-second waveform segments from the same person close together, which required no medical labels. Those segment embeddings are then averaged over several days of wear to smooth out noise such as sudden movements. They are combined with the person's height as a proxy for arterial path length, and fed into a regularized logistic regression that outputs a continuous probability of hypertension.

The design choice matters for how the output should be read. Rather than trying to estimate exact systolic and diastolic values from a single moment, which the paper says are highly susceptible to transient noise, the system targets a persistent hypertensive state. Existing commercial cuffless wearables, by the paper's account, usually need regular calibration against an arm cuff. They track changes from a baseline rather than identifying absolute hypertension, which limits their use as independent screening tools.

The validation was a prospective, multicenter study of 196 participants across five US sites who had no known cardiovascular disease and were not taking blood pressure medication. The reference standard was 24-hour ambulatory blood pressure monitoring. In the initial screening scenario, the smartwatch model reached a sensitivity of 65.8 percent, a specificity of 90.0 percent and a positive predictive value of 80.6 percent. Office blood pressure screening in the same group reached a sensitivity of 55.3 percent, a specificity of 90.0 percent and a PPV of 77.8 percent. The paper's confidence intervals are wide. The watch's sensitivity ranges from 54.0 to 76.3 percent and the office reading's from 43.4 to 66.7 percent, so the two intervals overlap.

The confirmatory scenario rests on a smaller group. Only 48 participants with elevated initial office readings were included. The smartwatch achieved a sensitivity of 78.4 percent, a specificity of 90.9 percent and a PPV of 96.7 percent. Multi-day home monitoring in that subset gave a sensitivity of 89.2 percent and a specificity of 81.8 percent. So the watch was more specific than home monitoring but less sensitive. The paper describes the watch as competitive with home monitoring rather than superior to it. It also reports that the specificity of the initial screen equals the office reading at 90.0 percent, so the improvement lies in sensitivity rather than in false alarms.

The authors say the model is meant to prompt at-risk people toward clinical assessment, not to diagnose them. Google's description of the feature in the Google Health app is less specific. The paper does not say how the blood pressure trends feature uses the model or what users would be shown. The interpretability analysis offers some physiological reasoning. Higher-risk waveforms had a narrower pulse width, a slower systolic upstroke and a lower d/a ratio, a measure of the reflected pulse wave relative to the forward wave. The accelerometer showed deeper K waves and wider IJK complexes. The authors read these as possible signs of stiffer arteries, but the figures show correlations between model scores and waveform shapes, not a demonstrated causal mechanism.

What changed

Type: Research · Applies to: Google Health app

Why it matters

AI-based health signals from wearables could change how health product teams design features, but clinical validation details are not given here.

Original source

Further reporting

Where we saw it

1 source

Matthias from THE DECODER

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works