The Apple Heart Study enrolled 419,297 people in eight months. It was run almost entirely through an app, with follow-up by telehealth visits and ECG patches sent in the post. Of all those people, 0.52% received a notification that their watch had detected an irregular pulse. When the investigators mailed ECG patches to those notified and got them back, atrial fibrillation showed up in 34% of cases. The results were published in the New England Journal of Medicine in 2019.
Read one way, that is a triumph: a consumer device finding a dangerous arrhythmia at population scale. Read another way, it means that about two in three people who got the alert did not have atrial fibrillation on the follow-up patch. Some may have had brief episodes the patch missed. Many simply got a worrying message and a trip into the health system.
The usual debate asks whether wearable data is signal or noise. That is the wrong frame. The data is often decent. The problem is that consumer wearables have quietly created a mass screening programme without any of the things a screening programme is supposed to have.
What a screening programme normally requires
When a health system decides to screen a population for a condition, it works through a well-established checklist. Is the condition important? Is there an accurate test? Is there an effective treatment that works better when started early? Is there a clear pathway for people who test positive? Do the benefits outweigh the harms of false positives and overdiagnosis?
Wearables skip that checklist. The “screening” population is whoever bought the device, which skews toward younger, healthier and wealthier people with lower prior risk of the conditions being flagged. The threshold for an alert is set by a company, cleared by a regulator for a specific claim, and then deployed to millions. And the pathway for a positive result is, in practice, “talk to your doctor.”
In the United States, the Preventive Services Task Force concluded in 2022 that there was insufficient evidence to recommend for or against screening for atrial fibrillation in adults aged 50 or over without symptoms. The same condition is being screened for continuously on millions of wrists anyway.
The costs land somewhere specific
Those costs are not abstract. A 2024 study in the Journal of the American Heart Association, led by Lindsey Rosman at the University of North Carolina, followed 172 patients who already had atrial fibrillation. The 83 who used a wearable reported more symptom monitoring and preoccupation and more worry about their treatment. About one in five wearable users said they felt anxious and always contacted their doctor after an irregular-rhythm notification. Wearable users also had higher rates of AF-related health care use, including ECGs, echocardiograms and ablation.
This is a small, observational study, and it cannot show that the devices caused the difference. People who choose to wear monitors may already be more anxious about their health. But it matches what many clinicians describe: inboxes filled with screenshots, and consultations spent explaining what a watch alert does and does not mean.
That work is mostly unpaid, and it lands on primary care and cardiology clinics that already have waiting lists. There is no billing code for “reviewed a patient’s watch tracing and reassured them.” The cost of the unplanned screening programme is paid in clinician time and patient anxiety, while the benefit is booked as product value by the device maker.
Where the signal is genuinely useful
None of this means clinicians should wave wearable data away. There are settings where it is clearly helpful:
- Known diagnosis, specific question. A patient with diagnosed paroxysmal AF using a watch ECG to capture an episode that correlates with symptoms can save weeks of ambulatory monitoring.
- Trends over snapshots. Resting heart rate, step counts and sleep patterns over months can tell a story about recovery, deconditioning or medication effects that a single clinic visit cannot.
- Clinician-prescribed monitoring. When a clinician chooses the device, the metric and the response in advance, the data has a destination.
The pattern is the same in each case. The data is useful when someone asked a question before it was collected.
What would make this work
Health systems and regulators could take a few steps to turn accidental screening into something more deliberate:
- Build a triage pathway. Decide who reviews consumer-device data, what counts as actionable and what gets a standard reassurance message.
- Separate patient-initiated from clinician-prescribed data. They carry different expectations, and mixing them in the same inbox guarantees overload.
- Ask for population-level performance data. Regulators and health systems should know how alerts perform in the actual user population, not just in validation studies.
- Pay for the work, or stop pretending it’s free. If reviewing wearable data is part of care, it needs time in the schedule.
The mechanism: prior probability does the damage
The core issue is a principle every clinician learns and most people outside medicine never do. The meaning of a positive test depends heavily on how likely the condition was before the test.
A wearable alert in a 75-year-old with hypertension and palpitations is a very different piece of information from the same alert in a healthy 30-year-old. The device is the same. The algorithm is the same. But in the older patient, atrial fibrillation was reasonably likely to begin with, so an alert substantially raises the probability. In the young patient, it was very unlikely, so even a fairly accurate alert is more often wrong than right.
Consumer devices are marketed and sold without regard to prior probability. They are worn disproportionately by people for whom many serious conditions are rare. That is not a flaw in the device’s engineering. It is a structural feature of consumer distribution, and it guarantees that a meaningful share of alerts will be false positives, however good the algorithm becomes.
