Can a smartwatch flag hypertension without cuff calibration? A 196-person study shows promise—and misses cases
A peer-reviewed Google-led study found that a week of Pixel Watch 3 pulse and motion data identified 65.8% of participants classified as hypertensive by 24-hour ambulatory monitoring while holding specificity at 90%. That could support opportunistic screening, but the system missed roughly one in three reference-positive cases, was tested only in an untreated US screening population and is not a diagnosis or a current product claim.
The short version
- What happened
- In 196 analysable adults at five US sites, the model's sensitivity was 65.8% and specificity 90.0% against 24-hour ambulatory blood-pressure monitoring; 26 of 76 reference-positive participants were missed.
- Who it affects
- A passive alert could reach people who do not routinely check blood pressure, but a person still needs a validated cuff and clinical assessment to confirm the result.
- What remains unproven
- Only 196 of 296 enrolled participants entered the main analysis; missing or incomplete 24-hour ambulatory data accounted for most exclusions.
Keep in mind
Only 196 of 296 enrolled participants entered the main analysis; missing or incomplete 24-hour ambulatory data accounted for most exclusions.
The five-site US cohort excluded people taking antihypertensive medication and those with known cardiovascular disease, limiting generalisability to treated and higher-risk populations.
The operating point was selected to hold specificity at 90%, and positive predictive value depends on the study's unusually high hypertension prevalence.
The confirmatory comparison involved only 48 participants and produced wide confidence intervals, especially for specificity.
Alphabet funded the work; most authors are employees who may own stock, and the complete model is not available for independent reproduction.
No trial tested diagnosis rates, treatment, cardiovascular outcomes, anxiety, workload or cost after real-world alerts.
Multi-source analysis · 3 linked sources. Primary analysed source date: 5 November 2026. Earliest linked supporting source: 15 December 2025.
Read the full report · 10 min The evidence, methods, implications and limitations
The answer: useful screening evidence, not a cuff-free diagnosis
A consumer smartwatch may be able to flag some people who should receive proper blood-pressure assessment without first calibrating the watch against a cuff. In a peer-reviewed Nature study published on 5 November, a Google-led model analysed optical pulse signals, wrist motion and the wearer's height. Among 196 adults with usable comparison data, it detected 50 of the 76 participants who met the study's definition of hypertension on 24-hour ambulatory blood-pressure monitoring. That is 65.8% sensitivity, with a 95% confidence interval of 54.0% to 76.3%. Its specificity was fixed at 90.0%, meaning 108 of 120 participants below the reference threshold were classified negative.
Those numbers describe a screening signal, not a measurement of systolic or diastolic pressure. The system produced a person-level likelihood after several days of wear; it did not show a blood-pressure reading and did not eliminate the need for a validated arm cuff or clinical evaluation. Its positive predictive value was 80.6% in this deliberately enriched study population, where 38.8% of the analysed participants met the ambulatory definition. Predictive value would change in a population with a different prevalence, so 80.6% should not be presented as a universal chance that any alert is correct.
The clinically important limit is the missed cases. Twenty-six reference-positive participants were classified negative, so a quiet watch could not safely reassure a person that hypertension was absent. Seventeen of those 26 were also missed by the initial office reading, illustrating why hypertension can be difficult to find with a single encounter, but overlap between failures does not make the watch a replacement. Google also says this is early feasibility research and not an announcement of a capability in a current or forthcoming product. The result is consequential health research, but it is neither Breaking news nor permission for autonomous diagnosis.[1][2][3]
How the model and prospective test were built
The research had distinct development and validation stages. A self-supervised waveform encoder first learned 256-dimensional representations from 492,347 hours of unlabelled smartwatch data contributed by 2,109 consenting participants in earlier Fitbit studies. A logistic-regression classifier was then developed with data from 663 people in 44 US states who had both smartwatch recordings and 24-hour ambulatory monitoring. The classifier combined the learned pulse and accelerometer representation with height, which the researchers used as a rough proxy for arterial path length. The prospective cohort reported in the headline result was separate from those development data.
At five US sites, adults wore a Pixel Watch 3 for a free-living week and an FDA-cleared ambulatory cuff for the first 24 hours. They also received guideline-based office measurements at two visits and took paired home readings morning and evening on days two through seven. The ambulatory monitor was programmed every 20 minutes during the day and hourly overnight. The analysis accepted at least eight valid daytime and four valid night-time readings and defined hypertension as a time-weighted 24-hour average of at least 130/80 mmHg. Participants wore the watch for an average 7.3 days and 14.7 hours per day.
