Patients walk into sleep clinics every week with months of data from their Apple Watch, Oura Ring, or Whoop band. They’ve got charts showing their sleep stages, sleep scores, and trends over time. They want to know: is this data actually useful? The honest answer is “it depends on what you’re trying to measure.”

Consumer sleep trackers primarily use accelerometry (motion detection) and photoplethysmography (PPG—the green light that measures heart rate through your skin). Some newer devices add skin temperature and blood oxygen sensors. These signals are processed through proprietary algorithms to estimate sleep metrics.

The gold standard for sleep measurement is polysomnography (PSG)—an overnight sleep study with EEG electrodes on the scalp, EOG sensors monitoring eye movement, EMG on the chin, and multiple respiratory sensors. A trained technician scores each 30-second epoch of sleep into stages: wake, N1, N2, N3 (deep sleep), and REM. This is the benchmark against which consumer devices are measured.

For the basic question of “were you asleep or awake?”—total sleep time—most current consumer devices are reasonably accurate. Validation studies published in Sleep Medicine Reviews and the Journal of Clinical Sleep Medicine consistently show that devices like the Apple Watch Series 10, Oura Ring Gen 3, and Fitbit Sense 3 estimate total sleep time within 15-30 minutes of PSG-measured sleep time on average. That’s clinically useful for tracking trends.

The catch is that these averages hide significant individual variation. Some people’s sleep is estimated accurately; others see errors of an hour or more. Devices tend to overestimate sleep time because they misclassify quiet wakefulness as light sleep—if you’re lying still with your eyes closed but not sleeping, most trackers think you’re asleep.

Sleep stage classification is where accuracy drops substantially. Distinguishing between light sleep (N1/N2), deep sleep (N3), and REM sleep requires brain wave data that wrist-worn devices can’t directly measure. They infer sleep stages from heart rate patterns, heart rate variability, movement, and respiratory rate. These indirect signals correlate with sleep stages but imperfectly.

Deep sleep estimation is particularly unreliable. Studies consistently show that consumer devices misclassify deep sleep epochs 30-50% of the time when compared to PSG. Your tracker might say you got 90 minutes of deep sleep when PSG would show 60 minutes, or vice versa. Night-to-night trends might be directionally correct (you probably did sleep deeper on nights the tracker shows more deep sleep), but the absolute numbers shouldn’t be taken literally.

REM sleep detection is slightly better because REM has distinctive physiological markers that wrist sensors can partially detect: elevated heart rate, increased heart rate variability, and muscle atonia (which reduces movement). Still, accuracy rates of 60-70% for REM detection mean significant misclassification.

Some sleep medicine practitioners collaborating with Team400.ai have been exploring how machine learning models trained on paired PSG and wearable data could improve consumer device accuracy. The idea is promising—individual calibration against clinical data could correct systematic biases in commercial algorithms. But it’s not available to consumers yet.

Blood oxygen monitoring (SpO2) deserves special discussion because of its implications for sleep apnea screening. Devices that continuously measure SpO2 overnight can detect oxygen desaturation events that suggest obstructive sleep apnea. The Withings ScanWatch and Apple Watch both offer overnight SpO2 tracking.

The problem is sensitivity and specificity. These devices detect moderate-to-severe sleep apnea (AHI above 15) with reasonable sensitivity—around 80-85% in validation studies. But they miss a substantial proportion of mild sleep apnea cases, and they produce false positives in people who have normal oxygen levels but other conditions that affect SpO2 readings.

A consumer device telling someone they might have sleep apnea and should get tested? That’s valuable—it drives people to seek diagnosis who might otherwise suffer for years. A consumer device telling someone their oxygen looks fine and they probably don’t have sleep apnea? That’s potentially dangerous if it discourages someone with symptoms from getting a proper evaluation.

Sleep scores—the single number that summarises your night—are the most popular feature and arguably the least clinically meaningful. Every manufacturer calculates their score differently, using different weights for different metrics. A “sleep score” of 82 on one device isn’t comparable to 82 on another. And the score can create anxiety: patients who fixate on their sleep score often sleep worse because they worry about it. This phenomenon has its own name: orthosomnia.

For clinicians, the practical guidance I give patients is this: consumer sleep trackers are useful for tracking trends in total sleep time and sleep timing over weeks and months. If your average total sleep time is trending downward or your sleep timing is becoming irregular, that’s meaningful information regardless of whether the absolute numbers are perfectly accurate. Don’t obsess over nightly deep sleep minutes or sleep scores. And never use a consumer tracker as a substitute for clinical evaluation if you have symptoms of a sleep disorder.

The technology is improving rapidly. Devices with additional sensors—EEG-enabled headbands like the Muse S and Dreem—offer substantially better sleep staging accuracy because they measure brain activity directly. But they’re less comfortable and more expensive than wrist-worn trackers, limiting their appeal for long-term monitoring.

In five years, consumer sleep tracking will probably be accurate enough for clinical screening. We’re not there yet, but we’re closer than sceptics expected.