Truace
Subscribe

Machine learning trained on children's data detects sleep better than adult tools

Study title: The Sleep-Wake Classification Performance of Pediatric-Trained Machine Learning Algorithms for Raw Accelerometer Data

Researchers trained eight machine-learning models to tell sleep from wakefulness using raw wrist-worn accelerometer data from children, then compared them with existing methods. Sixty-five children aged 5.3 to 17.7 years wore a GENEActiv device during overnight lab polysomnography, and the physician-scored brain-wave recordings served as the reference.

Read the study abstract

Study objectives Actigraphy sleep-wake classification methods increasingly seek to leverage raw acceleration data and machine-learning-based classification, but performance evaluation in pediatrics is limited. We trained machine-learning models using pediatric data and compared their sleep-wake classification performance with existing algorithms for children. Methods Sixty-five children (46% female, ages 5.3-17.7 years) completed in-lab overnight polysomnography and wore a GENEActiv device on their non-dominant…

By the Truvace desk2 min read1 sourceEvidence: Moderate · 64/100 — Controlled or validated comparison · n = 65 · 1 independent sourcePeer-reviewedTRV-2026-2267· cite & verify
The Sleep-Wake Classification Performance of Pediatric-Trained Machine Learning Algorithms for Raw Accelerometer Data"Medical laboratories and clinics in Kinwat" by Mouryan, CC BY-SA 4.0.
Trace impact readingContested
P 57The P score combines the specificity and measured human impact of the grounded problem claim with the strength of this Trace’s cited sources.

Both sides are scored from claims and sources, not community votes.

G 56The G score combines the specificity and measured human impact of the grounded gain claim with the strength of this Trace’s cited sources.

In brief

The work matters because wearable sleep scoring has mostly been built and tested on adults, and testing in children has been limited. A pediatric-trained model that does better than adult-based tools could make sleep measurement in children more dependable. The roughly 25-minute average underestimate of sleep time shows the tool is not yet a perfect stand-in for lab testing.

Main points

  1. The local-global LSTM classifier ranked best, with balanced accuracy of 0.85, sensitivity of 0.88 and specificity of 0.83.
  2. Its ROC-AUC was 0.95 and its Cohen's kappa was 0.67, and it beat both an adult-trained random forest and GGIR scoring algorithms.
  3. Total sleep time was underestimated by an average of 25 minutes, with no proportional bias.
  4. The data covered 560.1 hours of recordings, and 74.4% of epochs were scored as sleep.

The gain

A machine-learning classifier trained on children's data scored sleep and wake more accurately than an adult-trained random forest and GGIR algorithms when checked against lab sleep recordings.

The problem

The best-performing pediatric classifier still underestimated total sleep duration by an average of 25 minutes per night.

The rundown

Accelerometer data were converted into 30-second epochs and aligned with physician-scored sleep-wake labels from electroencephalography. Eight models were trained with leave-one-subject-out cross-validation, so each child was tested on a model that had not seen their data.

Models were ranked by a Euclidean distance score that combined performance and bias, where a lower score means closer to perfect accuracy and zero bias. Epoch-by-epoch analysis produced the accuracy metrics, and discrepancy analysis estimated overall sleep duration bias.

For benchmarking, the team included GGIR sleep scoring algorithms and an adult-trained random forest classifier. The authors concluded that the pediatric classifiers had a strong ability to detect sleep and wake, with the LSTM the most optimal.

What this doesn’t fix

The evaluation involved only 65 children, one wrist-worn device model and in-lab overnight recordings, so performance in other devices, settings or populations is not shown by the study.

Sources

  1. Peer-reviewedSLEEP2026-10-24

The debate