The validation group had a mean age of 41.9 years; 55.6% were female, 34.2% were non-Hispanic Black and 16.8% were Hispanic. Researchers measured wrist skin tone with a colorimeter rather than relying only on self-identification: 27.6% were grouped as light, 37.2% medium and 35.2% dark. Participants could not have known cardiovascular disease or be taking antihypertensive medication. More than two-thirds had either a previous elevated office reading without a hypertension diagnosis or an untreated diagnosis, so this was not an unselected sample of everyone who owns a watch.[1][2]
The distinction that matters
The study separates learning, development and prospective validation
The denominators answer different questions and should not be merged into one apparent sample.
2,109 participants
Contributed 492,347 hours of unlabelled watch signals
Self-supervised waveform representation
663 participants
Provided watch data plus 24-hour ambulatory reference labels
Classifier development and threshold selection
296 enrolled
Entered the five-site prospective study
196 had complete office and ambulatory data for the main analysis
196 analysed
Wore a Pixel Watch 3 for about a week
65.8% sensitivity and 90.0% specificity against 24-hour monitoring
What the comparisons do—and do not—show
At the same 90% specificity, the watch model's 65.8% sensitivity exceeded the 55.3% sensitivity of the first office screen in this cohort. Its area under the receiver-operating-characteristic curve was 0.856, compared with 0.865 for office blood pressure, but those summary curves do not mean the two methods are interchangeable. Office thresholds were based on cuff readings, while the model threshold was selected during development to target 90% specificity. The relevant question is whether adding passive screening helps find people who would otherwise remain untested, not whether a watch can be declared a winner over established measurement.
A smaller confirmatory analysis involved 48 people whose first office screen was elevated. In that subset, watch sensitivity was 78.4% and specificity 90.9%; three or more days of home readings reached 89.2% sensitivity and 81.8% specificity. The intervals were wide: watch specificity ranged from 58.7% to 99.8%, reflecting only eleven reference-negative people in the subset. The study therefore cannot establish that the watch is superior to home monitoring. It shows that the signal remained informative in a small, selected group that already screened positive.
An exploratory UK Biobank analysis asked a different question. A modified pulse-only encoder, needed because Biobank did not have matching accelerometer data, was trained or tuned on 58,117 records and evaluated across 76,268 people from different collection sites. Its classifications were associated with later hypertension diagnoses and several cardiovascular outcomes. That is useful evidence that the learned pulse features contain risk information, but it is observational association from a different model and finger-based signal. It does not validate the wrist model's diagnostic accuracy or show that alerts prevent strokes or deaths.[1][2]
Missing data and equity questions change the interpretation
The headline cohort is 196 people, but 296 enrolled. One hundred did not enter the main analysis, including 75 with missing or incomplete ambulatory data, often because the monitor was not worn during sleep or a file was empty or overwritten. The authors found no statistically significant differences in age, sex, body-mass index or office pressure between the full and analysed cohorts. That is reassuring on measured characteristics, but it cannot rule out bias from unmeasured differences in adherence, comfort, digital familiarity or health status. A real screening programme would face the same practical failures rather than analysing only complete comparisons.
The researchers reported no statistical evidence that classification errors were associated with age, sex, body-mass index, blood pressure or objectively measured skin tone in 196 observations. That is not proof of equal performance. The supplementary results show that 76.8% of recorded pulse-signal segments passed quality filtering overall, compared with 82.4% in the light, 77.2% in the medium and 71.9% in the dark skin-tone groups. A larger study should report sensitivity, specificity, false-alert burden and data availability separately for relevant groups, with enough participants to detect meaningful differences.
Company involvement also needs to stay visible. Alphabet funded the study; most listed authors are Alphabet employees and may own stock. Peer review raises the evidentiary status above the December 2025 medRxiv preprint, but it does not remove commercial interest or substitute for replication. The paper makes participant-level clinical-study data available in limited form and provides pseudocode rather than a deployable model. Independent teams therefore cannot yet reproduce the complete system from the publication alone.[1][2]
What would make this useful for patients rather than merely accurate on paper
The plausible benefit is opportunistic reach. A watch already collecting pulse and motion data could prompt someone with no symptoms to seek a validated check, particularly when one office reading is normal or healthcare contact is infrequent. But the intervention must be evaluated as a pathway: who receives an alert, how quickly confirmatory monitoring happens, whether people understand that silence is not reassurance, how many false alerts generate appointments, and whether diagnosis and treatment improve without widening access gaps. None of those patient outcomes was tested here.
