DON'T FORGET TO LIVE

The Standard

Wearables report numbers about your body. Almost nobody publishes how accurate those numbers are.
We do. Below is the bar for ten measurements, and every major maker graded against it — on published validation studies only.

Independent, based on published validation literature as of Oct 14, 2026. Not affiliated with the makers listed.

How this board works. For each measurement we publish THE BAR — the accuracy line, taken from the validation literature or a standards body, with its source. A maker MEETS the bar when at least one published validation of a device it sells reaches the bar under the conditions named. A maker FALLS SHORT when published validations exist and none reaches the bar. NO PUBLIC VALIDATION FOUND means we searched PubMed, Europe PMC, and the maker’s own documentation on Oct 14, 2026 and found no published validation — that is an absence of evidence, not a failure, and it is never graded as one. Where published results conflict, both are printed. A maker that does not report a measurement is marked, not graded.

The re-grade rule. Any maker can be re-graded the day it publishes validation that meets a bar. Corrections and new evidence are invited at gus@zohrxllc.com — send the study, and the board changes.

ZOH is held to the same board. ZOH sells no sensor of its own; what ZOH displays carries its source and its method — HRV shown as SDNN or RMSSD, never blended; body-fat percentage shown as an estimate — and every product ZOH sells carries a filed evidence grade: https://zohhealth.com/#/evidence. We don’t believe it. We measure it.

94graded cells
25MEETS
24FALLS SHORT
45NO PUBLIC VALIDATION FOUND
29studies cited by PMID

Sleep duration — total sleep time vs polysomnography

THE BAR: Mean absolute error ≤ 30 minutes vs PSG.

Basis: the standardized wearable-sleep evaluation framework of Menghini et al. (2021) works an example minimal detectable change of ≈39 minutes for a single night — a device whose typical error exceeds 30 minutes cannot resolve the smallest sleep change a person would act on. Context: a 2025 meta-analysis of 24 studies and 798 participants found pooled device–PSG total-sleep-time differences of ~17 minutes (PMID 39484805).

MakerGradePublished evidence
AppleMEETSApple Watch Series 8: MAE 27.75 min, bias +19.60 (Schyvens et al. 2025, SLEEP Advances, PMID 40303381). Note: Series 6 read through a third-party app showed absolute bias 48.1 min (Miller et al. 2022, PMID 36016077).
OuraMEETSRing Gen2: absolute bias 29.0 min (Miller et al. 2022, PMID 36016077).
FitbitFALLS SHORTCharge 5 MAE 31.24 min; Sense MAE 35.23 min (Schyvens et al. 2025, PMID 40303381).
GarminFALLS SHORTVivosmart 4 MAE 54.36 min (Schyvens 2025); Forerunner 245 absolute bias 45.3 min (Miller 2022).
WhoopFALLS SHORT4.0 MAE 50.31 min (Schyvens 2025); 3.0 absolute bias 30.3 min (Miller 2022).
WithingsFALLS SHORTScanWatch MAE 60.94 min (Schyvens 2025).
PolarFALLS SHORTVantage V absolute bias 31.2 min (Miller 2022).
SamsungNO PUBLIC VALIDATION FOUNDGalaxy Watch sleep studies located report desaturation/apnea indices, not total-sleep-time error vs PSG. Searched Oct 14, 2026.
COROSNO PUBLIC VALIDATION FOUNDSearched Oct 14, 2026.
AmazfitNO PUBLIC VALIDATION FOUNDSearched Oct 14, 2026.

Sleep staging — epoch-by-epoch agreement vs polysomnography

THE BAR: Four-stage Cohen’s kappa ≥ 0.60.

Basis: kappa 0.60 is the “substantial agreement” threshold of Landis & Koch (1977, Biometrics, PMID 843571).

MakerGradePublished evidence
AppleFALLS SHORTMulti-state kappa 0.20 (Miller et al. 2022, PMID 36016077); six-device range 0.21–0.53 (Schyvens et al. 2025, PMID 40303381).
FitbitFALLS SHORTSense 2 stage sensitivities 61.7–78.0% (Robbins et al. 2024, Sensors, PMID 39460013); kappa within the 0.21–0.53 six-device range (Schyvens 2025).
GarminFALLS SHORTMulti-state kappa 0.25 (Miller 2022).
OuraFALLS SHORTMulti-state kappa 0.43 (Miller 2022); per-stage accuracy 75.5% (light) to 90.6% (REM) (Svensson et al. 2024, Sleep Medicine, PMID 38382312).
WhoopFALLS SHORTFour-stage kappa 0.47 (Miller et al. 2020, J Sports Sci, PMID 32713257); 0.44 (Miller 2022).
PolarFALLS SHORTMulti-state kappa 0.28 (Miller 2022).
WithingsFALLS SHORTScanWatch was one of the six devices in Schyvens 2025; no device in that study exceeded kappa 0.53.
SamsungNO PUBLIC VALIDATION FOUNDSearched Oct 14, 2026.
COROSNO PUBLIC VALIDATION FOUNDSearched Oct 14, 2026.
AmazfitNO PUBLIC VALIDATION FOUNDSearched Oct 14, 2026.

