CoChat

Self-supervised learning of accelerometer data provides new insights for sleep and its association with mortality

Open access

Published version of a preprint

The DOI matches its Crossref record; 37 of the 37 references it cites were found in a registry or an index; 4 of the 24 sources its claims cite were checked for support.

Abstract

Sleep is essential to life. Accurate measurement and classification of sleep/wake and sleep stages is important in clinical studies for sleep disorder diagnoses and in the interpretation of data from consumer devices for monitoring physical and mental well-being. Existing non-polysomnography sleep classification techniques mainly rely on heuristic methods developed in relatively small cohorts. Thus, we aimed to establish the accuracy of wrist-worn accelerometers for sleep stage classification and subsequently describe the association between sleep duration and efficiency (proportion of total time asleep when in bed) with mortality outcomes. We developed a self-supervised deep neural network for sleep stage classification using concurrent laboratory-based polysomnography and accelerometry. After exclusion, 1113 participant nights of data were used for training. The difference between polysomnography and the model classifications on the external validation was 48.2 min (95% limits of agreement (LoA): −50.3 to 146.8 min) for total sleep duration, −17.1 min for REM duration (95% LoA: −56.7 to 91.0 min) and 31.1 min (95% LoA: −67.3 to 129.5 min) for NREM duration. The sleep classifier was deployed in the UK Biobank with ~100,000 participants to study the association of sleep duration and sleep efficiency with all-cause mortality. Among 66,262 UK Biobank participants, 1644 mortality events were observed. Short sleepers (<6 h) had a higher risk of mortality compared to participants with normal sleep duration 6–7.9 h, regardless of whether they had low sleep efficiency (Hazard ratios (HRs): 1.36; 95% confidence intervals (CIs): 1.18 to 1.58) or high sleep efficiency (HRs: 1.29; 95% CIs: 1.04–1.61). Deep-learning-based sleep classification using accelerometers has a fair to moderate agreement with polysomnography. Our findings suggest that having short overnight sleep confers mortality risk irrespective of sleep continuity.

Overview

CoChat Summary

TL;DR

The authors developed SleepNet, a self-supervised deep recurrent neural network that classifies wrist accelerometer data into wake, REM, and NREM sleep and validated it against laboratory polysomnography. Deployed in 66,262 UK Biobank participants, short overnight sleep (<6 h) was associated with higher all-cause mortality regardless of sleep efficiency.

Key findings

  • SleepNet achieved internal-validation F1 of 0.75 ± 0.1 for sleep/wake and 0.57 ± 0.11 for wake/REM/NREM; external-validation F1 was 0.66 ± 0.12 and 0.49 ± 0.10, respectively. Three-class epoch-level Kappa was 0.39 internally and 0.32 externally.
  • Against polysomnography, external-validation mean bias was 48.2 min for total sleep duration (95% LoA: −50.3 to 146.8 min), −17.1 min for REM duration (95% LoA: −56.7 to 91.0 min), and 31.1 min for NREM duration (95% LoA: −67.3 to 129.5 min). Internal-validation biases were 9.9 min, −24.4 min, and 34.4 min, respectively.
  • In 66,262 UK Biobank participants with 1,644 mortality events over 452,652 follow-up years, short sleepers (<6 h) had higher mortality risk than those sleeping 6–7.9 h, both with low sleep efficiency (HR 1.36; 95% CI 1.18–1.58) and high sleep efficiency (HR 1.29; 95% CI 1.04–1.61).
  • Overall UK Biobank sleep parameters were overnight sleep 6.8 ± 0.9 h/day, NREM 5.3 ± 0.9 h/day, REM 1.5 ± 0.6 h/day, and sleep efficiency 81.5 ± 8.6%. Longer overnight sleep was not associated with higher mortality risk than the reference group.

Methods

  • SleepNet used a ResNet-17 V2 1D convolutional feature extractor, a bidirectional LSTM, and fully connected layers to classify each 30-s accelerometry window. It was pre-trained by multi-task self-supervision on 700,000 person-days of unlabelled UK Biobank accelerometry, then fine-tuned using concurrent laboratory polysomnography as ground truth; 1,157,913 (~10,000 h) sleep windows were used to train the network.
  • Internal validation used subject-wise five-fold cross-validation in Raine Gen1, Raine Gen2, and Newcastle cohorts (1,113 participants/nights after preprocessing); external validation used Leicester and Pennsylvania cohorts (53 participants/nights). Performance was compared with a random forest model using hand-crafted spatiotemporal features.
  • For health association analysis, SleepNet was deployed in the UK Biobank using sleep windows detected by a random forest model; after exclusions, 66,262 participants were analyzed. Cox proportional hazards regression used age as the timescale and adjusted for sex, ethnicity, Townsend Deprivation Index, educational qualifications, smoking status, alcohol consumption, and overall activity.

