Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

A single deep network scores sleep from any EEG/EOG channel combination at adjustable temporal resolutions, from 30-second epochs down to sub-0.01 seconds, and extracts short wake events plus biomarker-relevant dynamics from the same pass.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

AnySleep is a channel-agnostic sleep-staging model that matches or exceeds U-Sleep at 30-s epochs and derives sub-30-s dynamics that help detect arousals and predict subject characteristics.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection Solid channel-agnostic 30-s sleep staging on ~20k recordings; the sub-30-s biomarker claims are suggestive but under-controlled, and the metadata abstract oversells them. the 4 major comments →

arxiv 2512.14461 v2 pith:RF2MQGT3 submitted 2025-12-16 cs.LG eess.SPq-bio.QM

AnySleep: a channel-agnostic deep learning system for high-resolution sleep staging in multi-center cohorts

Niklas Grieger , Jannik Raskob , Siamak Mehrkanoon , Stephan Bialonski This is my paper
classification cs.LG eess.SPq-bio.QM
keywords sleep stagingdeep learningpolysomnographyEEG/EOGchannel-agnostichigh-resolution sleep stagingarousal detectionmulti-center generalization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper introduces AnySleep, a deep network that scores sleep from arbitrary EEG or EOG channel combinations and outputs sleep stages at resolutions from 30 seconds down to roughly 8 milliseconds. Its central claim is that a model trained only on conventional 30-second expert labels can, by swapping the final temporal pooling at inference, produce high-frequency predictions that recover brief wake intrusions (arousals) that the 30-second grid misses, and that these fine-grained predictions improve prediction of age, sex, and sleep apnea status. Trained on roughly 20,000 overnight recordings from 21 datasets, the model matches or exceeds the established baseline at conventional 30-second staging while remaining robust to missing channels, EOG absence, and single-channel input. If true, this removes two long-standing barriers for large-scale sleep research: montage heterogeneity and the coarse 30-second scoring grid.

Core claim

AnySleep demonstrates that one encoder-decoder model with channel-attention modules at multiple depths can accept any set of EEG/EOG channels and, by changing only the kernel and stride of the final temporal average-pooling layer, produce sleep-stage predictions at any of 14 resolutions from 30 seconds down to 0.008 seconds. The paper reports macro-F1 scores of roughly 0.78-0.81 on held-out multi-center test sets, on par with or above the fixed-input baseline, and shows that at 2-8 second resolution the model detects roughly 53% of expert-annotated arousals (IoU F1 0.442) and that triplet features from sub-30-second stages outperform 30-second features for age, sex, and OSA prediction.

What carries the argument

The channel-attention module is the carrier of channel-agnostic behavior: each input channel's feature map is averaged over time, passed through a small network to produce a scalar, and softmax-normalized across channels to form a weighted sum, letting the model combine any number of channels at 13 depths of the encoder-decoder. The second key mechanism is the configurable temporal pooling in the segment classifier, kept at 30 seconds during training but varied at inference to generate multi-scale predictions from identical weights.

Load-bearing premise

The high-frequency predictions are treated as meaningful sleep stages even though the model was trained only on 30-second labels, and the paper itself says the absolute accuracy of high-frequency predictions remains uncertain.

