How far off is your watch when your heart rate changes quickly? We pulled 1270 device results from 139 studies and reran 7,518,782 paired readings from 42 people in two public datasets. Pick an activity below and open a device to see the tests behind the numbers. Search the full library of 223 studies
Heart rate accuracy by device
1,270 device results from 139 studies.
Error shows how far off the device was. Reads high or low shows whether its readings averaged above or below the reference.
| 1 | 0.67 bpm | 1Preliminary | Near even−0.18 bpm1 study | |
| 2 | 0.78 bpm | 1Preliminary | Low−1.52 bpm2 studies | |
| 3 | 0.9 bpm | 1Preliminary | Near even+0.1 bpm1 study | |
| 4 | 0.92 bpm | 2 | Near even−0.29 bpm2 studies | |
| 5 | 0.99 bpm | 1Preliminary | Low−1.34 bpm2 studies | |
| 6 | 1.18 bpm | 1Preliminary | Near even−0.31 bpm1 study | |
| 7 | Generation unspecified · Jeong Y et al. (2026) | 1.78 bpm | 1Preliminary | Near even+0.05 bpm1 study |
| 8 | 1.81 bpm | 1Preliminary | Near even+0.14 bpm4 studies | |
| 9 | 1.9 bpm | 1Preliminary | Low−0.99 bpm2 studies | |
| 10 | 1.94 bpm | 1Preliminary | Low−0.61 bpm1 study | |
| 11 | 2.08 bpm | 1Preliminary | Near even−0.03 bpm1 study | |
| 12 | 2.6 bpm | 1Preliminary | Low−2.1 bpm1 study |
Lower error ranks higher for the selected activity and error measure. A one-study rank is preliminary.
How the studies are combined
We extracted absolute errors from 98 studies. Each study contributes one estimate per device. Treadmill results include supervised clinical tests. We use the whole test when available, or the median across its speed or intensity stages.
We average the study estimates. The weight is the square root of the smallest reported participant count for those tests, capped at 100 participants. We use 10 participants as a weighting default when the count is unavailable. Manufacturer white papers and studies funded by a tested device maker receive half weight. With one study, that study supplies the whole estimate. Open a row to see each source’s share, or switch off weighting to give each study equal weight.
Direction summarizes signed differences in bpm against the reference. A negative value means the wearable read lower on average. Mixed means study estimates fell on both sides of zero outside a 0.5 bpm band. Near even means all the estimates fell within that band. The Studies column counts accuracy sources. Direction shows its own source count.
The ranks exclude one-person reports and results optimized by retrospectively shifting device timestamps. When a study gives only a family name or pools model variants, the row represents that study’s tested device or group. Protocols, wear sites, sampling windows and people vary. Open a row to see its original tests.
What a workout average can hide
Right after a lifting set, the ECG read 147 bpm on average. The Galaxy Watch 5 showed 121 bpm. A minute later, their averages were less than a beat apart. In this 70-person study, researchers checked what four devices displayed before a set, right after it and a minute later. Yun, Ko and Lee, 2026
The Galaxy Watch 5’s session-average error was 4.5 bpm, compared with 26.4 bpm right after a set. For the Apple Watch Series 8, it was 1.8 bpm across the session and 9.6 bpm right after. Those are two ways to measure the same workout. A session average can look close even when the reading you check between sets is far off. Resistance exercise study, Tables 1 and 3
All four models tended to read low immediately after sets. The Garmin Vivosmart 5 was 15.4 bpm off at that moment and the Fitbit Charge 5 was off by 22.0 bpm. A minute later, average errors for all four devices were back near where they started.
A follow-up put 60 different people through three visits. On the Galaxy Watch 4, the five-second error averaged 13.5 bpm after bench press and 35.4 bpm after squats. For individual five-second readings, the reported limits of agreement went from 52 bpm low to 2 bpm high after rounding. Averaging the eight readings per visit narrowed that to 37 to 13 bpm low. You can see the same device give a very different answer depending on when you look and how you average. Lee, Bae and Ko, Tables 3 and 5
Those were spot checks of the numbers people could see. A watch may lag while the pulse is changing, reject a noisy reading or update the screen at a different rate. The lifting test measures the combined effect. It can’t tell how much each of those mechanisms contributed. The study involved healthy young men, and everyone did the exercises in the same order. A late-set difference could reflect the exercise, the preceding workload or both. Protocol and limitations
What we found when we reran the treadmill data
Neudorfer and colleagues released one-second readings from four optical monitors paired with medical-grade ECG. We compared quiet stretches of heart rate with stretches where the ECG rate was changing quickly, within the same treadmill run and the same person. Each person counted once in the summary, whether they gave us 50 usable seconds or 500. Original recordings
The Garmin Venu 2S went from 2.15 bpm off in steadier stretches to 6.34 bpm off in rapid changes, using the median of each person’s average error. The typical person was more than 5 bpm off on 8.2% of steady readings and 45.2% of rapidly changing ones. Eighteen of the 21 people included for this model had larger average errors during rapid changes.