Heart rate — vs electrocardiography

THE BAR: Error ≤ 5% (median absolute percentage error, or MAPE) vs ECG in at least one named condition.

Basis: Shcherbina et al. (2017, J Pers Med, PMID 28538708) — 7 devices, 60 participants, telemetry reference. Six of seven devices met 5% cycling; three met it walking.

MakerGradePublished evidence
AppleMEETSWalking median error 2.5%; lowest overall heart-rate error of the 7 devices (Shcherbina 2017).
FitbitMEETSSense rest/recovery MAPE ≤ 3% vs ECG (Jamieson et al. 2024, Sci Rep, PMID 39362983). Also published: peak-exercise MAPE ~12% (Jamieson 2024); Surge walking median error in the 6.5–8.8% band (Shcherbina 2017).
GarminMEETSVivoactive 4 rest/recovery MAPE ≤ 3% vs ECG (Jamieson 2024). Peak-exercise MAPE ~7% (same study).
WhoopMEETSNocturnal heart-rate absolute bias 0.7 bpm, ICC 0.99 vs ECG (Miller 2022, PMID 36016077). Agreement is lower during dynamic resistance exercise (Støve & Hansen 2023, PMID 36803578).
PolarMEETSH10 RR-interval detection 99.6% correct vs visual ECG inspection; 99.4% at high exercise intensity (Gilgen-Ammann et al. 2019, PMID 31004219).
SamsungFALLS SHORTGear S2: highest overall heart-rate error of the 7 devices; cycling median error 5.1%, walking in the 6.5–8.8% band (Shcherbina 2017). No published validation at or under 5% located.
Amazfit (Xiaomi)FALLS SHORTMi Band 3 resting MAE 10.2 bpm — the highest of the consumer devices tested (Bent et al. 2020, npj Digit Med, PMID 32047863).
OuraNO PUBLIC VALIDATION FOUNDNo published heart-rate error statistics vs ECG for daytime/exercise conditions located. Nocturnal heart rate is scored on the resting board below. Searched Oct 14, 2026.
WithingsNO PUBLIC VALIDATION FOUNDNo published heart-rate error statistics vs ECG located. Searched Oct 14, 2026.
COROSNO PUBLIC VALIDATION FOUNDSearched Oct 14, 2026.

Resting / nocturnal heart rate — vs electrocardiography

THE BAR: MAE or |bias| ≤ 2 bpm, or ICC ≥ 0.90, vs ECG.

MakerGradePublished evidence
AppleMEETSAbsolute bias 1.5 bpm, ICC 0.96 (Miller et al. 2022, PMID 36016077).
FitbitMEETSCharge 4 nocturnal MAE 1.16 ± 0.42 bpm vs PSG-ECG (Doheny et al. 2024, J Clin Sleep Med, PMID 38450553; cohort n=10, Huntington’s disease).
OuraMEETSNightly heart-rate bias −0.63 bpm, r² = 0.996 (Kinnunen et al. 2020, Physiol Meas, PMID 32217820).
WhoopMEETSAbsolute bias 0.7 bpm, ICC 0.99 (Miller 2022).
PolarMEETSVantage V absolute bias 1.5 bpm, ICC 0.93 (Miller 2022).
GarminFALLS SHORTForerunner 245 absolute bias 5.4 bpm, ICC 0.41 (Miller 2022).
SamsungNO PUBLIC VALIDATION FOUNDNo resting-heart-rate-specific published error figure located. Searched Oct 14, 2026.
WithingsNO PUBLIC VALIDATION FOUNDSearched Oct 14, 2026.
COROSNO PUBLIC VALIDATION FOUNDSearched Oct 14, 2026.
AmazfitNO PUBLIC VALIDATION FOUNDSearched Oct 14, 2026.