What would settle it

Take a held-out cohort, have multiple experts annotate sleep stages at 5-s or 2-s resolution, and compare AnySleep's predictions at matching resolution against the consensus; if agreement is near chance or no better than resampling the original 30-s labels, the central high-frequency claim fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Multi-center studies can pool recordings with heterogeneous electrode montages without channel-exclusion or manual channel mapping, since performance stays high even with a single EEG channel or EOG only.
  • Sleep staging can be harmonized at 30-second epochs across sites: the model matches or beats the established baseline on held-out datasets and improves as more channels are added.
  • Short sleep events invisible to the 30-second grid, such as arousals, become detectable from wake predictions at 2-8 second resolution, enabling scalable arousal-like annotation in large cohorts.
  • Fine-grained sleep-stage sequences carry information about subject-level characteristics: triplet features from sub-30-second predictions improve age, sex, and sleep apnea classification compared to 30-second staging.
  • Scaling to many channels remains computationally reasonable because channel handling grows linearly with channel count, unlike baseline pair-wise majority-voting schemes that grow quadratically.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The most consequential untested implication is that the sub-30-second predictions are true physiological transitions rather than artifacts of sharpening the 30-second boundary; existing 5-second inter-rater studies suggest human agreement at that scale is low, so the model's high-frequency outputs will likely need consensus labels to be confirmed.
  • A natural next step the paper does not take is to use AnySleep's high-frequency predictions as pseudo-labels to fine-tune a model directly at 2-8 s resolution, which could improve fidelity beyond the re-pooling trick and make short-epoch training feasible.
  • Because the model's high-frequency predictions at 0.05-0.5 s best predict age, sex, and OSA, sleep fragmentation metrics could be redefined in a resolution-continuous way rather than via 30-second epoch transitions, a change that would affect how sleep continuity is quantified in clinical studies.
  • The channel-attention weights, though variable across training runs, offer a per-recording view of which channels matter at which abstraction depth; a future extension is to use these weights as interpretability tools for channel-optimized wearable sleep monitoring.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper introduces AnySleep, a U-Net-style encoder-decoder with channel-attention modules that accepts arbitrary EEG/EOG channel configurations and outputs sleep stages at adjustable temporal resolutions. The model is trained on approximately 19,900 recordings from 21 datasets using 30-s expert labels, with stochastic channel subsampling during training. At 30-s resolution, AnySleep matches or exceeds a retrained U-Sleep baseline across in-distribution and held-out datasets and remains stable under single- and two-channel inputs. The central additional claim is that, by changing the average-pooling window of the segment classifier at inference, the same model produces high-frequency sleep-stage predictions that capture short wake intrusions consistent with expert-annotated arousals and yield triplet-based features that improve prediction of age, sex, and sleep apnea status compared with 30-s staging.

Significance. If the sub-30-s claims hold, AnySleep would be a valuable contribution: a single channel-agnostic model that provides both conventional 30-s staging and multi-scale sleep dynamics from the same forward pass, with a large multi-center training set, external hold-out evaluation, three training seeds, a retrained U-Sleep baseline, and public code/model release. The 30-s channel-robustness results are well supported by the evaluation design and are a useful practical contribution for harmonized multi-center studies. The high-frequency and biomarker claims are, however, less secure: they rest on indirect proxies, and the paper itself acknowledges that the absolute accuracy of high-frequency predictions remains uncertain. Given that these claims appear in the abstract and are presented as headline results, they require stronger validation or substantial tempering.

major comments (4)
  1. [§2.3, §4.5] The central sub-30-s claim — that AnySleep 'captures short wake intrusions consistent with arousals' — is based on predictions produced by changing the pooling window at inference of a model trained only on 30-s labels. The paper itself states in the Discussion that 'the absolute accuracy of high-frequency predictions remains uncertain.' The arousal validation reports a maximum IoU F1 of 0.442 and an overlap of 57.7%, but no null-model or permutation control is provided to show that these values exceed what would be obtained by, e.g., expanding 30-s Wake labels to high resolution or by shuffling sub-epoch predictions within each 30-s epoch. Without such a control, the high-frequency predictions may be classifier artifacts rather than genuine micro-architectural events. Please add a control or direct validation on available short-epoch annotations (e.g., 5-s scored data) and temper the ab
  2. [§2.3, §4.5, Fig. 3] The triplet features used for age, sex, and OSA prediction are defined as absolute counts of stage triplets per 1.5-hour block. As the temporal resolution increases, the number of time steps per block increases by up to a factor of 3840, so the raw triplet counts scale up mechanically, independent of any true temporal structure. The improvements shown in Fig. 3 could therefore reflect count magnitude or noise-driven transitions rather than physiologically meaningful fine-grained dynamics. Please normalize counts by the number of windows, or include an explicit control that preserves the 30-s label sequence but randomizes the ordering of sub-epoch predictions, and report the effect of such a control on the RF performance. Without this, the claim that high-frequency staging 'improves prediction of pathophysiological conditions' is not established.
  3. [§2.3, Fig. 3] The best resolutions for arousal detection (2–8 s) and for age/sex/OSA prediction (0.05–0.5 s) are selected post hoc across 14 tested resolutions, with no correction for multiple comparisons or a held-out resolution-selection procedure. Since the abstract reports improvements at these chosen scales, please either pre-specify the resolution, apply a multiple-comparison correction, or demonstrate that the performance trend is consistent across a range of resolutions rather than a single peak that could arise by chance.
  4. [Abstract vs. Full Text] There is a serious inconsistency between the abstract reproduced at the top of the manuscript and the full-text abstract/Results. The top abstract states 'over 20,000 overnight recordings ... from 28 datasets' and claims improved prediction of 'obstructive sleep apnea, narcolepsy type 1, insomnia,' while the full-text abstract and Results report 19,909 recordings from 21 datasets and only evaluate age, sex, and sleep apnea status. Please reconcile these statements and ensure the abstract accurately reflects the datasets and outcomes actually analyzed.
minor comments (5)
  1. [§4.5] The list of 14 resolutions (1, 2, 4, 8, 16, 32, 64, 128, 256, 384, 640, 960, 1920, 3840) is correct but could be formatted more clearly as predictions per 30-s epoch, perhaps with corresponding time-step durations in seconds.
  2. [§2.2, Fig. 1] The caption for Fig. 1 states 'recordings in the reference montage that contained all investigated channels (563 recordings)' and later reports 643 recordings for the channel-count experiment; please verify these numbers are consistent with the dataset splits described in Table 2.
  3. [§2.3] The text 'around 23:12:05 and 23:13:30' in the description of Fig. 2a is vague; please specify the exact time axis or label the relevant intervals in the figure.
  4. [§4.3] The training section mentions that preliminary experiments showed that U-Sleep-style masking augmentation did not improve performance, but no details or quantitative results are given. A brief supplementary table or description would aid reproducibility.
  5. [§4.5] The random forest hyperparameter sampling ranges are given, but the number of trees is not specified. Please report the n_estimators value used in the scikit-learn models.