The two upper-arm sensors also had larger errors when heart rate moved quickly. The typical error went from 1.25 to 5.27 bpm for Polar Verity Sense and 1.22 to 4.64 bpm for Scosche Rhythm24. The Polar Vantage M2 moved in the same direction, from 6.04 to 8.20 bpm, with a wider interval for the within-person difference that included zero. Each comparison applies to the people who had usable data in both conditions for that device. Download the treadmill results
Changing the rapid cutoff from 8 to 10 or 12 bpm gave the same overall pattern in the medians and left fewer people with enough data. For Garmin, the stricter 10 bpm cut gave 2.08 vs. 9.87 bpm across 12 people. That’s a sensitivity check from the same study, and the smaller subset makes it a less stable estimate. Download the sensitivity checks
What if you compare at the same heart rate?
Five beats off at 75 bpm is a bigger percentage miss than five beats off at 150. So we ran the comparison again within each person, using only 10-bpm bands where we had enough readings in both steady and rapidly changing stretches.
For the Garmin Venu 2S, the typical increase in error went from 3.66 bpm before matching to 2.9 bpm after matching, using the same 16 people both ways. Every one of them had higher error when their heart rate was changing quickly. The typical increase was about 3 percentage points of the reference rate. The adjusted increase was also positive for most people on each of the other devices. The Polar Vantage M2’s percentage interval still included zero. Matched comparisons for all four models
Does this happen in other activities?
Sometimes. In a separate walking experiment, the Apple Watch SE moved from 4.5 bpm average error during steady walking to 6.3 bpm during stop-start walking, using the median across participants. The Fitbit Charge 6 stayed close to where it started, 3.8 to 3.4 bpm. In daily life, the median error for the Garmin Venu 2S was 2.65 bpm, with the middle half of participants between 1.89 and 3.47 bpm. Its typical participant was more than 5 bpm off on 11.3% of paired readings. Gielen et al., Table 3 · Daily-life recordings
Those bigger misses clustered at higher rates. For a typical Venu 2S wearer, about one in ten paired readings was off by more than 4 bpm below 80, versus 25 bpm at 110 to 139. The first band included 32 people and the second 29. These bands can contain different people and activities. See errors by reference rate
Data gaps can change how a number reads. In a 12-lead ECG study, the Garmin Vivosmart 4 had 4.12% median error at per-second matches, and roughly 93% of the ECG observations went unmatched at that resolution. With 60-second windows the median error was 3.23% and about 35% went unmatched. The export and the averaging window determine which readings the comparison can see. Van Oost et al., Table 3
The pool was a different story. The Garmin Venu Sq’s typical disagreement with the chest strap rose from 3.35 bpm running to 61.2 bpm swimming in the same 10-person study. For the Polar Vantage V2 it rose from 7.34 to 27.55 bpm. The strap can struggle underwater too, and the released file contains only times when all three devices logged a value. These numbers tell us how far apart the devices were during the pairs we can see. Swimming and running recordings
Same device, very different results
Take the Fitbit Charge 4. In a 26-person field study against a Polar H10, it was off by 1.2% on runs, 16.2% in badminton and 17.5% in soccer on average. The study looked at regular sports sessions. It did not measure how each person moved their wrist, so the numbers show where the device struggled. Ceugniez et al.
