Oura, Whoop, Apple Watch: what your sleep and HRV data are worth in a consultation

A ring or a watch has essentially three sensors: photoplethysmography (PPG), an accelerometer, and a skin temperature sensor.

PPG shines LEDs into tissue and measures how much light returns. Each heartbeat changes the blood volume in the capillary bed, which produces a pulse waveform. The accelerometer captures movement. The temperature sensor captures skin temperature at one point.

Everything else is computation. Sleep stages, HRV, recovery score, readiness, strain and sleep quality are model outputs that an algorithm derives from those three raw signals. Their quality depends on how well the model matches physiology and how clean the raw signal is.

In short

• What is measured directly: pulse waveform, movement, skin temperature.
• Sleep stages, HRV values and scores are derived quantities from an algorithm.
• Accuracy differs substantially between those categories.

The best independent comparison tested seven devices against polysomnography, the laboratory standard with EEG, eye and muscle recording (Chinoy et al., Sleep, 2021). Per 30-second epoch, sensitivity for sleep reached at least 0.93 on every device. Specificity for wake landed at 0.18 to 0.54.

That is the central finding. The devices reliably detect that you are asleep. Wake periods in bed they detect poorly, frequently logging them as sleep. Total sleep time bias ranged from −0.3 to +46.8 minutes.

A meta-analysis of 24 studies covering 798 participants and twelve devices (Lee et al., Journal of Clinical Sleep Medicine, 2025) quantifies the pooled differences against polysomnography: total sleep time −16.85 minutes, wake after sleep onset +13.26 minutes, sleep efficiency −4.69 percent.

Sleep stages are where it gets starker. An independent evaluation of five devices (Kainec et al., Sensors, 2024) found a mean absolute percentage error of 59 to 139 percent for wake after sleep onset. Devices overestimated wake on quiet nights and underestimated it on restless ones. A comparison of six devices against polysomnography (Miller et al., Sensors, 2022) found four-stage agreement of 50 to 65 percent (κ 0.25 to 0.52). Even the best device disagreed with the laboratory on roughly a third of epochs.

Oura Gen 3 with the OSSA 2.0 algorithm reports better figures: 91.7 percent accuracy for sleep versus wake and 90.6 percent for REM across 421,045 epochs (Svensson et al., Sleep Medicine, 2024). Two things belong alongside that: the study was conducted with Oura Health involvement, and a published methodological critique of it exists.

In short

• Sleep duration and sleep timing are usable, with a typical error in the region of a quarter of an hour.
• Wake periods in bed are systematically underestimated.
• Deep sleep and REM minutes are estimates with a wide error band and do not function as measurements.

That third line has an uncomfortable consequence. The people who lie awake at night are often the ones whose device tells them they slept well.

Heart rate variability describes the fluctuation in intervals between consecutive heartbeats. The metric most wearables use is rMSSD, which predominantly reflects vagal, parasympathetic activity (Laborde et al., Frontiers in Psychology, 2017).

Two hard constraints limit what a single value can tell you.

First, the spread among healthy people. A synthesis of 44 studies covering 21,438 healthy adults (Shaffer & Ginsberg, Frontiers in Public Health, 2017) gives a mean rMSSD of 42 ms with a range of 19 to 75 ms, and SDNN of 50 ms across 32 to 93 ms. An rMSSD of 28 ms sits comfortably in the healthy range, and so does one of 70 ms. The same paper states that 24-hour, short-term and ultra-short-term values are not interchangeable, which means an overnight ring value cannot be held against published reference ranges either.

Second, the measurement accuracy of the devices. The comparison study cited above (Miller et al., 2022) tested overnight rMSSD against ECG:

Heart rate itself was accurate on most devices, with absolute biases of 0.7 to 5.4 beats per minute. For rMSSD the errors are an order of magnitude larger. A bias of 10 to 22 ms against a healthy range spanning 19 to 75 ms explains why comparing your HRV with a training partner's yields nothing usable. For context: this study was funded by the Australian Institute of Sport, the authors disclose research support from Whoop Inc., and Whoop performed best.

A meta-analysis of PPG-derived pulse rate variability (Xu et al., Sensors, 2026) shows that agreement with ECG is good under controlled resting conditions. The authors state explicitly that these results should not be generalised to sleep, exercise, stress or free-living settings. Those are precisely the conditions wearables operate in.