Circularity Check

0 steps flagged

No significant circularity: the 30-s benchmark is external and the sub-30-s claims are explicitly presented as indirect evidence with acknowledged uncertainty.

full rationale

The paper's central derivation chain is not circular. AnySleep is trained on 30-s expert sleep-stage labels and its headline 30-s result is validated against held-out datasets and a retrained U-Sleep baseline; this is an external benchmark, not a quantity defined in terms of the model's own predictions. The sub-30-s analyses change only the pooling kernel/stride of the trained backbone and then use the resulting predictions to detect arousals and to predict age, sex, and OSA status. Although these analyses are self-referential in the sense that the model's own outputs are the objects being studied, the targets (expert arousal annotations and subject-level demographic/clinical labels) are external, and the triplet features are not fitted to those targets. The paper explicitly acknowledges the validation gap: 'the absolute accuracy of high-frequency predictions remains uncertain' and 'high-frequency predictions were assessed indirectly.' This is an honest limitation and a correctness risk, not a circular derivation. The methodological borrowings from Perslev et al. [17] are citations to prior external work, not a self-citation chain, and no equation or fitted parameter reduces to the claimed results by construction.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

No new physical entities are introduced; the ledger captures the training/architecture choices and domain assumptions. The strongest non-standard assumption is that sub-30-s outputs of a 30-s-trained network carry real physiological signal.

free parameters (6)
  • Dataset sampling balance alpha (Eq. 1) = 0.5
    Hand-chosen to balance equal dataset weighting vs recording count; affects training distribution and downstream generalization.
  • Channel-count sampling distribution p_n (Eq. 2)
    Defined over n=1..Nch; hand-chosen to expose model to varied channel counts; directly enables 'any channel' behavior.
  • Inference pooling resolutions = 1 to 3840 predictions per 30-s epoch (14 values)
    Grid chosen for high-frequency analysis; best resolution selected post hoc on test sets, central to arousal and biomarker claims.
  • Training hyperparameters = LR=1e-5, batch=64/32, 35-epoch windows, early stopping=100, max 10,000
    Hand-set following U-Sleep/Fiorillo; not fitted to validation but influence performance.
  • Attention module size/placement = MLP hidden=40, 13 modules, 4 heads for early fusion
    Architecture choices made by the authors; the placement ablation shows sensitivity.
  • Preprocessing normalization = median/IQR, clip [-20,20], resample 128 Hz
    Fixed choices that define input representation; not fitted to target labels.
axioms (5)
  • domain assumption 30-s AASM/R&K expert stage labels are reliable ground truth for model training and evaluation
    All training and the main 30-s claim rest on these labels; inter-rater variability is known (refs 1-3).
  • ad hoc to paper Changing the inference pooling window yields physiologically meaningful sub-30-s sleep stages from a network trained on 30-s labels
    Core to high-frequency claims; not directly validated against high-res expert annotations; paper acknowledges uncertainty.
  • domain assumption Expert-annotated arousals in MASS C1/C3 are a valid reference for evaluating Wake predictions
    Used in Section 2.3; arousal scoring is itself variable.
  • ad hoc to paper Triplet transition counts from model predictions capture sleep fragmentation relevant to age, sex, and OSA
    Following Perslev et al.; no physiological ground truth for transition counts.
  • domain assumption EEG/EOG channels without EMG suffice for competitive sleep staging
    Authors cite prior studies; preliminary experiments showed no EMG benefit.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of AnySleep: a channel-agnostic deep learning system for high-resolution sleep staging in multi-center cohorts." pith.science (2026). https://pith.science/paper/RF2MQGT3

