Newshot Can Google's smartwatch screen for hypertension without a cuff?
Google Research uses ordinary smartwatches to collect multi-day optical pulse wave and motion signals, then uses AI to identify a persistent hypertensive state; the system requires no cuff calibration and is validated against 24-hour ambulatory blood pressure monitoring as the clinical gold standard, but it targets the screening entry point, not a replacement for diagnosis.
- Hypertension is often asymptomatic, and existing cuffless devices mostly require cuff calibration
- The system aggregates multi-day PPG and ACC signals to identify a persistent state
- The model learns from and classifies 492,347 hours of data
- Accuracy is validated against 24-hour ambulatory blood pressure monitoring
From the official Google Research research blog, authored by Google Research researchers; the results were also used for the Health Guardian blood pressure trends feature in the Google Health app.
Hypertension affects more than 1 billion people worldwide, and more than half of patients do not know they have it. Google Research's new study has ordinary smartwatches continuously read multi-day pulse waves and wrist vibrations, with AI determining whether a persistent hypertensive state exists, validated against 24-hour ambulatory blood pressure monitoring, with accuracy benchmarked against the traditional cuff workflow. It aims to miss fewer at-risk cases, not to diagnose in place of a doctor.
Hypertension doesn't always raise an alarm; screening has to reach out proactively
More than half of adults with the condition do not know their status. Hypertension usually has no obvious symptoms and is often discovered only during an incidental measurement. With the body lacking a "sounding alarm" reminder, regular screening becomes the key entry point for finding the problem.
Underdiagnosis is especially severe among people aged 18 to 44. They may already have hypertension but rarely seek care proactively because of symptoms. To give people who have no obvious discomfort and cannot easily visit a hospital frequently a chance to be screened, both the place and the method of measurement have to change.
Traditional arm cuffs can provide a blood pressure reading, but they require a person to actively schedule a measurement. Although existing commercial cuffless devices can work continuously, they usually still require regular calibration with a traditional arm cuff, first obtaining a cuff reading as a baseline.
This brings a direct limitation: such devices are better suited to tracking changes relative to a person's baseline than to independently identifying an absolute hypertensive state. Without calibration, they cannot judge from their own readings alone that "this person is currently in a hypertensive state"; wearing them longer does not close this gap.
The opportunity for cuffless screening lies in reaching people who would not measure proactively, moving "is this worth further confirmation" forward from the hospital door. The key is not just whether a watch can produce a cuff-style number, but whether it can independently determine a persistent hypertensive state.
The watch doesn't report a blood pressure value; it looks for clues of persistent hypertension
Google's watch system reads two types of signals: light and mechanical vibration at the wrist. It does not measure a single systolic or diastolic pressure; it determines whether a person is persistently in a hypertensive state.
The two signal paths do not work separately but complement each other: one looks at blood flow, the other at when the body quiets down. The first comes from PPG (a sensor that tracks changes in blood volume using reflected light). Light shines on the wrist skin, and the system tracks blood volume fluctuations based on changes in reflected light while also capturing the pulse wave contour—the shape left behind each time blood surges against the vessels is all in there.
The second comes from ACC (a sensor that perceives wrist motion and vibration). It first picks out periods when the watch is still, then captures the faint vibration transmitted to the wrist by cardiac ejection, i.e., the ballistocardiogram signal. PPG describes blood flow and the pulse wave, while ACC marks the body states suitable for analysis—only combined are the two paths sufficient for a judgment.
These data are obtained passively by the watch during daily wear. The system aggregates across multiple days rather than directly predicting a blood pressure number at a given moment. Instantaneous readings are easily disturbed by the environment and body state, whereas multi-day signals are better suited to judging: do this person's physiological features persistently point to hypertension?
What ultimately participates in the judgment are the user's aggregated physiological representation and height. Height provides body-size context, and the aggregated representation summarizes the signal features collected over multiple days; together they are used to identify a hypertensive state. The system is named AI-PPG-ACC-HTN (a blood pressure identification system combining the two types of watch signals), which breaks down into optical signal, acceleration signal, plus hypertension.
How does it recognize hypertension from waveforms?
