Sleep stages from a wrist: what the device is really doing
The stage breakdown is the least reliable thing on the screen. The duration and the timing are the most.
Polysomnography — the real thing — scores sleep stages from an EEG, an EOG for eye movement, and an EMG for muscle tone. Electrodes on the scalp, the face and the chin. That is what it takes to see REM directly, because REM is defined by brain and eye activity.
Your watch has an accelerometer, an optical heart-rate sensor, and sometimes a skin temperature sensor. It sees none of those signals. What it does is infer.
The inference
Movement tells it whether you are still. Heart rate and, more usefully, beat-to-beat variability tell it something about autonomic state: parasympathetic tone rises in deep sleep, and REM has a characteristic irregular pattern with occasional heart-rate surges. Combine motion, rate and variability across the night, and a classifier can guess at the stage.
Validation studies of consumer devices against polysomnography generally find good agreement on sleep versus wake — often around 90% — and considerably worse agreement on the individual stages. Deep sleep and REM are frequently confused with each other and with quiet wakefulness. A device can be right about how long you slept and quite wrong about how much of it was deep.
Which leads to the practical hierarchy.
What to trust, in order
Total sleep duration. Reliable enough to act on. If your watch says you have averaged 6h10m for three weeks, that is true, and it is the most actionable sleep fact you have.
Timing and regularity. Also reliable, and underrated. The variance in your sleep midpoint from night to night is one of the better-supported predictors of how you will feel, and it is measured from onset and wake times the device genuinely detects.
Nocturnal heart rate and HRV. Directly measured, not inferred. Resting heart rate during sleep is one of the cleanest physiological signals a consumer device produces, and an elevation of five or more beats above your baseline is a real finding — illness, alcohol, a hard session, or accumulated load.
Stage percentages. Read the trend, ignore the night. If your deep sleep reads 40 minutes tonight and 70 last night, that difference is well inside the classifier’s error. If your monthly average has halved, something has changed.
The trap
The trap is optimising the number that is least reliable. People read a low deep-sleep percentage and go looking for interventions, when the same device is simultaneously telling them, with far more confidence, that they slept six hours and went to bed at a different time every night.
Sleep debt and irregularity are measured well and matter enormously. Stage architecture is measured poorly and, for a healthy person, is largely not something you can steer directly anyway — it follows duration and timing.
Fix the two things the device measures properly. The third tends to follow.
Sleep Analytics reports the stage breakdown because it is there, and puts duration, regularity and nocturnal HRV in front of it because that is the order they deserve.