Every sleep tracker on the market produces the same confident output: a score, a stack of colored stage bars, a percentage. What the marketing does not say is that the device is guessing — inferring brain states from movement and heart rhythm — and that the quality of that guess has now been measured against the clinical gold standard in dozens of validation studies. The pattern in the research is consistent enough to summarize in one sentence: modern trackers are genuinely good at the simple question (were you asleep, and for how long) and much weaker at the interesting one (what kind of sleep, and why you feel the way you do).
What “accurate” means: the polysomnography standard
Polysomnography (PSG) is the reference test. A lab sleep study records brain waves by EEG, eye movements by electrodes near the eyes, muscle tone at the chin, plus airflow, breathing effort, heart rhythm and blood oxygen. Sleep stages are defined by those brain and eye signals: wake, light sleep (N1 and N2), deep sleep (N3), and REM. That definition is the core problem for wearables — a wrist device or a ring measures none of those signals directly. It sees movement, heart rate, heart-rate variability and sometimes temperature, and a machine-learning model translates those proxies into its best estimate of a stage chart.
Before consumer wearables, researchers used actigraphy — a motion sensor worn on the wrist — and its performance established the baseline. Validation work found actigraphy detects sleep reliably, typically in the 86 to 94 percent sensitivity range, but systematically overestimates total sleep time and sleep efficiency, because lying still looks identical to sleeping from the sensor’s point of view. Modern consumer devices add optical heart rate and HRV to the motion signal, which improves the estimate without changing what it fundamentally is: an inference.
What validation studies actually show
The most-cited recent comparison, published in Sensors in 2024 by Robbins and colleagues at Brigham and Women’s Hospital, put an Oura Ring Gen3, a Fitbit Sense 2 and an Apple Watch Series 8 on 35 adults during a lab polysomnography night. The good news was real: all three devices detected sleep versus wake with sensitivity of 95 percent or higher — better than older research-grade actigraphy — and each estimated total sleep time within roughly 10 minutes of PSG. One disclosure that matters when weighing the results: the study was funded by Oura, though independently conducted, and Oura performed best in it.
The stage numbers tell the rest of the story. In that same study, four-stage agreement measured by Cohen’s kappa came out at 0.65 for Oura, 0.60 for Apple and 0.55 for Fitbit — moderate to substantial agreement, achieved in near-ideal conditions with healthy adults in a lab. Stage-level sensitivity ranged from about 50 to 86 percent across devices, and the per-stage bias was uneven: the Apple Watch overestimated light sleep by around 45 minutes per night while underestimating deep sleep by roughly 43 minutes; Fitbit overestimated light sleep by 18 minutes and underestimated deep sleep by 15.
Independent, less flattering studies agree. A 2023 multicenter validation in JMIR mHealth and uHealth tested 11 consumer trackers and found the Apple Watch Series 8 and Oura Gen 3 in the fair-agreement band (kappa 0.2 to 0.4) for stage classification, with the Fitbit Sense 2 reaching moderate agreement. A six-device validation published in SLEEP Advances found every device detected over 90 percent of sleep epochs, but wake detection specificity ran between just 29 and 52 percent, with overall kappa from 0.21 to 0.53. A 2024 JMIR systematic review of the Fitbit Charge 4, Garmin Vivosmart 4 and WHOOP found total sleep time disagreement as low as about a minute for WHOOP but REM disagreements up to about 21 minutes, and concluded all devices need improvement at stage classification.
Why stage breakdowns are the weakest part
The reason is physics, not sloppiness. Deep sleep and REM are defined by electrical activity in the brain and eyes; a wrist sensor receives mechanical motion and pulse waves several centimeters downstream of the heart. Algorithms can infer stages from heart-rate variability rhythms, breathing regularity and movement patterns, and they do it well enough to beat chance by a wide margin — but the margin between “better than guessing” and “clinical measurement” is exactly where the validation numbers land. Individual-night deep-sleep and REM estimates are the least reliable figures on the screen; concordance in the better studies is poor at the single-night level.
The practical translation: a month of deep-sleep bars can show a real trend (more deep sleep this week than last), because the same device making the same systematic error still tracks change. A single night’s “you got 52 minutes of deep sleep” is a number with no meaningful error bar, dressed as precision.
The wake problem: quiet wakefulness counts as sleep
The SLEEP Advances specificity finding — wake detected correctly only 29 to 52 percent of the time — explains the most common complaint tracker owners have. Lying awake in bed, still and frustrated, produces a motion signal and a heart-rate pattern close enough to sleep that the device scores it as sleep. Studies in insomnia populations found the same direction of bias: both research actigraphy and Fitbit devices overestimated total sleep time and sleep efficiency while underestimating how long falling asleep took and how often people woke.
If you sleep well, this barely matters. If you have insomnia, the tracker will tell you your sleep is better than your experience of it, which is a strange kind of gaslighting — and it is worth knowing the number is wrong before concluding that your exhaustion is imaginary.
Sleep scores, HRV and readiness claims
On top of raw metrics sits a layer of composite scores: sleep scores, readiness scores, daily energy estimates. The inputs — sleep duration, heart-rate variability, resting heart rate, movement — are measured acceptably well. The composites are where accountability disappears. Each vendor’s score is a proprietary weighting of those inputs with no published validation against any outcome: not next-day performance, not cognitive testing, not clinical measures. The same caveat applies to the recovery metrics sold to athletes, including the training-readiness scores on running watches covered in our running watch comparison.
HRV deserves specific mention because it sounds medical and is used loosely. Heart-rate variability is a real physiological measurement, and beat-to-beat precision matters for it — optical wrist sensors average pulses in ways that blunt the signal compared to a chest strap or the ECG-grade measurement in a ring. A nightly HRV trend from a wearable is a usable relative signal; an HRV number compared against some absolute “healthy range” in an app is not.
When tracking helps and when it fuels anxiety
The research supports tracking for exactly two uses. First, behavioral feedback: seeing bedtime drift, or the effect of late caffeine and alcohol on your numbers, changes behavior — sleep regularity is one of the better-validated correlates of feeling rested, and trackers measure timing and consistency well. Second, longitudinal trends after a real change (a new schedule, treatment, or training block), where consistent systematic error still shows direction.
The failure mode has a name in the sleep-medicine literature: orthosomnia — the pursuit of perfect sleep data that itself degrades sleep. People start optimizing the score, wake to check the chart, and turn a passive biological process into a performance review. If checking the app is the first thing you do in the morning and a bad score colors the day, the tracker has crossed from tool to stressor. The fix costs nothing: use the device for a month, learn your patterns, then stop reading the stages nightly.
If you are still deciding between device categories, the hardware differences and their practical weight are laid out in our fitness tracker versus smartwatch comparison. And whatever the device, treat the score the way the validation data does: as a well-informed guess about a night you mostly experienced correctly without it.