What remains is the trend. Sports science has worked with averaged weekly values rather than single measurements for years, because single values fluctuate too much (Plews et al., Sports Medicine, 2013). A seven-day average falling across three weeks is a signal worth following up. A single night's value stays outside that picture.

In short

• The healthy range for rMSSD spans roughly four-fold, so comparing across people is worthless.
• Device error sits in the same order of magnitude as genuine biological differences.
• What is usable: the change in your own averaged value over weeks.

For a trend worth trusting, keep the conditions constant: same position, no caffeine for at least two hours beforehand, no alcohol for 24 hours, and no hard session the previous evening that goes unrecorded (Laborde et al., 2017).

The criticism above concerns the precision of individual values. The usefulness lies elsewhere.

Behaviour change. An umbrella review of 39 systematic reviews covering 163,992 participants (Ferguson et al., The Lancet Digital Health, 2022) found that activity trackers produced roughly 1,800 additional steps per day, around 40 minutes more walking, and about one kilogram of weight reduction. For blood pressure, cholesterol and HbA1c, effects were "typically small and often non-significant." That is the honest picture: the devices measurably change behaviour, and the biomarkers do not automatically follow.

Making regularity visible. The most reliable number from a wearable is the time you go to bed. A two-and-a-half-hour spread across the week often explains more in a consultation than any deep sleep figure.

Catching physiological shifts early. In a Stanford cohort, 26 of 32 COVID-19 cases showed changes in heart rate, step count or sleep, with 22 of 25 analysable cases detected at or before symptom onset (Mishra et al., Nature Biomedical Engineering, 2020). That was a retrospective analysis with no reported false-positive rate, so it is not a validated screening test. As a demonstration that physiological change shows up in the data, it remains notable.

Resting heart rate and atrial fibrillation. Resting heart rate across weeks is among the more robust quantities. For rhythm disturbances, the EHRA practical guide (Svennberg et al., EP Europace, 2022) distinguishes ECG-based from PPG-based functions and holds that clinician overreading of the recording remains necessary in every case.

In 2017, Baron and colleagues coined the term orthosomnia in the Journal of Clinical Sleep Medicine: a case series of patients seeking help for self-diagnosed sleep disturbance, triggered by tracker reports of light or restless sleep. The authors describe how the data can become a perfectionistic quest for ideal sleep, and how patients may distrust a clinical assessment that contradicts the device.

Worrying about the sleep score lengthens sleep latency. The device's error in detecting wakefulness compounds it. When the data itself becomes the burden, a tracking break belongs among the therapeutic options.

The Apple Watch ECG app and irregular rhythm notification are CE-marked medical device functions in the European Economic Area (Apple, 2019). The pivotal trial of roughly 600 participants against a 12-lead ECG produced 98.3 percent sensitivity for atrial fibrillation and 99.6 percent specificity for sinus rhythm, with 87.8 percent of recordings classifiable at all. The remaining twelve percent belongs in the figure.

Oura and Whoop market their sleep, readiness and recovery outputs as general wellness features. How actively that boundary is policed was visible in an FDA warning letter to Whoop in July 2025 concerning its blood pressure feature, in which the agency held that blood pressure estimation is not a low-risk wellness function. The matter was subsequently closed.

If you bring your exports, we work through them in this order.

1. Regularity first. Bedtimes and wake times across 30 to 90 days, including the weekend spread.
2. Total sleep time as a weekly average. Individual nights stay out of it.
3. Resting heart rate as a trend. A resting heart rate climbing over weeks at unchanged training load is a concrete reason to look further.
4. HRV as a seven-day average against your own baseline. Absolute values and comparisons with other people stay out.
5. Deep sleep and REM minutes stay out. Where there is a reasoned suspicion of sleep apnoea or another sleep disorder, we arrange diagnostics built for that purpose.
6. Cross-check against labs and history. Where fatigue persists despite unremarkable tracker data, blood diagnostics is the next step.

Your wearable is a good behaviour sensor and a rough physiology sensor. Bedtimes, weekly average sleep duration, resting heart rate and activity are solid enough to inform a clinical assessment. Deep sleep minutes and a single night's HRV are not.

The most productive way to use them: treat the numbers as the starting point for a question to bring to us. If your data has been moving in one direction for weeks, or if you feel persistently exhausted despite good numbers, bring the exports to your first appointment. We go through them together and decide which measurement will actually answer the question.