Recovery
How Wearables Estimate Sleep And Where They Go Wrong
Consumer sleep trackers infer sleep from movement and pulse rather than measuring brain activity, which sets a hard ceiling on how accurately they can report sleep stages.
Sleep trackers report a number of hours and a breakdown into stages, and both are inferences. Understanding what the device actually senses explains which parts of the output are worth attention.
What the sensors actually detect
A wrist or ring device typically contains an accelerometer for movement and an optical sensor that measures blood volume changes at the skin to derive pulse rate.
Some add skin temperature and blood oxygen estimates. None of them measure brain activity, which is what sleep staging is clinically defined by.
The device therefore produces stages by feeding movement and cardiac data into a model trained to predict what a laboratory would have scored.
Why sleep and wake are the easy part
Distinguishing sleep from wakefulness using movement alone works reasonably well, because people move considerably more awake than asleep.
The characteristic failure is lying still while awake, which the movement signal reads as sleep. This is exactly what people with insomnia do, so the devices tend to overestimate their sleep.
Total sleep time is nonetheless the most trustworthy figure such a device produces, and it is the figure with the clearest connection to how a person functions.
Why staging is the weak part
Deep and rapid-eye-movement sleep are defined by patterns in brain electrical activity that have no reliable external signature a wrist sensor can read.
The models exploit correlations, since heart rate variability and movement do differ between stages, but the relationship is loose and varies between individuals.
Comparisons against laboratory scoring generally find stage-level agreement well below what the confident presentation in an app implies.
The readiness score problem
Composite scores combine sleep estimates with resting heart rate and heart rate variability into a single figure describing how prepared someone is.
The inputs are noisy, the weighting is proprietary, and the underlying measures vary day to day for reasons unrelated to training, including position, room temperature and hydration.
Treating a low score as an instruction is therefore acting on a number whose construction is unpublished and whose components are individually unreliable.
Using the output for what it can support
Trends across weeks are more defensible than any single night, because averaging reduces the noise that dominates individual readings.
Bedtime and wake time are recorded rather than inferred, which makes schedule regularity one of the more genuinely useful things these devices track.
Suspected sleep disorders are diagnosed with clinical testing, not with consumer devices, and persistent poor sleep is a reason to see a physician regardless of what an app reports.