Heart rate variability — vs ECG-derived RMSSD

THE BAR: The maker’s reported nightly HRV agrees with ECG-derived values at ICC or r ≥ 0.90 in a published validation.

Which statistic a maker reports — SDNN or RMSSD — is scored in the transparency set below. The two statistics are not interchangeable.

MakerGradePublished evidence
OuraMEETSNightly HRV r² = 0.980, bias −1.2 ms vs ECG (Kinnunen et al. 2020, PMID 32217820). Note: a 5-minute-window comparison found ICC 0.63 (Miller et al. 2022, PMID 36016077).
WhoopMEETSRMSSD ICC 0.99 vs ECG; limits of agreement ±7.6 ms (Miller 2022). The paper discloses that the raw data were provided by the manufacturer.
PolarMEETSH10 RR intervals — the intervals Polar HRV is computed from — 99.6% correct vs ECG (Gilgen-Ammann et al. 2019, PMID 31004219). Note: Vantage V optical nightly HRV ICC 0.65 (Miller 2022).
AppleFALLS SHORTApple Watch HRV vs ECG ICC 0.67, limits of agreement ±55.2 ms (Miller 2022). Apple reports SDNN, not RMSSD — see the transparency set.
GarminFALLS SHORTHRV ICC 0.24, limits of agreement ±92.0 ms (Miller 2022).
FitbitFALLS SHORTCharge 4 HRV ICC 0.69 (healthy) / 0.87 (COPD) vs Polar H10 (Hermans et al. 2024, JMIR mHealth, PMID 39504450).
SamsungNO PUBLIC VALIDATION FOUNDSearched Oct 14, 2026.
WithingsNO PUBLIC VALIDATION FOUNDSearched Oct 14, 2026.
COROSNO PUBLIC VALIDATION FOUNDSearched Oct 14, 2026.
AmazfitNO PUBLIC VALIDATION FOUNDSearched Oct 14, 2026.

Blood oxygen (SpO2) — vs co-oximetry / arterial reference

THE BAR: Arms ≤ 3.5% across the 70–100% SaO2 range.

Basis: the FDA premarket accuracy criterion for reflectance pulse oximeters, as applied in Walzel et al. (2023, PMID 38005550). ISO 80601-2-61 permits Arms ≤ 4.0%.

MakerGradePublished evidence
AppleMEETSWatch Series 8 Arms 2.2%, bias −0.1 vs reference oximeter, full range (Walzel et al. 2023, Sensors); Series 7 MAE 2.2% vs clinical oximeter (Jiang et al. 2023, PLOS Digital Health, PMID 37437005). Caveat: under controlled hypoxemia (SaO2 < 88%), Series 7 Arms was 5.82% (Jiang et al. 2026, JMIR Form Res, PMID 41894520).
SamsungMEETSGalaxy Watch 4 RMSD 1.8% vs arterial SaO2 under an FDA/ISO-style protocol (Browne et al. 2025, Sensors, PMID 40096012). Note: Galaxy Watch 5 Arms 3.8% vs reference oximeter (Walzel 2023).
WithingsMEETSScanWatch RMSE 2.97% / 3.00% (right/left wrist) vs co-oximetry (Kirszenblat & Edouard 2021, JMIR, PMID 33857011 — the paper’s authors include Withings employees); Arms 3.5% vs reference oximeter (Walzel 2023). Note: measurement success rate 69.5% in one clinical cohort (Jiang 2023).
GarminFALLS SHORTFenix 5X Plus at 4,559 m: bias +7.0% vs arterial blood gas, limits of agreement −6.5 to +20.5; the authors state the device did not meet acceptable validity criteria (Schiefer et al. 2021, Sensors, PMID 34640680).
FitbitFALLS SHORTCharge 4 SpO2 agreement poor in a COPD cohort (ICC 0.32), with overestimation (Hermans et al. 2024, PMID 39504450). Scored on agreement; no Arms-protocol study located.
PolarFALLS SHORTElixir sensor as shipped: RMSE 4.13% vs arterial SaO2 under an ISO-style protocol (Russell et al. 2024, Med Sci Sports Exerc, PMID 38635406). A post-hoc recalibration in the same paper reached 2.67%; this board grades the shipped algorithm.
OuraNO PUBLIC VALIDATION FOUNDSearched Oct 14, 2026.
WhoopNO PUBLIC VALIDATION FOUNDSearched Oct 14, 2026.
COROSNO PUBLIC VALIDATION FOUNDSearched Oct 14, 2026.
AmazfitNO PUBLIC VALIDATION FOUNDSearched Oct 14, 2026.

