Sleep detection
Wake, light and deep sleep through the night, from breathing and movement
Sleep detection turns a night in NAPPA into a picture of the infant's sleep: every 30 seconds classified as wake, light sleep or deep sleep, the standard 30-second epoch of clinical sleep scoring. It needs no electrodes and no sleep laboratory — only the movement sensor at the front of the diaper cover, which follows breathing and body movement through the night.
The result is a hypnogram, the night's sequence of sleep states, and a sleep depth trend that shows the same night as a continuous curve rather than a staircase (Figure 1A). From it come the totals a study reports: sleep time, wake time, time in light and deep sleep, and how the states cycle across the night (Figure 1B).
AHow a night becomes a hypnogram
BOne recording, as the classifier scored it
Show the totals for panel B
| State | Epochs | Time | Share of recording |
|---|---|---|---|
| Wake | 395 | 3 h 17.5 min | 35% |
| Light sleep (N1/REM) | 524 | 4 h 22 min | 47% |
| Deep sleep (N2/N3) | 203 | 1 h 41.5 min | 18% |
| Total | 1,122 | 9 h 21 min | 100% |
What you get
- Per night
- The hypnogram in 30-second epochs, the sleep depth trend, and the totals: sleep time, wake time, and time in light and deep sleep.
- Per child
- The same measures across nights, for following sleep over a study or comparing nights under different conditions.
- Alongside
- Breathing rate, body position, activity and arousals from the same night — see respiration, arousals and position.
What you need
- Garment
- NAPPA, tightened at the waist so the sensor moves with the belly.
- Sensor
- One Movesense sensor: Flash, MD or HR2. The features are computed in the sensor, so no streaming is needed with any of them.
- Recording
- A night's sleep. The published studies logged a 12-hour window, typically 7 pm to 7 am.
How it is derived
The classifier was trained on nights recorded with NAPPA during clinically indicated polysomnography in 33 infants aged 2 weeks to 18 months, at the sleep laboratory of the Children's Hospital in Helsinki. A clinical expert, unaware of the study, scored every 30-second epoch of the polysomnograms by the American Academy of Sleep Medicine rules, and those scores were the training target: 16,649 epochs in all. Performance was tested by leaving each infant out in turn (Heliyon 2024).
The model is a small bidirectional recurrent neural network — two layers of gated recurrent units — working on five features that the sensor computes itself: activity from the accelerometer, and breathing rate, regularity and the size and variability of the breathing movement from the gyroscope, which picks up the belly's breathing movement more reliably than the accelerometer. Adding the infant's age, or correcting breathing rate for age, did not change the result. The classifier is published as open source.
Why three states and not five? Telling REM sleep from N1, or N2 from N3, needs signals a diaper cover cannot record: eye movements and muscle tone for the first, brain activity for the second. Grouping them into light and deep sleep keeps the distinctions the signals can actually support.
Against polysomnography
Across the 33 infants the classifier agreed with the polysomnography scoring at a median of 78% of epochs (interquartile range 74–83%; Matthews correlation 0.65). Deep sleep was recognised best (82%), light sleep and wake at 76% and 75%. Most disagreements fell between neighbouring states — 21% of wake epochs were scored as light sleep — and many sat at the boundaries, where the classifier shows a gradual change and the clinical hypnogram a step (Heliyon 2024).
For comparison, two trained experts scoring the same polysomnogram with its full set of brain, eye and muscle signals typically agree on 60–80% of epochs.
The sleep depth trend
Sleep does not switch between states the way a hypnogram draws it. The sleep depth trend combines the classifier's state and its confidence into one continuous curve, so the gradual descent into deep sleep, the cycling across the night and the brief arousals are visible at a glance. It also carries the classifier's confidence, which a staircase hypnogram hides.
The training data came mostly from young infants, so there are no normative reference values for age yet; the measures are best used to compare nights, conditions or groups within a study.
Available on
Publications
No publication here covers all of the selected products.