The system first learns stable waveform features from a large amount of watch signals, then performs classification; it does not first memorize the answer "hypertension."
The research team used 492,347 hours of sensor data from 2,109 consenting participants. That is roughly equivalent to one person's sensor records continuing for 21 years, except they are spread across 2,109 people, covering multi-day, multi-person daily activity.
The training phase does not touch blood pressure labels. The team used a one-dimensional deep convolutional neural network (a neural network that extracts features from continuous waveforms) and self-supervised contrastive learning (bringing similar signals closer in representation space) to map 15-second PPG and ACC waveforms from the same participant with overlapping time to nearby positions. PPG (observing pulse waves with light) and ACC (recording wrist motion with an accelerometer) first learn to correspond to each other; the answer only appears at the classification stage.
A single 15-second signal is not directly judged as high or low blood pressure. The system averages 15-second feature blocks across multiple days to reduce short-term noise such as sudden movement and temperature changes, leaving information closer to a persistent state.
The aggregated representation, together with height, is fed into a regularized logistic regression classifier (a classification model that uses height to help compute a continuous probability), producing a continuous hypertension probability score. What is output is a probability, not a specific systolic or diastolic value.
Without a cuff, is the accuracy sufficient?
This system directly uses ambulatory blood pressure monitoring as the comparison. The official conclusion is that its accuracy in both initial screening and confirmatory screening is comparable to the traditional cuff diagnostic workflow.
This validation used a prospective, multicenter clinical study. The study first determined the validation protocol, then observed subjects and collected results; it was not a post-hoc test using existing data, and the validation was also completed jointly by multiple clinical centers. For a screening tool, these two points bear on whether the conclusions can extend beyond a single institution or a single batch of subjects.
The comparison standard is 24-hour ambulatory blood pressure monitoring, ABPM (continuously worn device that measures blood pressure at scheduled times), which Google calls the clinical gold standard. Its task differs from a single office measurement: ambulatory monitoring takes readings repeatedly at different times of day to determine a person's blood pressure state rather than leaving only a single reading. The validation of the watch system therefore has a clear clinical reference.
The official statement of accuracy covers two steps: initial screening identifies candidates for further evaluation among people not yet diagnosed; confirmatory screening is used for re-identification after an existing positive finding. Both are comparable to the traditional cuff diagnostic workflow, indicating that these two key steps, screening and recheck, both reached the benchmarking level set officially.
This evidence comes from the project's official research materials, and no independent third-party retest of the same validation results has been seen so far. "Comparable accuracy" can serve as the basis for this system having passed clinical validation, but it cannot yet be understood as having been repeatedly confirmed across all brands, all populations, and all daily-use conditions. What can be confirmed is the validation method and the official conclusion; how far it can go outside the research setting still depends on retest evidence in real-world use.
It can't measure a single blood pressure accurately, but it can track long-term trends
The results of this study have already entered the product path: they helped shape the Health Guardian blood pressure trends feature in the Google Health app. "Trends" here means seeing how a personal state changes over time, not an absolute reading given by a single hospital examination.
Why does the product only do trends and not readings? The calibration limitation mentioned earlier is key—if existing cuffless solutions lack absolute blood pressure calibration, they can only track changes relative to a baseline, and the baseline usually still has to come from a traditional arm cuff. Google's system instead uses multi-day signals to identify a persistent state, bypassing per-measurement absolute measurement, at the cost that what it provides is still a probability, not a diagnosis.
Returning to that everyday scenario of someone wearing an ordinary smartwatch, not yet aware of abnormal blood pressure but worried about missing a diagnosis: what it can do is remind you to measure your blood pressure, not draw conclusions in place of a doctor.
One question remains to be verified: in different populations, with different smartwatches, and over longer periods of use, how will the actual performance of missed detections and false alarms in this screening turn out?
Source: Google Research Blog (November 5, 2026); the framing is an introduction to the official research blog, and this article explains its technical path and validation standards based on publicly available content; the complete sensitivity, specificity, and external validation data were not disclosed in the provided original text.
Glossary · plain-language notes appearing in this piece
Full texts of in-depth pieces, research interpretations, and tool tutorials, plus all podcasts, are in the members' area.