Step count — vs direct observation / hand count

THE BAR: MAPE ≤ 10% at normal walking speeds (4.0–6.4 km/h).

Basis: the CADENCE-Adults validity catalog (Mora-Gonzalez et al. 2022, PMID 36076265) and the 158-publication systematic review of Fuller et al. (2020, PMID 32897239). Slow-speed and free-living results degrade for most devices; free-living findings are printed where published.

MakerGradePublished evidence
AppleMEETSWatch Series 1 MAPE 4% at normal speeds (CADENCE catalog).
FitbitMEETSIonic MAPE 4% (CADENCE catalog); Sense MAPE 6.8–8.0% vs hand tally (Jamieson et al. 2024, PMID 39362983). Free-living caveat: Charge 2 overcounted by +12.4% (mean percentage error) in older adults (Tedesco et al. 2019, PMID 31219048).
GarminMEETSvivoactive 3 MAPE 1–2% (CADENCE catalog); Vivoactive 4 MAPE 0.9–1.8% vs hand tally (Jamieson 2024).
SamsungMEETSGear Fit2 MAPE 1–2% at normal speeds (CADENCE catalog).
PolarMEETSM600 MAPE 6% at normal speeds (CADENCE catalog).
WhoopDOES NOT REPORTWhoop does not publish a step count. Not scored.
WithingsNO PUBLIC VALIDATION FOUNDFor current devices. Searched Oct 14, 2026.
OuraNO PUBLIC VALIDATION FOUNDSearched Oct 14, 2026.
COROSNO PUBLIC VALIDATION FOUNDSearched Oct 14, 2026.
AmazfitNO PUBLIC VALIDATION FOUNDSearched Oct 14, 2026.

Body composition — vs DEXA

THE BAR: A published DEXA validation whose %body-fat bias sits within ±3.5 percentage points AND whose own agreement/equivalence assessment passes.

Basis: the ±3.5% standard-error-of-estimate criterion for %fat prediction (Lohman, 1992, as carried in ACSM body-composition guidance). A correlation coefficient alone is not agreement and does not meet this bar.

MakerGradePublished evidence
SamsungFALLS SHORTGalaxy Watch 5 BIA vs DXA (n=108): bias −0.88 pp, MAE 2.87 pp, limits of agreement −7.85 to +6.10 pp — and the study’s equivalence test was not met (Carrier et al. 2025, Front Sports Act Living, PMID 41341278).
WithingsFALLS SHORTBody Cardio vs DEXA: fat-mass median absolute error −3.7 kg (IQR −8.0 to +0.28) (Frija-Masson et al. 2021, JMIR mHealth, PMID 33929337). Withings’ own Body Scan study reports r = 0.99 vs DEXA — a correlation, published without bias or limits of agreement, so it does not meet this bar.
GarminNO PUBLIC VALIDATION FOUNDIndex S2: no peer-reviewed DEXA validation located (reviews and forums only). Searched Oct 14, 2026.
FitbitNO PUBLIC VALIDATION FOUNDAria line: no peer-reviewed DEXA validation located. Searched Oct 14, 2026.
Amazfit (Xiaomi)NO PUBLIC VALIDATION FOUNDMi Body Composition scales: no DEXA validation located on PubMed. Searched Oct 14, 2026.
AppleDOES NOT SELLNo body-composition device in the current line. Not scored.
OuraDOES NOT SELLNo body-composition device in the current line. Not scored.
WhoopDOES NOT SELLNo body-composition device in the current line. Not scored.
PolarDOES NOT SELLNo body-composition device in the current line. Not scored.
COROSDOES NOT SELLNo body-composition device in the current line. Not scored.

Respiratory rate — vs reference measurement

THE BAR: Published validation with MAPE ≤ 10%, or |bias| ≤ 1 breath/min, or ICC ≥ 0.90.

MakerGradePublished evidence
WhoopMEETSWithin ~1 breath/min of polysomnography on average (Berryhill et al. 2020, J Clin Sleep Med, PMID 32043961); bias 1.8%, precision error 6.7% as tabulated in the npj Digital Medicine 2021 wearable respiratory-rate review.
FitbitFALLS SHORTCharge 4 respiratory-rate ICC 0.84 (COPD) / 0.77 (healthy) vs a Polar H10-derived reference (Hermans et al. 2024, PMID 39504450).
AppleNO PUBLIC VALIDATION FOUNDSearched Oct 14, 2026.
GarminNO PUBLIC VALIDATION FOUNDSearched Oct 14, 2026.
SamsungNO PUBLIC VALIDATION FOUNDSearched Oct 14, 2026.
OuraNO PUBLIC VALIDATION FOUNDSearched Oct 14, 2026.
WithingsNO PUBLIC VALIDATION FOUNDSearched Oct 14, 2026.
PolarNO PUBLIC VALIDATION FOUNDSearched Oct 14, 2026.
COROSNO PUBLIC VALIDATION FOUNDSearched Oct 14, 2026.
AmazfitNO PUBLIC VALIDATION FOUNDSearched Oct 14, 2026.

