Skip to content
Cardiovascular healthValidation study/review9 min read

Apple Watch irregular-pulse alerts were rare; 34% of those tested had atrial fibrillation

MoveVitals Editorial Collective · Study published · Reading updated

Original paper:Large-Scale Assessment of a Smartwatch to Identify Atrial Fibrillation · The New England Journal of Medicine · 2019 · Original paper ↗

Key takeaway

Among 419,297 Apple Watch users with no known atrial fibrillation (AF), 0.52% received an irregular-pulse alert, and 34% of the 450 who returned a usable ECG patch had AF (97.5% CI 29 to 39). Alerts that coincided with patch recording matched AF with a positive predictive value of 0.84 (95% CI 0.76 to 0.92), but only 20.8% of alerted people returned a patch and nobody without an alert was monitored, so a quiet watch does not rule AF out.

Study design

Validation study/review

Reading basis

Full text

Reading focus

Reference & error

The research question

Atrial fibrillation (AF; the paper also counts atrial flutter) is the most commonly diagnosed clinically significant heart rhythm disorder and is linked to roughly five times the risk of stroke. It often comes and goes, so an ECG can look normal between episodes, and it can cause few or no symptoms, which delays diagnosis. The authors note that continuous heart monitors and implantable devices find more AF in high-risk groups but cover limited periods and need either an invasive procedure or action by the user. Smartwatches measure pulse with optical sensors, and algorithms exist that flag irregular pulses, but how well one works during ordinary consumer use was unknown. The Apple Heart Study set out to test how well an irregular-pulse notification algorithm in an Apple Watch app identifies AF, and whether a fully remote design with no study sites could do it at very large scale.

Who and how

The Apple Heart Study was a prospective, single-group, open-label, siteless, pragmatic study: there was no control group and no clinic visit to enrol. People in the United States with a compatible iPhone and Apple Watch, aged 22 years or older and proficient in English, could download the study app between November 29, 2017 and August 1, 2018. Anyone who reported previous AF or current oral anticoagulant use was not eligible, and eligibility was self-reported. In all, 419,297 participants joined over 8 months from 50 states and the District of Columbia; their mean age was 41 years (SD 13), 42% were female and 5.9% were 65 or older. The watch's photoplethysmography sensor, which uses light to track changes in blood flow, intermittently and passively sampled the pulse while the wearer was at rest and built 1-minute series of pulse intervals (tachograms) that the algorithm classified as regular or irregular. After a participant's first irregular-pulse notification, a telemedicine visit with a physician from American Well was offered, and unless symptoms were urgent an ECG patch (ePatch) was mailed to be worn for up to 7 days. Two clinicians read each patch report, and 3-minute ECG strips time-aligned to the sampled tachograms were read separately, with a third clinician and a committee settling disagreements. Apple sponsored the study and owns the data; Stanford scientists ran the analyses independently of the sponsor.

There were two coprimary outcomes: AF lasting more than 30 seconds on the patch in a notified participant, and AF on the patch at the same moments as an irregular tachogram. Key secondary outcomes were AF on the patch at the time the notification algorithm itself detected an irregular pulse (the positive predictive value of a notification) and self-reported contact with a health care provider within 3 months. The median monitoring time was 117 days (interquartile range 113 to 186), with surveys 90 days after a notification and at the end of the study. The analysis was descriptive. The authors targeted 503 analysable patches in each of two age groups (under 65 and 65 or older) so that confidence intervals for the AF yield in people 65 or older and for the tachogram positive predictive value would be no wider than 0.10, and they gave 97.5% confidence intervals for those two quantities and 95% for the rest. The patch analysis counted notified people with no reported AF history or anticoagulant use and no urgent symptoms at the first visit, who wore the patch for at least 1 hour within 14 days of shipment and returned it within 45 days.