There is a subtler problem too. Wearables are increasingly measuring things for which there is no agreed clinical threshold at all: heart rate variability, sleep stages, blood oxygen trends, “readiness” scores. When a patient brings one of these into a consultation, the clinician often has no evidence base to interpret it against. The data is not wrong, exactly. It is unanchored.
A worked scenario: one alert, two pathways
Imagine two people who receive the same irregular rhythm notification on the same evening.
The first is a 68-year-old man with high blood pressure who has noticed occasional palpitations. He mentions the alert to his GP, who arranges an ECG and ambulatory monitoring. Paroxysmal atrial fibrillation is confirmed. His stroke risk score indicates that anticoagulation is appropriate. The alert may well have prevented a stroke. This is the case the technology’s advocates describe, and it is real.
The second is a 34-year-old woman who runs regularly and has no symptoms. She is alarmed by the notification and books an urgent appointment. Her ECG is normal. She is offered ambulatory monitoring, which is also normal. Over the next month, she receives two more notifications and books another appointment. She is reassured each time but begins checking her watch many times a day. Eventually a cardiologist reviews her results and explains that the alerts were likely artefacts or benign rhythm variations.
Both stories are plausible outcomes of the same product feature. The first produces clear benefit. The second produces cost and anxiety with no benefit. The question for health systems is not which story is true. It is what the ratio between them looks like in the population actually wearing the devices, and nobody is systematically measuring that.
The strongest counterargument
Advocates of consumer wearables make a strong case. Atrial fibrillation is often silent until it causes a stroke. Many people with it are never diagnosed until too late. Wearables put continuous monitoring in the hands of millions at no cost to the health system, and some of those people will be found and treated early. Anxiety and extra appointments are real costs, but they are small compared with a disabling stroke. And people have a right to information about their own bodies, whether or not a screening committee has approved it.
This argument partly holds. Individual cases of wearables detecting serious arrhythmias are real, and for some users the benefit is life-changing. The autonomy point is also important. Paternalistic refusal to engage with patient data will not work and should not.
But the argument is about individuals, while the costs fall on systems. A health service does not experience one alert. It experiences all of them, and has to decide how much clinical time to spend on each. The fact that some alerts save lives does not tell us whether the overall balance is positive, or whether the clinician time would have done more good elsewhere. Those questions can only be answered with population data, which is why the lack of it matters so much.
Second-order effects to watch
Equity. If wearable users get faster access to cardiology because their devices generate referrals, people without devices, often older, poorer and at higher risk, could find themselves further back in the queue.
Defensive medicine. Clinicians who dismiss an alert that later turns out to be significant face obvious liability risk. That pressure pushes toward more testing, even when the prior probability is low.
Insurance and employment. As wearable data becomes more clinically accepted, questions about who can access it, and whether it can affect insurance or employment decisions, will become more pressing.
Drift toward diagnosis. Device makers have strong incentives to seek regulatory clearance for more conditions. Each new clearance adds another category of alerts to the clinical workload.
What this means for different players
Health system leaders should decide explicitly how patient-generated data enters the clinical record, who triages it and what the standard responses are. Leaving it to individual clinicians guarantees inconsistency and burnout.
Clinicians can use prior probability as a practical tool, explaining to patients why the same alert means different things for different people. A short, standard explanation can reduce repeat visits.
Regulators could require manufacturers to report real-world positive predictive value in user populations, not just accuracy in validation studies, and to publish it.
Ministries and payers should consider whether, for some high-risk groups, deliberately prescribed monitoring would deliver more benefit per clinical hour than responding to consumer alerts in the general population.
Researchers and funders should prioritise studies of the downstream effects of consumer alerts on health service use and outcomes. That evidence gap is the main reason the debate remains stuck.
What to watch next
Watch whether health systems begin publishing how many wearable-triggered contacts they handle and what share lead to a new diagnosis. Watch whether device makers share real-world performance data with regulators and clinicians. And watch whether any payer starts funding a structured triage service for consumer-device alerts. The first organisation to do that well will learn more about the real value of these devices than a decade of marketing claims.
The takeaway
The debate about whether wearable data is accurate enough is largely settled for some measurements and still open for others. The more important problem is design. Consumer wearables are running a screening programme with no eligibility criteria, no agreed pathway and no one accountable for the downstream costs.
Clinicians did not choose to join that programme. But they are the ones running its follow-up clinic.
How does your practice handle patient-generated wearable data today, and who actually reviews it? Share what’s working, or not, in the comments.
