Wearable device validation for sleep and night-time HR/HRV
Six devices vs polysomnography/ECG in 53 healthy adults: sleep/wake agreement 86-89%, sleep-stage agreement only 50-65% (κ 0.20-0.52)
Participants & Protocol
n=53 healthy adults (26 F, 27 M; 25.4 ± 5.9 y), one night in a sleep laboratory, 9 h in bed (23:00-08:00)
Protocol
Validation study (from the abstract): one night in a sleep laboratory with six wearables worn at the same time and compared with gold-standard polysomnography (sleep) and electrocardiography (heart rate); outcomes = agreement and Cohen's kappa for two-state (sleep/wake) and multi-state (sleep stage) categorisation, with heart rate and HRV as the secondary aim.
Devices worn simultaneously: Apple Watch S6, Garmin Forerunner 245 Music, Polar Vantage V, Oura Ring Gen 2, WHOOP 3.0, Somfit; reference = polysomnography (sleep) and ECG (heart rate)
Key Findings
Sleep vs wake (two-state) agreement with PSG: Apple Watch 88% (κ 0.30), Garmin 89% (κ 0.35), Polar 87% (κ 0.44), Oura 89% (κ 0.51), WHOOP 86% (κ 0.44), Somfit 87% (κ 0.48) → all six valid for field assessment of sleep timing and duration
Sleep stages (multi-state) agreement with PSG: Apple Watch 53% (κ 0.20), Garmin 50% (κ 0.25), Polar 51% (κ 0.28), Oura 61% (κ 0.43), WHOOP 60% (κ 0.44), Somfit 65% (κ 0.52) → all six "require improvement" for specific sleep stages
Heart rate and HRV during / just before sleep were the secondary aim; the abstract reports no HR/HRV numbers (full text, checked 2026-09-27)
Inference from the design: within-device trends are the usable signal; cross-device comparison of absolute values is not supported by this study
Not from this paper: "readiness" / "recovery" scores are proprietary composites — no validation of the composite against performance outcomes is claimed here
Practical Application
In 53 healthy young adults measured for one night against polysomnography, six consumer wearables (Apple Watch, Garmin, Polar, Oura, WHOOP, Somfit) classified sleep vs wake with 86-89% agreement (valid for sleep timing and duration) but identified sleep stages with only 50-65% agreement (κ 0.20-0.52, best Somfit/WHOOP/Oura, worst Apple Watch/Garmin) — so trust a wearable's bedtime, wake time and total sleep, and treat its deep/REM minutes as a rough guess.
Use WHOOP / Oura / Garmin for trend tracking of sleep duration and morning HRV, not for absolute comparisons across devices.
Not from this paper: trust resting supine morning HRV, be skeptical of nap- or motion-contaminated values; "readiness score" as a weak heuristic combined with subjective wellness; chest straps (Polar H10) and finger-sensor apps as HRV references — house practice, see REF-361.
Limitations
Single night, laboratory, 53 healthy young adults — not athletes, not home conditions
Algorithms update over time — 2022 validation of these device generations may not apply to 2026 devices
Sleep-stage validation modest across all six consumer devices
The abstract gives no HR/HRV accuracy figures; "recovery" scores were not validated here
How this entry was checked and changed
Content correction 2026-09-27
"HRV bias < 5 ms, ICC 0.8-0.9" -> not traceable to the abstract; removed. "Sleep-stage kappa 0.4-0.6 typically" -> per-device κ 0.20-0.52 (Apple 0.20, Garmin 0.25, Polar 0.28, Oura 0.43, WHOOP 0.44, Somfit 0.52); two-state sleep/wake 86-89%, κ 0.30-0.51.
Participants added: n=53 healthy adults (26 F), one night, PSG + ECG reference; Tier N/A -> Tier 0-1 (healthy young adults).
Morning-HRV, readiness-score and chest-strap advice marked as not from this paper.
Lines labelled "full text; not checked" were read against the open-access full text (PMC9412437): 1 confirmed and relabelled.
2026-09-28: added Protocol heading and design bullet from the abstract (Europe PMC PMID 36016077).
Full text (checked 2026-10-12)
Night-time HRV (RMSSD) bias versus ECG: Apple Watch −9.6 ms, Garmin −22.4 ms, Polar −8.7 ms, Oura −10.2 ms, WHOOP −4.5 ms, Somfit −20.5 ms, with intraclass correlations of 0.67, 0.24, 0.65, 0.63, 0.99 and 0.69 respectively (full text, Table 6).
Heart rate bias versus ECG: Apple Watch 0.5 bpm (ICC 0.96), Garmin 5.0 bpm (ICC 0.41), Polar −1.1 bpm (ICC 0.93), Oura 0.1 bpm (ICC 0.85), WHOOP −0.3 bpm (ICC 0.99), Somfit 2.2 bpm (ICC 0.65) (full text, Table 6).
HRV limits of agreement were ±7.6 ms for WHOOP but ±55.2 to ±92.0 ms for the other devices; for most devices HRV was overestimated at lower ECG values and underestimated at higher ECG values (full text, Table 6).
Sampling differed: Apple Watch, Oura, WHOOP and Somfit used the entire sleep period, Polar a 4-h period within sleep, and Garmin a 3-min manual test during wake ~30–60 min before lights out; WHOOP and Somfit supplied raw data, the others were read from apps (full text).
Total sleep time bias versus PSG: Apple Watch +39.5 min, Garmin +43.8 min, Polar −0.8 min, Oura +1.5 min, WHOOP −12.2 min, Somfit −5.5 min (full text, Table 4).
All devices detected >90% of sleep epochs, but wake detection was 26% (Apple Watch) and 27% (Garmin) versus 51–57% for Polar, Oura, WHOOP and Somfit (full text, Table 3).
Participants were an active population reporting 7.5 ± 0.4 exercise sessions per week (399 min; 30% vigorous intensity) (full text).