What it found

  1. 1
    Alerts were uncommon overall but became more frequent with age: 2161 of 419,297 participants (0.52%) received an irregular-pulse notification, ranging from 0.16% of those aged 22 to 40 to 3.1% of those 65 or older, and half of the alerts came by day 38 after enrolment and 90% by day 133.
  2. 2
    A little over a third of alerted people who wore a patch had AF on it: AF was found in 153 of 450 participants, a yield of 34% (97.5% CI 29 to 39), with 35% (97.5% CI 27 to 43) in those 65 or older and 18% (95% CI 6 to 31) in those younger than 40, and the patches went on a mean of 13 days (SD 16) after the alert and were worn for an average of 6.3 days.
  3. 3
    When a patch happened to be recording during an alert, the alert was usually confirmed: 72 of the 86 participants in that situation had AF on the concurrent strips, a positive predictive value of 0.84 (95% CI 0.76 to 0.92), or 0.78 (95% CI 0.64 to 0.92) in people 65 or older.
  4. 4
    A single irregular reading was less reliable than the notification, which the authors link to notifications requiring confirmation from multiple irregular tachograms: 1489 of 2089 sampled irregular tachograms from notified participants coincided with AF (positive predictive value 0.71, 97.5% CI 0.69 to 0.74), falling to 0.60 (97.5% CI 0.56 to 0.64) in people 65 or older. Of the 600 irregular tachograms without AF, frequent premature atrial contractions were present in 77%, frequent premature ventricular contractions in 16% and atrial tachycardias in 38% (not mutually exclusive), and sinus arrhythmia alone in 28 (4.7%).
  5. 5
    The AF that patches found was often substantial: of the 153 participants with AF, 20% had continuous AF, most of the rest had AF less than 50% of the time they were monitored, and 89% had an episode lasting at least 1 hour. A total of 20 participants were contacted urgently, 18 for AF with ventricular rates above 200 beats per minute for more than 30 seconds, 1 for a pause longer than 6 seconds and 1 for nonsustained ventricular tachycardia lasting more than 6 seconds.
  6. 6
    Alerts led many people to seek care, and the app caused little serious harm: of 1376 notified participants (64%) who returned the 90-day survey, 787 (57%) contacted a provider outside the study, 28% were prescribed a new medication, 33% were recommended to see a specialist and 36% were recommended additional testing. No serious app-related adverse events were reported, and 16 (1.5%) of 1038 reviewed adverse events were app-related, 15 of them anxiety, none leading to hospitalization or urgent medical attention.

Limits & uncertainty

  1. 1
    The authors note that only 450 of 2161 notified participants (20.8%) returned an analysable patch, fewer than anticipated, so the planned precision for the AF yield (503 patches per age group) was not met and the result may not carry over to notified people who did not return one, although they saw no qualitative differences between the two groups. Patches were also applied a mean of 13 days after the alert, so an AF episode may already have ended.
  2. 2
    The authors note that the study was not designed to evaluate the watch as a screening tool or to measure sensitivity, specificity or false positives. Nobody without an alert wore a patch, the algorithm was built to minimise false alarms and not to detect short episodes, and the positive predictive values describe people who had already been notified, so an absent alert does not exclude AF and an alert is not a diagnosis.
  3. 3
    The authors note that eligibility and outcomes were self-reported with no in-person contact, that several notified participants later reported earlier AF despite the screening, that loss to follow-up was substantial, and that the cohort skewed young (mean age 41 years), as smartwatch owners do.
  4. 4
    The authors add that uncertainty remains about the benefits of diagnosing and treating asymptomatic AF, particularly when episodes last 6 hours or less. Being single-group, the study cannot show whether alerts prevented strokes or improved any outcome.
  5. 5
    Our observations: Apple funded the study and owns the data, although Stanford scientists did the analyses, and the results describe the algorithm and watch of 2017 and 2018, which the authors say is a fast-moving technology. The end-of-study survey was completed by 929 of 2161 notified participants (43%) but 293,015 of 417,136 never-notified participants (70%), so its comparison of new AF diagnoses (404 of 929 versus 3070, or 1.0%) rests on self-report and unequal response.