Limitations

  • Stated: validation datasets mainly consisted of healthy populations from Caucasian ethnic backgrounds; validation in chronic disease and different ethnic populations is needed. Laboratory polysomnography may suffer from the first-night effect, and future validation could assess within-person variability using multi-night polysomnography. Sleep stage identification from actigraphy is highly challenging, especially for wake periods in bed without wrist movement.
  • Stated: the analysis did not include daytime naps in the measurement of overnight sleep duration, which may explain why long overnight sleep was not associated with increased mortality risk.
  • Evident from the text: external validation was much smaller than internal validation (53 vs. 1,113 participants/nights) and showed lower agreement, with the model tending to underestimate REM and short sleep and overestimate NREM and long sleep.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1038/s41746-024-01065-0

Cites

37 of 37 references verified0 citation errors · 0 not checkable

Statistics

Recomputing p-values needs the paper’s full text.

Figures

Fig. 1 The SleepNet development pipeline. 1. We use multi-task self-supervised learning to obtain a feature extractor by learning from 700,000 person-days of tri-axial accelerometry data in the UK Biobank. 2. The pre-trained feature extractor was then fine-tuned with a deep recurrent network to train a sleep-stage classifier using polysomnography as the ground truth. 3. We deploy the sleep prediction model on the UK Biobank and investigate the association between device-measured sleep and mortality outcomes.
Fig. 1 The SleepNet development pipeline. 1. We use multi-task self-supervised learning to obtain a feature extractor by learning from 700,000 person-days of tri-axial accelerometry data in the UK Biobank. 2. The pre-trained feature extractor was then fine-tuned with a deep recurrent network to train a sleep-stage classifier using polysomnography as the ground truth. 3. We deploy the sleep prediction model on the UK Biobank and investigate the association between device-measured sleep and mortality outcomes.CC BY · PubMed Central
Fig. 2 Agreement assessment via Bland-Atman plot for total sleep duration, rapid eye movement sleep (REM) duration and non-rapid eye movement sleep (NREM) duration on internal and external validation. a the agreement for the internal validation; b the agreement for the external validation. The internal validation consists of 1113 polysomnography nights from the Raine Study and the Newcastle cohort, whereas the external validation consists of 53 polysomnography nights from the Leicester and Pennsylvania cohorts.
Fig. 2 Agreement assessment via Bland-Atman plot for total sleep duration, rapid eye movement sleep (REM) duration and non-rapid eye movement sleep (NREM) duration on internal and external validation. a the agreement for the internal validation; b the agreement for the external validation. The internal validation consists of 1113 polysomnography nights from the Raine Study and the Newcastle cohort, whereas the external validation consists of 53 polysomnography nights from the Leicester and Pennsylvania cohorts.CC BY · PubMed Central
Fig. 3 Device-measured sleep probability trajectories throughout the day for the UK Biobank participants. Top: variations of the average overnight sleep probability for the participants with self-reported ‘morning’ and ‘evening’ chronotype (a) and the overnight sleep distributions across thirds of device-measured physical activity level (b). Bottom: variations of the average REM (c) and NREM (d) probability in participants with a history of self-reported insomnia symptoms versus those without. REM rapid-eye-movement sleep, NREM non-rapid-eye-movement sleep.
Fig. 3 Device-measured sleep probability trajectories throughout the day for the UK Biobank participants. Top: variations of the average overnight sleep probability for the participants with self-reported ‘morning’ and ‘evening’ chronotype (a) and the overnight sleep distributions across thirds of device-measured physical activity level (b). Bottom: variations of the average REM (c) and NREM (d) probability in participants with a history of self-reported insomnia symptoms versus those without. REM rapid-eye-movement sleep, NREM non-rapid-eye-movement sleep.CC BY · PubMed Central
Fig. 4 Associations of overnight sleep duration with all-cause mortality for groups with low and high sleep efficiency. a participants with less than 85% sleep efficiency; b participants with greater or equal to 85% sleep efficiency. The model used 1644 events among 66,262 participants. We used age as the timescale and adjusted for sex, ethnicity, Townsend Deprivation Index of baseline address (split by quarter in the study population), educational qualifications, smoking status, alcohol consumption (Never, <3 times/week, 3+ times/week), overall activity (measured in milli-gravity units). The median was used to separate groups with low and high sleep efficiency. Areas of squares represent the inverse of the variance of the log risk. The I bars denote the 95% confidence interval for the floated risks.
Fig. 4 Associations of overnight sleep duration with all-cause mortality for groups with low and high sleep efficiency. a participants with less than 85% sleep efficiency; b participants with greater or equal to 85% sleep efficiency. The model used 1644 events among 66,262 participants. We used age as the timescale and adjusted for sex, ethnicity, Townsend Deprivation Index of baseline address (split by quarter in the study population), educational qualifications, smoking status, alcohol consumption (Never, <3 times/week, 3+ times/week), overall activity (measured in milli-gravity units). The median was used to separate groups with low and high sleep efficiency. Areas of squares represent the inverse of the variance of the log risk. The I bars denote the 95% confidence interval for the floated risks.CC BY · PubMed Central