Even wearing the same sensor differently can move the result. In one cycling test, the Polar Unite’s coefficient of variation fell from 20.3% to 6.2% when researchers increased the band tension. A different study put two Polar Verity Sense bands on the same arm, one over a tattoo and one just beside it. At rest, mean absolute percentage error was 22.9% over the tattoo and 2.9% on nearby skin. Bretonneau et al. · Navalta et al., Table 1
For newer devices, Apple’s 2026 white paper covers outdoor running, cycling, intervals and lifting, and uses a Polar H10 reference. It reports the difference between Series 12 and each comparator, with Apple using its own logger. Temple’s white paper gives absolute error across treadmill zones for the Series 12, Fitbit Air, WHOOP MG, Forerunner 970 and Oura Ring 5. We show those results with the manufacturer label and the paper’s main exclusion rule visible. The different protocols answer different questions about the hardware. Apple white paper · Temple white paper
When the difference matters
If you want a rough workout average, look for evidence about the average. If you’re timing short intervals or watching how quickly your heart rate drops, look for a comparison that checks those moments. For someone deciding whether a watch is good enough for that job, a side-by-side run with an ECG chest strap and the same workout mode is more informative than a correlation number from a long steady ride. Overnight resting heart rate is its own measurement. INTERLIVE validation guidance · Overnight validation study
How we checked the evidence
Which studies went into the table?
The ranking compares the same activity and error measure. One-study ranks are preliminary. Turn on Require 2 studies to rank only exact models with a second result in that view. Open a device for the test conditions, reporting windows and signed differences. The source ledger also includes studies that report other outcomes, such as Apple’s paired differences between devices.
The study library has one finding from each of 223 original studies. We used the original abstract or named table to check each finding. It includes results such as correlation, bias and limits of agreement in the form the authors reported. The library keeps those measures with their conditions, so readers can see what was measured. Study library and inclusion notes
How the treadmill analysis works
We used only the published treadmill-running rows. We labeled a reading steady when the ECG changed by no more than 2 bpm over the preceding 10 seconds, and rapid when it changed by at least 8 bpm in either direction over that window. Each window needed contiguous ECG data, and each participant needed at least 30 paired readings in each condition for the device in question. We chose these cutoffs after a pilot and checked 10 and 12 bpm as alternatives.
For each person we calculated mean absolute error, mean signed error and the share of usable readings more than 5 bpm off. The table shows the median of each of those person-level values. For paired change, we subtracted each person’s steady error from their rapid error, took the median, and resampled people 10,000 times for a percentile interval. Overlapping seconds are observations within a person. The number of people doing the tests is the denominator for the uncertainty calculation.
The source researchers had already aligned the recordings. This analysis measures residual error and cannot reconstruct what the watch displayed in real time or recover missing readings. The 5 bpm tolerance is descriptive. Age, placement, movement, fit, device firmware and how a number is exported all matter for applying the results to a different workout.
How we calculated spread and matched the heart rates
In the activity reanalysis, a person needed 30 usable pairs in a condition. We averaged the magnitude of their errors in bpm and the magnitude divided by the reference rate for each reading. We kept the signed errors as their own measures. The activity results show the median across people, the middle 50% of people, and a 95% interval for the median from 10,000 participant resamples. The middle half describes how much people differed. The interval describes how precisely we estimated their median. Published limits of agreement, where shown, describe the spread of paired errors under the source’s method.
To see whether larger misses cluster at certain rates, we also show each band’s median of participant 90th-percentile absolute errors. Within a person, about one in ten paired readings exceeds their 90th percentile. A band needs five people with at least 30 pairs each. The people can differ between bands, so those rows describe the available data at that rate. Download the results by heart-rate band
For the matched treadmill comparison, we needed at least five readings of each type in a shared 10-bpm band. We weighted both conditions by the smaller count in each band and required at least 30 observations’ worth of overlap. We calculated the rapid-minus-steady difference for each person and resampled people. This exploratory check addresses the mix of reference heart rates within one treadmill protocol. Fit, motion and device-specific data gaps can still affect the result. Reproduction code
Each pair is a wearable reading alongside its reference. Two devices can share the same ECG timestamp, and one person can contribute many readings. The paired-reading total counts the free-living, lab and swimming analysis once per device and condition. The treadmill transition analysis uses a subset of those recordings.
Disclosure: Vora has a partnership with WHOOP. WHOOP’s evidence is included under the same criteria as the other devices. Vora ran the public-data analysis and the evidence review. The original researchers collected the measurements. Vora’s editorial policy
Get the full dataset, original studies and reproduction code. If you have a published workout with a synchronized watch and ECG chest-strap recording, send us the source with the exact model and activity.
Vora’s step-count study covers how we checked wearable readings against a manual count.