A heart-rate number on a smartwatch looks as exact as one from a hospital monitor, but it is made differently. Published studies suggest it is often within acceptable error in laboratory tests, less reliable during some kinds of movement, and different from one device to the next.
How a wrist sensor counts heartbeats
Most wrist wearables estimate heart rate with photoplethysmography (PPG). Allen describes it as a simple, low-cost optical technique that detects changes in blood volume in the small blood vessels of the skin. Each heartbeat adds a small pulse to the signal. That pulse rides on a slowly varying baseline, which also carries the effects of breathing, the nervous system and temperature regulation. The device’s software has to pick the heartbeat out of that mix. An electrocardiogram (ECG) records the heart’s electrical activity instead, and studies of wearables use it, as a clinical-grade gold standard, for the reference.
How close the readings come
Bent and colleagues tested six wearables, four consumer and two research-grade, against an ECG patch. The 53 people who completed the study were spread roughly evenly across the six groups of the Fitzpatrick skin-tone scale. Each session of about an hour included seated rest, paced deep breathing, walking and a typing task. Accuracy differed significantly between devices and between activities, and absolute error during activity was on average 30% higher than during rest. Averaged across the skin-tone groups, the mean absolute error at rest was 9.5 beats per minute; at rest, the consumer devices averaged 7.2 and the research-grade devices 13.9. The physiology program page charts these results device by device.
Shcherbina and colleagues tested seven wrist devices on 60 volunteers, against continuous telemetry, while they sat, walked, ran and cycled. Six devices achieved a median heart-rate error below 5% during cycling, and errors were highest for walking. Calories were another matter: no device estimated energy expenditure with an error below 20%.
Nelson and Allen took a different approach. One participant, the first author, wore two consumer devices and an ambulatory ECG through 24 hours of ordinary life, 102,740 heartbeats in all. The two devices had mean absolute percentage errors of 5.86% and 5.96%, and stayed within the authors’ ±10% threshold except for one device during daily activities such as chores and brushing teeth. Results from one person, though, may not generalize to others.
Where readings drift
Movement. Bent and colleagues describe a “signal crossover” effect: the sensor can lock on to the rhythm of a repeated motion, such as walking or jogging, and mistake it for the heartbeat. Walking raised errors significantly in every device except the Apple Watch 4. For typing, the authors report that error “was higher during typing compared with rest in all devices, and often nearly as high as during walking, except for the Apple Watch and the Empatica E4.”
Skin tone. The two laboratory studies point in different directions. Shcherbina and colleagues found device error was higher for darker skin tones, as well as for men, people with a higher body mass index, and during walking. Bent and colleagues, who recruited similar numbers in each skin-tone group, saw no statistically significant difference in accuracy across skin tones. Bent and colleagues sized their study from a power analysis, which called for at least eight people per skin-tone group; even so, a result that is not statistically significant does not show that no difference exists.
The device. Both studies found differences between devices. They tested specific models and software, such as the Apple Watch 4 and Fitbit Charge 2, and Shcherbina and colleagues point to the challenge of “rapid technological advance and algorithm/product cycle upgrades”; results need not carry over to later models.
What a heart-rhythm alert means
The Apple Heart Study, funded by Apple, enrolled 419,297 people who reported no atrial fibrillation. Over a median of 117 days, 2,161 of them (0.52%) received an irregular-pulse notification. A notification led to a telemedicine visit and an ECG patch in the mail. Among the 450 participants who returned a patch with usable data, applied on average 13 days after the notification, 34% showed atrial fibrillation. When a later notification arrived while the patch was being worn, 84% of the time the ECG showed atrial fibrillation at the same moment. Of 1,376 notified participants who answered a 90-day survey, 57% had contacted health care providers outside the study.
Wellness device or medical device
FDA guidance issued on 6 January 2026 says the agency does not intend to examine low-risk general wellness products to determine whether they are devices or whether they meet device requirements such as premarket review. Its examples of general wellness claims include enhancing participation in recreational activities by monitoring their consequences, “such as to monitor heart rate,” and a product that monitors pulse rate and/or oxygen saturation during exercise and hiking. FDA may also consider optical estimates of measures such as blood pressure or oxygen saturation to be general wellness products when they are intended solely for wellness, provided, among other conditions, that they do not prompt or guide specific clinical action or medical management and do not include values that mimic clinical ones unless validated. The guidance is nonbinding: it describes the agency’s current thinking.
INSTAR’s research interest
INSTAR’s interest is in the questions a careful reader can ask before relying on wearable data. Which reference was the device compared with, and during which activities? Were people of every skin tone included in numbers large enough to detect a difference? Does a single average hide the activity where the device did worst? Is the device in the study the model actually in use?
These questions apply to any research that uses wearable data, including the topics on the physiology and kinesiology program pages. The page on how this site checks a fact describes the checks applied to the sources cited here.