How Accurate Is Your Smartwatch at Telling Deep Sleep From REM?

Your smartwatch shows 47 minutes of deep sleep last night, and you either feel smug or cheated. The problem is that number is the softest one on the screen. When researchers put three popular consumer devices head to head against polysomnography, the lab standard that reads brain waves directly, the wearables nailed the easy question and fumbled the hard one. Telling asleep from awake, all three hit sensitivity at or above 95 percent (see [1]). Telling deep from light from REM is where they came apart. The Apple Watch Series 8 underestimated deep sleep by 43 minutes and overestimated light sleep by 45 minutes on the same nights. The Fitbit Sense 2 overestimated light by 18 minutes and undercut deep by 15. The Oura Ring Gen3 was the outlier that showed no significant stage differences from the lab, which is worth knowing if stage data is why you bought the thing.

This is not one flaky brand. A separate validation ran six wrist wearables against polysomnography and found the same tell: when the algorithm is unsure, it defaults to light sleep (see [2]). Per-stage accuracy ranged from 33 to 83 percent across the devices, and agreement with the lab landed at a Cohen's kappa of 0.21 to 0.53, which sleep researchers read as fair to moderate, not good. Wake detection was the weakest link, with specificity as low as 29 percent, meaning a lot of the minutes you spent staring at the ceiling got quietly filed as sleep.

Here is the mechanism, because it changes how you should use the number. A wrist device infers stages from movement and heart rate, not from the cortical signals that actually define deep and REM sleep. That inference is a decent guess in aggregate and a bad one on any single night, especially at the transitions where stages blur. So treat the stage breakdown the way I treat any single reading: as one noisy sample, not a verdict.

What the wearable does reasonably well is total sleep and consistency over weeks, and that trend is the thing worth optimizing. If you want to know whether a change actually moved your sleep, ignore last night's deep-sleep bar and watch the fourteen-day line instead. In Wellness Project you can log a protocol change and read the /sleep trend across the window rather than reacting to one graph. The night-to-night stage split is entertainment. The slope is the signal.