@misc{pith2026251214461,
  author       = {Pith},
  title        = {Pith review of: AnySleep: a channel-agnostic deep learning system for high-resolution sleep staging in multi-center cohorts},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RF2MQGT3}},
  note         = {Machine review of arXiv:2512.14461}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Sleep is essential for health, yet studying its dynamics requires manual sleep staging, a labor-intensive step in research and clinical care. Across centers, polysomnography (PSG) recordings are traditionally scored in 30-s epochs for pragmatic, not physiological, reasons and vary in electrode count, montage, and subject characteristics. These constraints challenge harmonized multi-center studies and the discovery of robust biomarkers on shorter timescales. We present AnySleep, a deep neural network that scores sleep from any electroencephalography (EEG) or electrooculography (EOG) data at adjustable temporal resolutions. We trained and validated the model on over 20,000 overnight recordings (> 200,000 hours of EEG and EOG) from 28 datasets across multiple clinics to promote robust generalization across sites. The model attains state-of-the-art performance and surpasses or equals established baselines at 30-s epochs. Performance improves with more channels, yet remains strong when EOG is absent or only EOG or single EEG derivations (frontal, central, or occipital) are available. On sub-30-s timescales, the model captures short wake intrusions consistent with arousals and improves prediction of pathophysiological conditions (obstructive sleep apnea, narcolepsy type 1, insomnia) over 30-s scoring. We make the model publicly available to facilitate large-scale studies with heterogeneous electrode setups and accelerate biomarker discovery in sleep.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Rethinking Random Transformers as Adaptive Sequence Smoothers for Sleep Staging

    cs.LG 2026-05 unverdicted novelty 5.0

    Randomly initialized Transformers act as adaptive sequence smoothers for sleep staging via a Random Attention Prior Kernel, with gains mainly from inductive bias rather than training.

Reference graph

Works this paper leans on

69 extracted references · 28 canonical work pages · cited by 1 Pith paper · 1 internal anchor

  1. [1]

    Cesari, M.et al.Interrater sleep stage scoring relia- bility between manual scoring from two European sleep centers and automatic scoring performed by the artificial intelligence–based Stanford-STAGES algorithm.J. Clin. Sleep Med.17, 1237–1247, DOI: https://doi.org/10.5664/jcsm.9174 (2021)

    work page doi:10.5664/jcsm.9174 2021

  2. [2]

    J., Lee, J

    Lee, Y. J., Lee, J. Y., Cho, J. H. & Choi, J. H. Interrater reliability of sleep stage scoring: A meta- analysis.J. Clin. Sleep Med.18, 193–202, DOI: https://doi.org/10.5664/jcsm.9538 (2022)

    work page doi:10.5664/jcsm.9538 2022

  3. [3]

    Sleep Res.33, DOI: https://doi.org/10.1111/jsr.13956 (2023)

    Nikkonen, S.et al.Multicentre sleep-stage scoring agreement in the Sleep Revolution project.J. Sleep Res.33, DOI: https://doi.org/10.1111/jsr.13956 (2023)

    doi:10.1111/jsr.13956 2023

  4. [4]