Verification

Claims the paper makes

4 Valid · 0 Not valid · 12 Not checked · 62 Own results and interpretations

41 more claims

Create a free account to see all 78 claims and to verify the 20 references still open.

See all claims — free account
  • Probably not a substantial claim. 24
  • It may state two separate claims as one. 14

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

3 of 3

Citations

Cites

Show all 37 references

Research works printed but not linked: 1

  1. Huang, Z., Xu, W. & Yu, K. Bidirectional lstm-crf models for sequence tagging. arXiv preprint arXiv:1508.01991 (2015). · arXiv

Cited by

See all 54

Metadata

Authors
Yuan, H., Plekhanova, T., Walmsley, R., Reynolds, A. C., Maddison, K. J., Bucan, M., Gehrman, P., Rowlands, A., Ray, D. W., Bennett, D., McVeigh, J., Straker, L., Eastwood, P., Kyle, S. D., & Doherty, A.
Affiliations
  • University of Oxford (GB)
  • University of Leicester (GB)
  • Flinders University (AU)
  • The University of Western Australia (AU)
  • University of Pennsylvania (US)
  • John Radcliffe Hospital (GB)
  • Curtin University (AU)
  • Murdoch University (AU)
Corresponding author
Doherty, Aiden
Venue
npj Digital Medicine
Published
2024-06-30
Volume
7
Issue
1
Pages
86-86
ISSN
2398-6352 (ISSN-L)
OpenAlex
W4398139961
Cited by
40
References
36
Citations by year
2026: 17 · 2025: 15 · 2024: 8
FWCI
16.68
Type
journal article
Language
en
Open access
Licence
Locations
  • npj Digital Medicine (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • Murdoch University Research Portal (submittedVersion, cc-by)
  • UWA Profiles and Research Repository (UWA) (submittedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
Funders
  • Wellcome Trust (223100/Z/21/Z, 227093/Z/23/Z)
  • Novo Nordisk
  • British Heart Foundation (RE/18/3/34214)
  • Swiss Re | Swiss Re Foundation
  • DH | National Institute for Health Research (NIHR203316)
  • U.S. Department of Health & Human Services | NIH | National Institute of Mental Health
Topics
  • Social Sciences › Psychology › Experimental and Cognitive Psychology › Sleep and related disorders
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Context-Aware Activity Recognition Systems
  • Life Sciences › Neuroscience › Endocrine and Autonomic Systems › Circadian rhythm and melatonin
Keywords
sleep stage classification, wrist-worn accelerometers, self-supervised deep learning, polysomnography validation, sleep duration, sleep efficiency, all-cause mortality, short sleepers, REM sleep, NREM sleep, UK Biobank, wearable devices

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-05, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-05 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol and qwen3.8-omni-flash).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Self-supervised learning of accelerometer data provides new insights for sleep and its association with mortality. https://app.cochat.ai/papers/10.1038/s41746-024-01065-0
The paper
Yuan, H., Plekhanova, T., Walmsley, R., Reynolds, A. C., Maddison, K. J., Bucan, M., Gehrman, P., Rowlands, A., Ray, D. W., Bennett, D., McVeigh, J., Straker, L., Eastwood, P., Kyle, S. D., & Doherty, A. (2024). Self-supervised learning of accelerometer data provides new insights for sleep and its association with mortality. npj Digital Medicine. https://doi.org/10.1038/s41746-024-01065-0

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.