Skin temperature — vs reference thermometry

THE BAR: A published validation of the maker’s temperature output against a reference thermometer.

Every maker that reports temperature reports deviation from the wearer’s own baseline, not an absolute clinical temperature. The board grades the validation, and prints that fact.

MakerGradePublished evidence
OuraMEETSMeets for relative tracking: nocturnal finger skin temperature tracked menstrual-cycle phase change comparably to oral temperature — follicular-to-luteal difference 0.30 °C (skin) vs 0.23 °C (oral); daily correlation with oral temperature r = 0.56 (Maijala et al. 2019, BMC Women’s Health, PMID 31783840).
AppleNO PUBLIC VALIDATION FOUNDReports wrist-temperature deviation; no published validation vs reference thermometry located. Searched Oct 14, 2026.
FitbitNO PUBLIC VALIDATION FOUNDSearched Oct 14, 2026.
GarminNO PUBLIC VALIDATION FOUNDSearched Oct 14, 2026.
SamsungNO PUBLIC VALIDATION FOUNDSearched Oct 14, 2026.
WhoopNO PUBLIC VALIDATION FOUNDReports skin-temperature deviation. Searched Oct 14, 2026.
WithingsNO PUBLIC VALIDATION FOUNDSearched Oct 14, 2026.
PolarNO PUBLIC VALIDATION FOUNDSearched Oct 14, 2026.
COROSNO PUBLIC VALIDATION FOUNDSearched Oct 14, 2026.
AmazfitNO PUBLIC VALIDATION FOUNDSearched Oct 14, 2026.

Transparency — how makers handle your data

Not accuracy — conduct. Facts from vendor documentation, checked Oct 14, 2026.

MakerPublishes its HRV statistic?Raw data export?Third-party API / sharing
AppleYes — SDNN (Apple Support). An RMSSD “Recovery HRV” was added for Series 12 / Ultra 4 in 2026.Yes — Health app exportNo cloud API — HealthKit is on-device only, read with the member’s permission
Fitbit (Google)Yes — RMSSD (Fitbit help)Yes — account exportFree API (Fitbit Web API / Google Health API). The legacy Fitbit Web API path ends Oct 30, 2026.
GarminYes — RMSSD (Garmin support, HRV Status)Yes — account exportApplication-gated — Garmin Health API partner application
SamsungNot named in consumer documentationYes — Samsung Health data downloadApplication-gated — Health Data SDK partner approval
OuraYes — RMSSD (Oura support)Yes — Oura on the Web exportFree API — Oura API v2 (a member’s Oura membership governs whether data exists)
WhoopYes — RMSSD (Whoop support)Yes — app exportFree API — Whoop developer API
WithingsNot published — no HRV method documentation locatedYes — account exportAPI exists; accounts created after Oct 12, 2026 require a Withings+ subscription to share data with third-party apps
PolarYes — RMSSD (Polar support, Nightly Recharge)Yes — Polar Flow exportFree API — Polar AccessLink
COROSNot named in consumer documentationYes — Training Hub exportApplication-gated — API access by application
Amazfit (Zepp)Yes — RMSSD (HRV4Training reference compilation, Oct 2026)Yes — app exportApplication-gated — partner API
Hume—Via Apple Health / Health Connect syncNo API. (Hume is listed for transparency only; no published validation literature was located for its devices, so it is not on the accuracy boards.)

What would change these grades? Published validation. A maker that funds an independent study, posts the protocol, and reports the error against a reference standard moves the day the paper is public — in either direction. Until then, the grades above are what the published record supports, no more and no less.

Board dated Oct 14, 2026. It will be re-run as new validation publishes. Corrections: gus@zohrxllc.com.

zohhealth.com · Watch the Films · Evidence Grades
Independent, based on published validation literature as of Oct 14, 2026. Not affiliated with the makers listed.
ZOH — Know Your Biology. Don’t Forget to Live.