What to take from it

An irregular-pulse notification from a watch like this is a good reason to get checked, not a diagnosis. In the study, a little over a third of alerted people who wore a patch had AF on it, an alert that coincided with a recording matched AF with a positive predictive value of 0.84, and the authors stress that a normal later recording does not mean the alert was wrong because AF can come and go. More than half of survey respondents (57%) went on to contact a provider outside the study, which is the sensible next step.

The study does not show that a quiet watch means a healthy rhythm, that wearing a watch prevents strokes, or that finding AF through an alert improves outcomes. Decisions about starting or stopping medicines such as anticoagulants belong with your clinician. Confidence is moderate for the alert performance figures in this largely younger, self-selected group of US users and low for any claim of health benefit.

Reading your own data

The useful signal here is the watch's own irregular-rhythm notification, not daily heart-rate or resting heart-rate averages, which cannot show AF. If an alert appears, note the date, time, what you were doing and any symptoms, and bring that record to a clinician who can arrange an ECG; because episodes are intermittent, a recording soon after the alert is likelier to catch AF than one weeks later (in the study, patches went on a mean of 13 days after the alert). The sensor sampled the pulse only intermittently at rest, so wearing the watch regularly gives it more chances to sample, and these results apply to this algorithm on the Apple Watch, not to other brands or software versions.

How to read this evidence

Agreement is specific to the device, firmware, reference, population and setting. Correlation is not agreement, and validating a metric is not proof of diagnostic or treatment benefit.

Sources & attribution

Large-Scale Assessment of a Smartwatch to Identify Atrial Fibrillation

The New England Journal of Medicine · 2019-11-14

DOI: 10.1056/NEJMoa1901183 ↗PubMed · 31722151 ↗

MoveVitals editorial interpretation · educational content · no individual clinical review recorded

Sources checked 2026-11-01

Practical guides

Read in context

·Observational meta-analysis

Disrupted daily rest-activity rhythm is linked to higher mortality in 109,365 people

Across seven observational studies with 109,365 participants, a weaker day-night activity contrast (lower relative amplitude) was associated with higher cardiovascular mortality (hazard ratio [HR] 1.50, 95% CI 1.13 to 1.98) and all-cause mortality (HR 1.46, 95% CI 1.27 to 1.68). The evidence base is small and observational, so it cannot show that improving the rhythm lowers risk, and illness may itself disturb the rhythm.

Journal of the American Heart Association

In-depth explainersRead study →
·Trial meta-analysis

Five exercise types lowered resting blood pressure, isometric training the most

Pooling 270 randomised trials with 15,827 participants, aerobic, resistance, combined, interval and isometric training all lowered resting blood pressure compared with non-intervention controls (P<0.001 for each), with the largest fall for isometric training: 8.24 mm Hg systolic and 4.00 mm Hg diastolic. The ranking between modes is partly indirect and the trials differ, so it is not a personal prescription.

British Journal of Sports Medicine

In-depth explainersRead study →
·Randomised trial

A salt substitute lowered stroke, major cardiovascular events and death in rural China

In 20,995 people in 600 rural Chinese villages who had a history of stroke or were 60 or older with high blood pressure, a salt substitute (75% sodium chloride, 25% potassium chloride) lowered the stroke rate versus regular salt over a mean 4.74 years: 29.14 versus 33.65 events per 1000 person-years (rate ratio 0.86, 95% CI 0.77 to 0.96; P = 0.006). Death was also lower (rate ratio 0.88) and serious hyperkalemia events were not significantly higher, but the trial was open-label and set in rural China.

The New England Journal of Medicine

In-depth explainersRead study →
·Randomised trial

Evolocumab added to statins lowered LDL by 59% and reduced cardiovascular events

In 27,564 people with atherosclerotic cardiovascular disease already on a statin, adding evolocumab lowered LDL cholesterol by 59% versus placebo at 48 weeks and reduced the primary composite of cardiovascular events to 9.8% from 11.3% over a median 2.2 years (HR 0.85, 95% CI 0.79 to 0.92). Amgen funded the trial, follow-up was short, and only people with established disease took part, so the result does not speak to healthy people or to long-term safety.

The New England Journal of Medicine

In-depth explainersRead study →
Cardiovascular health →