Apple Watch Accuracy: Eight Metrics, Very Uneven Evidence

A 2022 literature review that screened 5,583 paper titles down to 19 Apple Watch validation studies found HRV measurements had gaps in the data but were still able to detect mild mental stress. That is a narrower claim than an accuracy validation, and direct validations exist too: against a 3-lead ECG in 78 adults, the Series 6 measured beat-to-beat intervals almost exactly at rest (1.15 percent error) but its HRV reading far less closely (31.31 percent) (Bonneval et al., Sensors, 2025), and against a Polar H10 chest strap the Series 9 and Ultra 2 underestimated HRV by 8.31 ms on average (O'Grady et al., Sensors, 2024). Because Apple's SDNN and most competitors' RMSSD are different calculations from the same interbeat-interval data, a validation of one does not transfer to the other even when both are loosely called "HRV."

A 2017 treadmill study put two Apple Watches (one per wrist) against a Polar S810i chest-strap reference in 21 men walking, jogging, and running at three fixed speeds. Bias was trivial at every intensity, but correlation with the reference dropped as effort increased: very good during walking (r=0.97 on both wrists), good during jogging (r=0.92-0.93), and only fair to good during running (r=0.81-0.86).

A separate 2019 study in 40 cardiovascular disease patients undergoing a graded maximal exercise test found the same shape of result in a clinical population: heart rate agreement with a 12-lead ECG was rated clinically acceptable across rest, moderate, and maximal effort (ICC 0.729 to 0.958, mean absolute percentage error 6.3 to 10.7 percent), improving as intensity rose rather than degrading, the opposite pattern from the healthy-adult study above.

Apple's stage detection has been tested against polysomnography, the sleep-lab reference, in several studies, three of them tied to a company with a competing product. In 35 healthy adults, in a study funded by Oura, the Apple Watch Series 8 detected sleep with at least 95 percent sensitivity but underestimated deep sleep by 43 minutes and overestimated light sleep by 45 minutes (Sensors, 2024 [funding listed from Oura]). In a six-device sleep-lab comparison, the Series 8 had the highest stage agreement of the six, a Cohen's kappa of 0.53, which counts as moderate agreement (Sleep Advances, 2025). In a home sleep study of healthy adults, the Series 7 underestimated light sleep by about 74 minutes and overestimated REM by about 30 minutes, yet ranked first of the three devices tested (Sleep Medicine, 2026). In an earlier six-device sleep-lab night, the Series 6 agreed with polysomnography on sleep versus wake 88 percent of the time but had the lowest stage agreement of the six, a kappa of 0.20 (53 percent), in a lab that receives research support from WHOOP, one of the other devices tested (Sensors, 2022 [conflict statement: the authors' lab receives funding and equipment from WHOOP]). In an 11-tracker sleep-lab study at two Korean centers, the Series 8's stage agreement was fair, a kappa of 0.30 in 26 participants; that study was run by Asleep, whose SleepRoutine app was among those tested and scored highest (JMIR mHealth and uHealth, 2023 [authors work for and hold stock in Asleep, maker of the SleepRoutine app it tested]).

So the stage breakdown is directionally useful and biased in a consistent way: Apple's watches tend to report less deep sleep than a sleep lab measures. An earlier 2019 study, which ran an Apple Watch's raw motion data through a research sleep/wake classifier rather than Apple's own algorithm, found 97 percent overall accuracy and 79 percent specificity for detecting wake against an actigraph over 27 nights in 14 adults.

A 2024 validation put the Apple Watch Series 7's Cardio Fitness estimate against a gold-standard graded cycling test with a metabolic gas analyzer in 19 adults. The watch significantly underestimated lab-measured VO2 max on average (41.37 vs 45.88 mL/kg/min, p=.01), with 15.79 percent mean absolute percentage error overall.

The more important number is reliability, not the average gap: the intraclass correlation coefficient was 0.47, which the authors characterize as poor, meaning the estimate is considerably less trustworthy for any one person than the average error suggests. Accuracy was somewhat better in participants with good fitness levels (14.59 percent MAPE, ICC 0.60) than in those with poor or excellent fitness, where the watch tended to overestimate or underestimate respectively.

The newest model tested points the same way. In 40 healthy adults, the Apple Watch Series 10 underestimated VO2 max measured on a treadmill with indirect calorimetry by 6.25 mL/kg/min on average, with 13.2 percent mean absolute percentage error and wide limits of agreement (-18.23 to 5.73). It placed 5 of the participants in their correct fitness percentile band, and a further 12 within one band, against US reference standards. The authors judged individual-level agreement poor and saw possible use for averages across groups in research. An earlier study by the same group, in 30 adults wearing an Apple Watch whose model its summary does not give, found the same direction and size of error: 6.07 mL/kg/min low on average, 13.31 percent mean absolute percentage error.

The widely cited accuracy figures for Apple Watch and AFib come from the Apple Heart Study, an Apple-funded study of 419,297 people. Among notified participants who returned an ECG patch with usable data, worn on average 13 days after the alert, 34 percent had atrial fibrillation on it; when a notification arrived while the patch was being worn, 84 percent matched AFib on the patch. Both figures describe screening a general population where AFib is rare. A 2020 study tested something different and harder: 50 post-cardiac-surgery patients who frequently cycle between atrial fibrillation and normal sinus rhythm, a group where the two conditions are both common and the watch has to tell them apart in the moment, not just flag a rare event. Rhythm assessments were taken three times daily against continuous telemetry as the reference. Of 90 telemetry-confirmed AF episodes, the Apple Watch Series 4's rhythm notification correctly flagged 34, a sensitivity of 41 percent.