    L., Harvey, E

    Loomis, A. L., Harvey, E. N. & Hobart, G. Electrical potentials of the human brain.J. Exp. Psychol.19, 249–279, DOI: https://doi.org/10.1037/h0062089 (1936)

    work page doi:10.1037/h0062089 1936

  5. [5]

    Decat, N.et al.Beyond traditional visual sleep scoring: massive feature extraction and unsupervised clustering of sleep time series.bioRxivDOI: https: //doi.org/10.1101/2021.09.08.458981 (2021)

    work page doi:10.1101/2021.09.08.458981 2021

  6. [6]

    & Peter-Derex, L

    Lambert, I. & Peter-Derex, L. Spotlight on Sleep Stage Classification Based on EEG.Nat. Sci. Sleep Volume 15, 479–490, DOI: https://doi.org/10.2147/ nss.s401270 (2023)

    2023

  7. [7]

    Korkalainen, H.et al.Detailed Assessment of Sleep ArchitectureWithDeepLearningandShorterEpoch- to-Epoch Duration Reveals Sleep Fragmentation of Patients With Obstructive Sleep Apnea.IEEE J. Biomed. Heal. Informatics25, 2567–2574, DOI:https: //doi.org/10.1109/JBHI.2020.3043507 (2021)

    arXiv 2020

  8. [8]

    Neurophysiol.131, 1187–1203, DOI: https://doi.org/10.1016/j.clinph.2020.02.027 (2020)

    Brink-Kjaer, A.et al.Automatic detection of cortical arousals in sleep and their contribution to daytime sleepiness.Clin. Neurophysiol.131, 1187–1203, DOI: https://doi.org/10.1016/j.clinph.2020.02.027 (2020)

    work page doi:10.1016/j.clinph.2020.02.027 2020

  9. [9]

    Neu- 14 rosci.15, DOI: https://doi.org/10.3389/fnins.2021.5 64098 (2021)

    Malafeev, A.et al.Automatic Detection of Mi- crosleep Episodes With Deep Learning.Front. Neu- 14 rosci.15, DOI: https://doi.org/10.3389/fnins.2021.5 64098 (2021)

    work page doi:10.3389/fnins.2021.5 2021

  10. [10]

    Cesari, M.et al.A data-driven system to iden- tify REM sleep behavior disorder and to predict its progression from the prodromal stage in Parkin- son’s disease.Sleep Med.77, 238–248, DOI: https: //doi.org/10.1016/j.sleep.2020.04.010 (2021)

    work page doi:10.1016/j.sleep.2020.04.010 2020

  11. [11]

    Younes, M.et al.Odds Ratio Product of Sleep EEG as a Continuous Measure of Sleep State.Sleep38, 641–654, DOI: https://doi.org/10.5665/sleep.4588 (2015)

    work page doi:10.5665/sleep.4588 2015

  12. [12]

    E.et al.Examining Initial Sleep Onset in Primary Insomnia: A Case-Control Study Using 4-Second Epochs.J

    Moul, D. E.et al.Examining Initial Sleep Onset in Primary Insomnia: A Case-Control Study Using 4-Second Epochs.J. Clin. Sleep Med.03, 479–488, DOI: https://doi.org/10.5664/jcsm.26912 (2007)

    work page doi:10.5664/jcsm.26912 2007

  13. [13]

    Follin, L. F.et al.An inter-rater variability study between human and automatic scorers in 5-s mini- epochs of sleep.Sleep Med.128, 139–150, DOI: https: //doi.org/10.1016/j.sleep.2025.02.005 (2025)

    work page doi:10.1016/j.sleep.2025.02.005 2025

  14. [14]

    F.et al.Optimizing automated sleep stage scoring of 5-second mini-epochs: a transfer learning study.bioRxivDOI: https://doi.org/10.1101/2025.0 6.28.661238 (2025)

    Follin, L. F.et al.Optimizing automated sleep stage scoring of 5-second mini-epochs: a transfer learning study.bioRxivDOI: https://doi.org/10.1101/2025.0 6.28.661238 (2025)

    work page doi:10.1101/2025.0 2025

  15. [15]

