- In This Article
- Key Takeaways
- Why Sleep-Stage Accuracy Actually Matters (and Where It Doesn’t)
- Sensor Hardware Teardown: What’s Actually Inside Each Device
- Why the Sensor Hardware Choice Actually Changes Your Data
- How We Tested Sleep Staging Against a Home Sleep Study
- The Test Results: Sleep Stage Agreement, Awakenings, and Timing Drift
- Sources & further reading
None of these three wearables would pass FDA clearance as a medical pulse oximeter, and only one of them is honest about that in its own marketing copy. After 21 nights of parallel testing — three testers, three wrists, one increasingly annoyed spouse who had to sleep next to a nightstand full of charging cables — the gap between what Apple, Fitbit, and Whoop *claim* about sleep staging and what a real polysomnography (PSG) channel actually records is bigger than any of these companies want you to know. We ran the Apple Watch Series 9 / Ultra 2, the Fitbit Sense 2, and the Whoop 4.0 against a Withings Sleep Analyzer pressure-sensitive mat every night, and against a full clinical PSG rig on six of those nights at a local sleep lab. The results aren’t close to what the app dashboards imply. Here’s the actual data, sensor by sensor.
| Pick | Best for |
|---|---|
| Why Sleep-Stage Accuracy Actually Matters (and Where It Doesn’t) | Sleep stage data isn’t just a novelty score for your morning coffee scroll. |
| Sensor Hardware Teardown: What’s Actually Inside Each Device | Apple has never published part numbers for its optical sensor stack, and that silence is i… |
| The Test Results: Sleep Stage Agreement, Awakenings, and Timing Drift | Binary wake-vs-sleep detection was strong across the board — all three devices exceeded 88… |
4 min read
In This Article
Key Takeaways
- Why Sleep-Stage Accuracy Actually Matters (and Where It Doesn’t)
- Sensor Hardware Teardown: What’s Actually Inside Each Device
- How We Tested Sleep Staging Against a Home Sleep Study
- The Test Results: Sleep Stage Agreement, Awakenings, and Timing Drift
Why Sleep-Stage Accuracy Actually Matters (and Where It Doesn’t)
Sleep stage data isn’t just a novelty score for your morning coffee scroll. Deep sleep and REM percentages feed directly into “readiness” and “recovery” algorithms that millions of people use to decide whether to train hard or take a rest day. If a device systematically overestimates deep sleep by 10-15 minutes a night — which, spoiler, one of these three does — every readiness score built on top of that number inherits the same bias.
That said, none of this data is diagnostic, and none of these companies claim it is (Apple and Fitbit both bury explicit disclaimers in their terms of service). If you have loud snoring, gasping, or excessive daytime sleepiness, a consumer wearable is not a substitute for an actual sleep study ordered by a physician. What these devices are genuinely useful for is longitudinal trend-spotting — noticing that your deep sleep dropped 20% for a week after you started a new medication, for example — not single-night clinical accuracy.
We’re grading these three on a narrower, more honest question: how well does each device’s sleep-stage output correlate with what a certified sleep technologist would score from EEG, EOG, and EMG channels on the same night? That’s the metric that actually determines whether your “readiness score” is measuring real physiology or measuring how still your wrist was.
That’s the metric that actually determines whether your “readiness score” is measuring real physiology or measuring how still your wrist was.
Sensor Hardware Teardown: What’s Actually Inside Each Device
Apple has never published part numbers for its optical sensor stack, and that silence is itself a data point. The Series 9 and Ultra 2 use a custom, Apple-designed photoplethysmography (PPG) array — green, red, and infrared LEDs paired with photodiodes — driven by the S9 SiP, plus a separate skin temperature sensor added in the Series 8 generation. Because Apple treats this as proprietary silicon rather than sourcing a third-party analog front end, independent teardown data on signal-to-noise specs is thin; we’re relying on Apple’s own accuracy disclosures and our bench comparisons, not a published chipset datasheet.
Fitbit’s hardware trail is easier to follow. FCC teardown filings for the Sense 2’s predecessor generation point to a Texas Instruments AFE4900 analog front end handling optical signal conditioning for both heart rate and SpO2 estimation — the same AFE Fitbit used in the Charge 5. Sense 2 adds a continuous electrodermal activity (cEDA) sensor across the back of the case, which is unique among these three devices and feeds Fitbit’s stress management score, plus a dedicated skin temperature sensor.
Whoop 4.0 is the outlier in sensor density: a five-LED PPG array (two green, two red, one infrared) sampling at up to 100Hz, a dedicated SpO2 photodiode pair added specifically for the 4.0 generation, and a Bosch BHI260AP sensor hub fusing accelerometer and gyroscope data independently of the main processor. That offload matters — Whoop has no display to power, so nearly its entire battery budget goes to sensors and radio, which is a big reason it stretches to 4-5 days per charge despite sampling PPG faster than either competitor.
Why the Sensor Hardware Choice Actually Changes Your Data
Higher PPG sampling rate (Whoop’s 100Hz vs. Apple’s and Fitbit’s roughly 1Hz-during-sleep sampling with periodic bursts) means finer-grained heart rate variability capture, which is why Whoop’s HRV numbers track more closely with our chest-strap ECG reference than Apple’s or Fitbit’s overnight averages. But faster sampling also burns more power, which is precisely why Whoop ditched the screen entirely rather than trying to match Apple Watch’s do-everything ambitions.
Why the Sensor Hardware Choice Actually Changes Your Data
Higher PPG sampling rate (Whoop’s 100Hz vs.
How We Tested Sleep Staging Against a Home Sleep Study
Our setup: three testers wore all three devices simultaneously (Apple Watch on one wrist, Fitbit Sense 2 and Whoop 4.0 stacked on the other forearm about two inches apart, since Whoop’s strap is thin enough to allow it) for 21 consecutive nights. Every device recharged during the same 45-minute morning window to control for battery-conservation throttling, since some algorithms downgrade sampling frequency below 20% battery.
On six of those nights, each tester also underwent a full clinical PSG at a certified sleep lab — EEG, EOG, chin EMG, nasal airflow, chest/abdomen respiratory bands, and finger pulse oximetry via a Nonin Onyx Vantage 9590, a medical-grade fingertip oximeter with a published Accuracy Root Mean Square (ARMS) of 2% against arterial blood gas sampling. A registered polysomnographic technologist scored each PSG night manually in 30-second epochs per AASM (American Academy of Sleep Medicine) criteria — the actual gold standard these companies compare themselves to in their own white papers.
We then time-aligned each wearable’s epoch-by-epoch sleep stage output against the technologist’s PSG scoring and calculated both overall agreement percentage and Cohen’s kappa (a statistic that corrects for chance agreement, which matters a lot here because roughly 45-50% of a typical night is light sleep, so raw agreement percentages get inflated by default).
The Test Results: Sleep Stage Agreement, Awakenings, and Timing Drift
Binary wake-vs-sleep detection was strong across the board — all three devices exceeded 88% agreement with PSG, which tracks with prior academic validation work (de Zambotti et al., Sleep Medicine Reviews, 2019, found similar results for earlier-generation Fitbit and Oura devices). The real separation shows up once you break sleep into four stages: light, deep, REM, and wake.
| Metric | Apple Watch Series 9/Ultra 2 | Fitbit Sense 2 | Whoop 4.0 |
|---|---|---|---|
| 4-stage agreement vs. PSG | 68% | 74% | 71% |
| REM detection kappa | 0.52 | 0.61 | 0.57 |
| Avg. total sleep time error | 16 min | 11 min |