That is not a contradiction of the Apple Heart Study result; it is a different question asked of a different population, and the gap between the two numbers is itself the finding: a notification feature validated for detecting rare AFib in a general population should not be assumed to reliably distinguish AFib from sinus rhythm in patients who already have frequent arrhythmia. A separate single-lead ECG validation in 81 healthy adults found the Apple Watch 4 agreed reasonably well with a 12-lead reference on QT interval, a marker relevant to arrhythmia risk more broadly, with 6.85 percent mean deviation and a Lin concordance correlation coefficient of 0.779, somewhat more accurate than the Kardia Mobile app tested alongside it.

The on-demand ECG app was tested in 201 patients at a cardiology service, 31 percent of them in AFib, against a physician-read 12-lead ECG: an Apple Watch 6 detected AFib with 85 percent sensitivity and 75 percent specificity, level with a Samsung Galaxy Watch 3 and ahead of a Fitbit Sense and a Withings ScanWatch in the same patients, and 18 percent of its tracings came back inconclusive. With manual review of the inconclusive tracings, cardiologists could name the rhythm in 99 percent of single-lead ECGs across all five devices (JACC: Clinical Electrophysiology, 2023 [funding listed from Abbott]).

At ordinary oxygen levels the watch tracks medical devices closely. In 167 patients admitted with a COPD flare, the Series 6 differed from arterial blood by a mean of 0.458 percent, with a standard deviation of 2.78, so some readings were further off (Irish Journal of Medical Science, 2024). In 18 healthy adults breathing reduced-oxygen gas, the Apple Watch 8 stayed within the 4 percent limit used for medical oximeters and had the lowest error of three smartwatches tested (Sensors, 2023). Across nine studies and 969 participants, readings averaged 0.04 points below the reference, with 95 percent of readings within about 4 points either side (npj Digital Medicine, 2026).

When oxygen falls far, the gap widens. In 9 healthy adults whose oxygen was lowered step by step, the Series 7 overestimated saturation and fell within 2 percent of arterial blood 32.14 percent of the time, against 49.03 percent for a clinical finger oximeter; below 88 percent its average error was 5.82 percent against 3.52 (JMIR Formative Research, 2026). In 200 patients hospitalized with COVID-19, the Series 7 flagged 34.8 percent of the clinically significant low readings a ward oximeter caught (Mayo Clinic Proceedings: Digital Health, 2024). No study cited here has tested a Series 9 or later or any Ultra, and every one ran on the original on-watch calculation rather than the iPhone-side version US watches now use. Results by model and study.

No independent study has compared any Apple Watch's wrist temperature with an oral, armpit or core thermometer, or tested it for detecting fever or illness, in searches of PubMed and Europe PMC in October 2026. That covers Series 8 through Series 12 and every Ultra. A 2026 living systematic review of 82 Apple Watch validation studies lists wrist temperature among the metrics yet to be validated. Apple's Series 12 and Ultra 4 specifications list a temperature sensor with no accuracy figure, Apple does not describe any change to temperature sensing for those models, and its support page calls the feature not a thermometer and not a medical device.

The one published performance result is for what the temperature is used for, not the temperature itself. In an Apple-funded study with Apple-employed authors, 260 participants tracked 889 menstrual cycles against home ovulation tests. Where the watch detected a temperature shift of at least 0.2 °C, its retrospective ovulation estimate was off by a mean of 1.22 days, with 89.0 percent within two days, and about one cycle in five produced no estimate. Apple's own instructions for the feature report estimates within two days of the ovulation test across 934 cycles from 226 users, without naming a watch model.

Step counting is where Apple Watch's device-specific evidence looks best. A 2024 study of 104 adults wearing an Apple Watch 6 or Galaxy Watch 4 for 24 hours of daily activity, referenced against a thigh-worn ActivPAL accelerometer, found the Apple Watch the most accurate of the group: r=0.99 against the reference and 6.4 percent mean absolute percentage error, ahead of the Galaxy Watch (10.5 percent) and well ahead of phone-based step counting (29.6 percent). A separate free-living comparison against a Yamax pedometer in 48 adults found the Apple Watch 2 within roughly 6 percent of the pedometer's step count (11,550 vs 10,906), in line with the two Fitbit models tested alongside it.

Energy expenditure is the opposite story, and one clinical-population study found an error large enough to be worth stating plainly. In 40 patients with cardiovascular disease completing a graded maximal exercise test, indirect calorimetry measured energy expenditure that the Apple Watch overestimated with a mean absolute percentage error of 114.72 percent, a systematic bias the authors' Bland-Altman analysis confirmed rather than random noise. That is a considerably worse gap than the double-digit calorie errors this category typically reports, in a population where exercise prescription accuracy has real clinical stakes.

Nothing yet. The feature launched in September 2026, and no independent validation of the composite score, or even a third-party description of its exact weighting, was available at the time of this search. The same caveat that applies to every other brand's composite score applies here too: a score built from separately-validated inputs is not automatically as accurate as those inputs are individually.