Measuring the nap#
Two signals, and a hypnogram that implies more
Where we have been, in one line#
A nine-point scale and a reaction-time task — while most of you wear a device that claims to do better.
This week is about whether it does. Nearly every popular claim about sleep now rests on data of this kind.
The reference standard#
Polysomnography defines the stages: brain activity, eye movement, muscle tone.
- needs a lab, electrodes and a scorer
- everything else is an approximation of it
The size of that approximation is the lecture.
What each one actually has#
| Signal | Polysomnography | Wrist device |
|---|---|---|
| brain activity | yes | no |
| eye movement | yes | no |
| muscle tone | yes | no |
| movement | — | yes |
| heart rate, optical | — | usually |
That is the whole input on the right.
From movement to a hypnogram#
Actigraphy separates sleep from wake reasonably well — sleeping people move less.
Stages are a different job: inferred from those two signals by a proprietary model.
The confidence in the display comes from the inference, not from anything observed.
The known failure mode#
Lying still while awake reads as sleep.
- worst for the people most worried about their sleep
- their devices reassure them precisely when they should not
Below the resolution#
A 20-minute nap is nearly invisible to the instrument
Why that lands here#
Short sleep is the hard case: a nap of that length sits near or below what these systems can resolve.
The course’s central object is the thing the popular instrument sees least well.
Validation, and its absence#
- some devices publish agreement studies against polysomnography; most do not
- agreement on stages is consistently worse than on total sleep time
- algorithms change in firmware updates
So last year’s validation may not describe this year’s device.
Why the number looks so certain#
Clean coloured bands are a design decision.
The estimate underneath carries uncertainty the display does not show.
Showing it would make the product worse and the data better.
Tell me where this is unfair#
Your sleep tracker is a reasonably good movement sensor wrapped in a presentation layer that implies a measurement it did not take.