Paperdoi:10.1093/sleep/zsag282
Sleep–wake classification performance of pediatric-trained machine learning algorithms for raw accelerometer data (SLEEP 2026)
SLEEP, zsag282 (advance article, 2026-10-24)
Licence: CC BY-NC 4.0
TL;DR
65 children (5.3–17.7 y, 46 % female) had in-lab overnight PSG while wearing a GENEActiv on the non-dominant wrist (560.1 h; 74.4 % of epochs sleep). Eight ML models were trained on 30-s epochs with leave-one-subject-out CV and benchmarked against GGIR algorithms and an adult-trained random forest. The pediatric-trained local-global LSTM was best: BA 0.85, sensitivity 0.88, specificity 0.83, ROC-AUC 0.95, κ 0.67. It underestimates total sleep by about 25 min on average, with no proportional bias.
- Form factor: wrist (GENEActiv, non-dominant)
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 wrist. The acceleration data were converted into 30-second epochs and aligned with physician-scored sleep-wake data from electroencephalography. Eight machine-learning models were trained using leave-one-subject-out cross-validation. Epoch-by-epoch analyses generated performance metrics (e.g., balanced accuracy [BA]) and discrepancy analyses provided overall sleep duration bias estimates. The combination of highest performance and least bias was used to rank using Euclidean distance scores - where a lower score represents closer to perfect performance and zero bias. For benchmarking, we included GGIR sleep scoring algorithms and an adult trained random forest classifier.
Results
Overall, 560.1 hours of polysomnography and actigraphy data were collected (74.4% of epochs were scored as sleep). The pediatric-trained local-global long-short term memory (LSTM) classifier had the most optimal epoch-by-epoch performance (e.g., BA=0.85, sensitivity=0.88, specificity=0.83, ROC-AUC=0.95, and Cohen’s kappa=0.67). These metrics exceeded that of an adult-trained random forest classifier and GGIR-based algorithms. Discrepancy analyses revealed that overall sleep duration was underestimated by an average of 25 minutes using the LSTM classifier with no proportional bias.
Conclusions
We trained eight pediatric sleep-wake classifiers that had a strong ability to detect sleep and wake, with the LSTM classifier being most optimal.
Verbatim from the publisher's DOI record (opens in a new tab) · Publisher licence: https://creativecommons.org/licenses/by-nc/4.0/
Results on ChatPPG leaderboards
All leaderboardsNo results from this paper on ChatPPG leaderboards yet. Browse the leaderboards.
Tasks & signals
Tasks
Signals
Keywords
How it works
Cite this paper
APA
Chen, P.-W., Cielo, C. M., Walch, O., McDonald, M., Song, P. X. K., Goldstein, C. A., Moreno, J. P., Jansen, E. C., & Mitchell, J. A. (2026). Sleep–wake classification performance of pediatric-trained machine learning algorithms for raw accelerometer data (SLEEP 2026). SLEEP, zsag282. https://doi.org/10.1093/sleep/zsag282
BibTeX
@article{chen2026sleepwake,
title = {{Sleep–wake classification performance of pediatric-trained machine learning algorithms for raw accelerometer data (SLEEP 2026)}},
author = {Chen, Pin-Wei and Cielo, Christopher M. and Walch, Olivia and McDonald, Morgan and Song, Peter X. K. and Goldstein, Cathy A. and Moreno, Jennette Palcic and Jansen, Erica C. and Mitchell, Jonathan A.},
year = {2026},
journal = {SLEEP, zsag282},
doi = {10.1093/sleep/zsag282}
}Built from the hub record. Check it against the publisher before submitting.
Added to the ChatPPG research hub
