Focused research · 7 November 2026

Wrist heart rate
versus ECG

Three historical primary studies in adults during aerobic exercise.

The question is how closely optical wrist readings agree with concurrent ECG during aerobic exercise. The evidence supports protocol-specific conclusions; no pooled accuracy estimate is justified here.

What the evidence supports

Agreement depends on the tested activity and sampling procedure; there is no single transferable accuracy number for a wrist device.

Heterogeneous cohorts and historical models; no pooled estimate or causal comparison of populations.

Evidence: S1 · S2 · S3 · C1 claim record

A small average signed error can hide substantial absolute error.

40 participants assigned this wrist model; repeated observations are not 120 independent people. No confidence interval inferred from SD.

Evidence: S3 · C2 claim record

Steady-state readings and handrail-assisted measurements do not establish accuracy during unrestricted exercise or rapid transitions.

No claim that every device fails during transitions; direct transition validation needed.

Evidence: S1 · S2 · S3 · C3 claim record

These studies do not establish current hardware or firmware performance, calorie accuracy, clinical benefit, or safe heart-rate treatment thresholds.

No recommendation to substitute a device for clinical monitoring.

Evidence: S1 · S2 · S3 · C4 claim record

Included studies and their limits

Two newly inspected Methods sections and one hash-reconciled retained Methods section. “Inspected” records access and assessment, not low risk of bias.

S1 · Shcherbina et al., 2017

60 healthy adults; convenience recruitment with diversity selection. Concurrent 12-lead ECG, Quark CPET.

Selected Results report less heart-rate error during cycling than walking in this protocol.

Limit: No accuracy percentages reproduced: walking threshold text is internally inconsistent; supplements and broader notice reconciliation remain pending.

Methods, source access and unresolved checks
Design
Laboratory criterion validation; two device batches; fixed activity sequence.
Independent unit
Participant; 80 tests are not 80 independent people.
Tested models
Apple Watch (generation unspecified); Basis Peak v1; Fitbit Surge; Microsoft Band v1; Mio Alpha 2; PulseOn; Samsung Gear S2
Sampling
Walking, running, cycling; final minute per stage, device-specific extraction.
Endpoints
Relative error; repeated-measures modeling; selected Results only.
Calibration and version
Manufacturer instructions; no independently verified calibration or firmware versions.
Bias concerns
Convenience selection; excluded malfunctioning device; device-specific acquisition; missing-data exclusion in PCA.
Actual access
Retained Methods 151–228 and selected Results 229–247; native current PMC challenge and PubMed footer.
Notices
Exact DOI search found publisher notes; notes request returned 429. No notice clearance.
Registry
No registry body inspected.
Capture locators
historical rawPointers in provenance-manifest.json; original captures October 31, hash rechecked November 7

S2 · Pasadyn et al., 2019

50 healthy athletic adults, age 18–56; 68% male, 76% White. Three-lead ECG, Quinton Q-tel RMS; trained visual reading.

Authors report declining wrist-device agreement at faster treadmill speeds.

Limit: Rendered concordance decimals are malformed; exact coefficients and device ranking withheld pending figure/publisher reconciliation.

Methods, source access and unresolved checks
Design
Prospective laboratory validation; randomized device and wrist assignment.
Independent unit
Participant; 25 participants per wrist model.
Tested models
Apple Watch III; Fitbit “Iconic” (source spelling; model identity unresolved); Garmin Vivosmart HR; TomTom Spark 3
Sampling
Six treadmill speeds, readings after two minutes; handrail gripping during roughly five-second observation.
Endpoints
Concordance, paired differences, Bland–Altman; mixed model for repeated observations.
Calibration and version
Placement described; proprietary algorithms and firmware not verified.
Bias concerns
Selected fit volunteers; altered arm motion; visual timing; missingness/registry agreement not established.
Actual access
Main Methods/statistics 135–197, Results 199–216 and limitations 223–225 read.
Notices
PubMed identity matched; exact DOI notice search found no identity-matched notice. Bounded screen only.
Registry
NCT03612063 reported by paper; direct registry returned shell, so prespecification unverified.
Capture locators
native-03-block-0:135–225; native-05-block-27:177–251

S3 · Etiwy et al., 2019; erratum 2020

80 adults in cardiac rehabilitation; mean age 62, 19% female. Torso limb-lead ECG, Quinton Q-tel RMS; cardiologist-supervised visual reading.

Fitbit treadmill mean signed difference was 0.5 bpm but mean absolute difference 8.2 bpm (Table 2); near-zero bias can coexist with individual error.

Limit: Descriptive paired readings, not an exercise prescription or participant-level confidence interval.

Methods, source access and unresolved checks
Design
Prospective criterion validation; randomized wrist-device allocation.
Independent unit
Participant; repeated time points are not additional patients.
Tested models
Apple Watch (generation unspecified); Fitbit Blaze; Garmin Forerunner 235; TomTom Spark Cardio
Sampling
Individually prescribed treadmill/cycle exercise; readings at 3, 5, 7 minutes.
Endpoints
ECG-minus-device differences, absolute differences, concordance, Bland–Altman.
Calibration and version
Electrode preparation and wrist placement described; firmware/calibration not independently verified.
Bias concerns
Visual timing; model-specific missing displays; selected single-center cohort; limited subgroup precision.
Actual access
Main Methods 175–204, Results 206–317, limitations 335–338 read.
Notices
Identified erratum read: corrects coauthor name to Zade Akras; no numerical change in that notice.
Registry
NCT03182439 reported by paper; registry shell prevents prespecification check.
Capture locators
native-07-block-0:175–204; native-08-block-20:275–299; native-09-block-2:335–338; native-06-block-2:152–164

Search boundary and missing evidence

The protocol was frozen before retrieval on 7 November 2026. Eligible publication dates were 1 January 2010 through 7 November 2026. Seven targeted web-search operations returned 75 result blocks, including duplicates and irrelevant records; this was not a complete PubMed database search. Six primary candidates were attempted, and three were included.

Ten of 22 open operations returned challenges, rate limits, footers or registry shells. A newer 2025 candidate could not be assessed beyond its abstract and early text. Another candidate remains held because its minimum participant age was unspecified and its error table requires reconciliation. These gaps limit currency and coverage.

The identified Etiwy erratum was read and corrects a coauthor name. Shcherbina’s publisher notes were inaccessible; broader notice assessment remains unresolved. No inaccessible correction is treated as harmless, and no comprehensive absence-of-notices claim is made.

Raw native responses remain outside Git. A repository clone supplies authored notes, source links and hashes, not publisher article bodies or the machine-specific evidence archive.

Continue the review

Resolve source-text inconsistencies, registry records, missingness and firmware details before extending quantitative claims. Retrieve the blocked recent Methods through supported public tools or lawfully supplied text. A qualified external reviewer should assess extraction, agreement statistics, selection bias and clinical applicability.