What estimated sleep stages really mean

Every sleep app draws you a chart with coloured bands for light, deep and REM, and almost none of them tell you where those bands came from. On a phone only night, what you are looking at are estimated sleep stages, not measured ones, and the difference is not a technicality. It sets a ceiling on the accuracy you should read into that chart, and it changes which questions the chart can answer at all.

What a real stage measurement involves

The reference standard is polysomnography, the overnight study a sleep lab runs. It scores your night from three signals at once: brain activity from EEG electrodes on the scalp, eye movement from electrodes at the corners of the eyes, and muscle tone from an electrode under the chin. A trained scorer then works through the night in thirty second windows and assigns each one a stage.

Those three signals are not incidental. They are the definition. Deep sleep is defined by slow, high amplitude waves across a certain proportion of the window. REM sleep is defined by the combination of a low amplitude mixed frequency trace, bursts of rapid eye movement, and muscle tone dropping to its lowest point of the night. Those are the criteria. Not “you were quite still” or “your breathing evened out”.

A phone on your nightstand has none of the three. No electrodes on your scalp, no view of your eyes, no read on your chin muscles. It cannot measure a sleep stage, and no amount of processing conjures a signal that was never captured.

Where estimated sleep stages are accurate, and where they are not

What a phone does have is movement and sound. Tock reads both: the accelerometer for the movement half, which is the phone version of actigraphy, and the microphone for breathing sounds, snores, coughs and room noise.

From those, a model infers the shape of a night. Long stillness with slow, even breathing sounds is the kind of thing that tends to accompany deeper sleep. Brief small movements with more variable breathing is the kind of thing that tends to accompany lighter stages and REM. The model uses the tendency, and the tendency is real, which is why the estimate is better than nothing. It is also just a tendency, which is why it is an estimate.

Here is where that lands on accuracy, honestly.

Sleep versus wake is where phone tracking is strongest. Was I asleep, roughly when did I drop off, roughly when did I surface, how long was that gap in the middle. Those are questions about stillness and noise, and stillness and noise are exactly what the sensors capture.

Splitting light from deep from REM is where it is weakest. Two stages that look completely different on an EEG can look nearly identical from the mattress. The single largest failure mode of any movement based approach is the same one it has always had: lying awake perfectly still reads as sleep, because nothing moved and nothing made a sound.

So the bands on the chart are a plausible model of a night that produced your movement and audio, not a reading of the night you had. Tock labels them as estimates in the interface for that reason, rather than hoping you assume it.

When they stop being estimates

If you wear something on your wrist, the picture improves, and Tock does not need to be the thing reading it. When Apple Health has stage data in it from a wearable, Tock uses that instead of its own estimate, and the night detail shows measured stages alongside the vitals that came with them: heart rate, resting heart rate, HRV, blood oxygen, respiratory rate, wrist temperature.

That is still not a lab. A wrist device is inferring stages too, mostly from heart rate variability and movement, and the honest description of consumer wearables is that they are better at this than a phone and worse at it than electrodes. But it is a real second signal from the body rather than a proxy read through a mattress, and it is a step up.

There is no Tock app on the Watch. Tock reads your Watch’s data through Apple Health, which is the same route any other wearable takes to get there.

How to read an estimated hypnogram without fooling yourself

A hypnogram is the stepped chart of stages across a night, and there is a right and a wrong way to use an estimated one.

The wrong way is to treat a single night’s deep sleep number as a score. It is the most tempting number on the screen, it is the least reliable one a phone produces, and comparing it against a friend’s app is comparing two different models of two different nights. Nobody learns anything from that except how to feel bad.

The right way is longitudinal. Over three or four weeks your own estimate is consistent with itself, because it is the same model reading the same sensors in the same bedroom. That makes direction of change meaningful even when the absolute numbers are soft. If the estimated deep sleep block that usually sits early in your night keeps shrinking across a fortnight, something changed, and the estimate has done its job by pointing at it. Chasing yesterday’s number is noise. Watching a month bend is signal.

The other thing worth trusting is shape rather than totals. Sleep architecture has a normal profile, with deeper sleep loaded toward the front of the night and REM getting longer toward the morning, and a phone estimate reproduces that overall arc reasonably well even when it misplaces individual blocks. If your chart is fragmented into confetti every night, that is a pattern to take seriously regardless of what the specific labels say.

For the underlying biology, sleep cycles explained covers what actually happens across a night and why the ninety minute figure everyone quotes is an average rather than a rule. For where phone tracking sits against a wrist device on everything else, sleep tracking without a wearable has the comparison in full.