A stronger next study would prospectively lock the model and threshold, recruit a much larger international population and include older adults, treated patients, people with arrhythmias and established cardiovascular disease. It should compare an alert-enabled group with usual care, measure confirmed new diagnoses and treatment, and follow harms such as anxiety, inappropriate self-treatment and unnecessary investigations. Device-wear failures and subgroup performance should be reported from everyone enrolled, not only those with complete reference data. Independent replication on other watch hardware would show whether the finding is a general physiological signal or a product-specific pipeline.
Until then, the practical reading is narrow. This peer-reviewed study supports further testing of passive smartwatch screening without cuff calibration. It does not support replacing ambulatory, home or office cuffs; it does not show improved health outcomes; and it does not establish a market-ready feature. Anyone concerned about blood pressure should use validated measurement and clinical advice, regardless of whether a wearable produces an alert. The assessment would change if a preregistered, independently replicated trial showed that alerts lead to more confirmed diagnoses and better outcomes with acceptable false-alert and equity profiles.[1][2][3]
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Continue the story
Read next
Continue exploring hypertension: compare the finding and its evidence limits.
Can a wearable AI warning predict a hypertensive crisis three days early?
A preliminary one-hospital study in Pakistan followed 4,190 adults with stage 2 hypertension and reported alerts up to 72 hours before a crisis. About 32% of positive alerts were followed by a crisis, but the model protocol is incomplete and treatment comparisons were not randomised.
Read the next reportExplore the full source and evidence record
What this means for people
- A passive alert could reach people who do not routinely check blood pressure, but a person still needs a validated cuff and clinical assessment to confirm the result.
- Because the model missed 26 of 76 reference-positive participants, absence of an alert cannot safely rule out hypertension.
- Lower retained pulse-signal quality in the darker skin-tone group makes larger subgroup studies important before equitable deployment can be assumed.
Global context
The prospective clinical validation took place at five US sites. The separate UK Biobank analysis used a different pulse-only model and finger signal, so it cannot substitute for validation of the watch system outside the United States. Blood-pressure thresholds, access to ambulatory monitoring, device ownership and clinical follow-up differ across countries; international studies should test the full screening pathway rather than importing the US estimates unchanged.
What the evidence does not yet show
- Only 196 of 296 enrolled participants entered the main analysis; missing or incomplete 24-hour ambulatory data accounted for most exclusions.
- The five-site US cohort excluded people taking antihypertensive medication and those with known cardiovascular disease, limiting generalisability to treated and higher-risk populations.
- The operating point was selected to hold specificity at 90%, and positive predictive value depends on the study's unusually high hypertension prevalence.
- The confirmatory comparison involved only 48 participants and produced wide confidence intervals, especially for specificity.
- Alphabet funded the work; most authors are employees who may own stock, and the complete model is not available for independent reproduction.
- No trial tested diagnosis rates, treatment, cardiovascular outcomes, anxiety, workload or cost after real-world alerts.
What to watch next
- Independent, preregistered prospective validation with a locked model and threshold.
- Performance and usable-data rates by skin tone, age, sex, health status and device-wear pattern with adequately powered denominators.
- Trials in treated patients, older adults and people with cardiovascular disease or arrhythmias.
- Randomised evidence that alerts produce more confirmed diagnoses or better outcomes without excessive anxiety and workload.
- Regulatory review, product-specific labelling and a clear confirmatory-care pathway before any consumer release.
Living evidence record
Impact record IAI-1WDJY98
Evidence stage
Studied
Confidence
Supported
Reporting basis
Multi-source analysis
Independent or research support
Not yet
Record status
Monitoring
Last checked
5 November 2026
Source trail
3 direct sources across 2 source types.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Evidence trail
Sources used for this report
Links checked 5 November 2026
Research question, topics and regions
Smartwatch screening for hypertension
Topics: Hypertension · Wearables · Screening · Clinical AI · Health equity
Regions: United States · United Kingdom
This report is labelled multi-source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.