Your Watch Thinks You Have AFib. Read the Fine Print.

The full story

The notification is a small thing: a tap on the wrist, a card on the screen, the phrase "irregular rhythm." It is also one of the most carefully hedged consumer features ever shipped. The FDA's 2018 de novo order creating the category, named "photoplethysmograph analysis software for over-the-counter use," states plainly that the feature is not intended to diagnose atrial fibrillation or replace traditional diagnostic methods, is not for people under 22 or already diagnosed with AF, and cannot detect all instances of AF. It adds that all notifications should be reviewed by a medical professional.

None of that is on the watch face; what is on the watch face reads like a result. So it is worth knowing what the very large studies actually measured, because the distance between finding atrial fibrillation and making someone better off is the entire subject.

What the big studies measured

The Apple Heart Study, published in the New England Journal of Medicine in 2019 as a pragmatic trial, enrolled 419,297 participants and followed them a median of 117 days. Only 0.52 percent, 2,161 people, ever received an irregular-pulse notification. Of the 450 who returned an analysable ECG patch afterwards, atrial fibrillation was present in 34 percent (97.5% CI 29 to 39), rising to 35 percent among those 65 or older. Fifty-seven percent of notified participants who returned a survey contacted a clinician outside the study.

The Fitbit Heart Study, in Circulation in 2022, ran the same shape of experiment on 455,699 users with a median age of 47 over a median 122 days. One percent received an irregular heart rhythm detection. Of the 1,057 with an analysable subsequent ECG patch, AF was present in 32.2 percent. It also produced the figure that gets quoted out of context constantly: the positive predictive value of a repeat detection during patch monitoring was 98.2 percent (95.5 to 99.5). Read the qualifier. That is the algorithm agreeing with itself while a medical-grade patch is already recording, not the odds your first notification means disease.

The honest summary: roughly a one-in-three chance of atrial fibrillation among people motivated enough to wear a patch afterwards, in cohorts that skewed young. Two-thirds of those patches showed no AF.

A notification is not a diagnosis, and not even specifically AF

A 2021 Apple Heart Study analysis in Circulation: Arrhythmia and Electrophysiology reported that arrhythmias other than atrial fibrillation also turned up on subsequent patch monitoring in participants whose watch had flagged an irregular pulse. That matters: the sensor detects irregularity in the interval between pulses, and irregularity has many causes. The word on the screen is a hypothesis about a category broader than the one the app names.

Where the sensors do look strong is in people already suspected of the disease. A 2021 systematic analysis in the Journal of Atrial Fibrillation pooled nine prospective studies and 1,559 patients with a mean age of 63.5, of whom 39.5 percent had a history of AF, and found smartwatch single-lead ECG and photoplethysmography detected AF with 95 percent sensitivity and 94 percent specificity, non-inferior to standard monitoring strategies (odds ratio 1.06, 95% CI 0.93 to 1.21). In the Huawei Heart Study, reported in JACC: Asia in 2021, a machine-learning model on photoplethysmography reached 81.9 percent sensitivity, 96.6 percent specificity and 96.4 percent positive predictive value against 72-hour Holter monitoring in 50 patients with known paroxysmal AF. Excellent numbers, in enriched cohorts where a large fraction already carried the diagnosis. Move the same algorithm into a population where AF is rare and the arithmetic of predictive value turns against you.

What the evidence does not show: that finding more AF prevents strokes

This is the part the marketing skips, and it is not a subtle gap: it is three randomised trials deep.

mSToPS, a randomised clinical trial published in JAMA in 2018, gave 2,659 high-risk adults with a mean age of 72 either immediate or delayed home ECG patch monitoring. Immediate monitoring diagnosed new AF in 3.9 percent at four months against 0.9 percent, an absolute difference of 3.0 percent. At one year, monitored participants had more AF diagnoses than matched unmonitored controls (6.7 versus 2.6 per 100 person-years), more anticoagulant starts (5.7 versus 3.7) and more cardiology visits (33.5 versus 26.0). There was no difference in AF-related emergency visits or hospitalisations.

The LOOP study, in the Lancet in 2021, went further with better hardware: 6,004 adults aged 70 to 90 with at least one stroke risk factor, randomised to continuous implantable loop recorder screening or usual care, median follow-up 64.5 months. Screening tripled AF detection (31.8 versus 12.2 percent, hazard ratio 3.17) and more than doubled anticoagulation starts (29.7 versus 13.1 percent). Stroke or systemic embolism: 4.5 versus 5.6 percent, hazard ratio 0.80 with a confidence interval of 0.61 to 1.05. Not significant. The authors' conclusion is unusually direct: not all AF is worth screening for, and not all screen-detected AF merits anticoagulation.

STROKESTOP, also in the Lancet in 2021, is the positive trial, and its effect size sets the scale. Among 28,768 people aged 75 to 76 randomised to an invitation to intermittent ECG screening or to control, after a median 6.9 years the screening arm showed a small but significant reduction in a composite of stroke, systemic embolism, bleeding-related hospitalisation and death: 31.9 versus 33.0 percent, hazard ratio 0.96 (0.92 to 1.00), p = 0.045. Only 51.3 percent of those invited took part.

Hence the USPSTF's 2022 recommendation statement in JAMA: for asymptomatic adults 50 or older without prior stroke or transient ischaemic attack, the evidence is insufficient to assess the balance of benefits and harms of screening. An I statement, neither for nor against. The accompanying evidence report, spanning 26 studies and 113,784 people, found screening detects substantially more AF than no screening (risk difference 1.0 to 4.8 percent across four randomised trials totalling 32,491 people), and that anticoagulation in clinically detected AF reduced ischaemic stroke (pooled relative risk 0.32) and all-cause mortality (0.68) while increasing major bleeding. And the sentence that belongs on every marketing deck: no trial assessed the benefits and harms of anticoagulation specifically in screen-detected populations. The treatment evidence comes from people who arrived with symptoms, and it is being extrapolated to people who arrived with a notification.

What to buy, and what to do with a tap on the wrist

• The spec that matters is a cleared irregular rhythm notification feature plus on-demand single-lead ECG. The notification is a prompt; the tracing is the artefact a clinician can read.
• Yield rises with age: in the Apple Heart Study, 35 percent of notified participants 65 and older had AF on a patch, against 34 percent overall in a much younger cohort.
• Expect almost nothing to happen: notification rates were 0.52 percent and 1 percent in the two big studies.
• Read the labelling limits: under 22, or already diagnosed with AF, and the feature is outside its intended use.

If a notification arrives, note what every meaningful number here has in common: it came from putting a recording device on somebody. The FDA's order says notifications should be reviewed by a medical professional, which is also the practical route from an ambiguous flag to an interpretable rhythm strip.

What makes this technology interesting is also what makes it uncomfortable: it has decoupled detection from benefit. We can now find atrial fibrillation at a scale no health system has managed, in people who feel entirely well, years before they would have presented, and the trials say we do not yet know whether that helps them. Not a reason to switch the feature off. A reason to hold two ideas at once: the sensor on your wrist is a real instrument, and the question it answers is not quite the question you were asking.