    N., Jørgen Jennum, P., Mignot, E

    Olesen, A. N., Jørgen Jennum, P., Mignot, E. & Sorensen, H. B. D. Automatic sleep stage classifica- tion with deep residual networks in a mixed-cohort setting.Sleep44, DOI: https://doi.org/10.1093/slee p/zsaa161 (2020)

    work page doi:10.1093/slee 2020

  16. [16]

    & Thorey, V

    Guillot, A. & Thorey, V. RobustSleepNet: Trans- fer Learning for Automated Sleep Staging at Scale. IEEE T. Neur. Sys. Reh.29, 1441–1451, DOI: https: //doi.org/10.1109/tnsre.2021.3098968 (2021)

    arXiv 2021

  17. [17]

    Medicine4, DOI: https: //doi.org/10.1038/S41746-021-00440-5 (2021)

    Perslev, M.et al.U-Sleep: Resilient high-frequency sleep staging.npj Digit. Medicine4, DOI: https: //doi.org/10.1038/S41746-021-00440-5 (2021)

    doi:10.1038/s41746-021-00440-5 2021

  18. [18]

    & Walker, M

    Vallat, R. & Walker, M. P. An open-source, high- performance tool for automated sleep staging.eLife 10, e70092, DOI: https://doi.org/10.7554/elife.70092 (2021)

    doi:10.7554/elife.70092 2021

  19. [19]

    & Flöel, A

    Hanna, J. & Flöel, A. An accessible and versatile deep learning-based sleep stage classifier.Front. Neuroinform.17, DOI: https://doi.org/10.3389/ FNINF.2023.1086634 (2023)

    arXiv 2023

  20. [20]

    Berry, R. B.et al. The AASM manual for the scoring of sleep and associated events: Rules, terminology and technical specifications, Version 2.6(American Academy of Sleep Medicine, Darien, Illinois, 2020)

    2020

  21. [21]

    N., Arnal, P

    Chambon, S., Galtier, M. N., Arnal, P. J., Wainrib, G. & Gramfort, A. A deep learning architecture for temporal sleep stage classification using multivariate and multimodal time series.IEEE T. Neur. Sys. Reh.26, 758–769, DOI: https://doi.org/10.1109/tn sre.2018.2813138 (2018)

    arXiv 2018

  22. [22]

    Phan, H., Andreotti, F., Cooray, N., Chén, O. Y. & Vos, M. D. Joint Classification and Prediction CNN Framework for Automatic Sleep Stage Classification. IEEE Trans. Biomed. Eng.66, 1285–1296, DOI: ht tps://doi.org/10.1109/TBME.2018.2872652 (2019)

    arXiv 2018

  23. [23]

    & Brox, T

    Ronneberger, O., Fischer, P. & Brox, T. U-Net: Convolutional Networks for Biomedical Image Seg- mentation. InLecture Notes in Computer Science, 234–241, DOI: https://doi.org/10.1007/978-3-3 19-24574-4_28 (Springer International Publishing, 2015)

    doi:10.1007/978-3-3 2015

  24. [24]

    W.et al.Representations of tem- poral sleep dynamics: Review and synthesis of the literature.Sleep Med

    Hermans, L. W.et al.Representations of tem- poral sleep dynamics: Review and synthesis of the literature.Sleep Med. Rev.63, 101611, DOI: https://doi.org/10.1016/j.smrv.2022.101611 (2022)

    arXiv 2022

  25. [25]

    Bonnet, M. H. & Arand, D. L. EEG Arousal Norms by Age.J. Clin. Sleep Med.03, 271–274, DOI: https: //doi.org/10.5664/jcsm.26796 (2007)

    work page doi:10.5664/jcsm.26796 2007

  26. [26]

    M., Carskadon, M

    Ohayon, M. M., Carskadon, M. A., Guilleminault, C. & Vitiello, M. V. Meta-Analysis of Quantitative Sleep Parameters From Childhood to Old Age in Healthy Individuals: Developing Normative Sleep Values Across the Human Lifespan.Sleep27, 1255– 1273, DOI: https://doi.org/10.1093/sleep/27.7.1255 (2004)

    work page doi:10.1093/sleep/27.7.1255 2004

  27. [27]

    A., Winer, J

    Mander, B. A., Winer, J. R. & Walker, M. P. Sleep and Human Aging.Neuron94, 19–36, DOI: https: //doi.org/10.1016/j.neuron.2017.02.004 (2017)

    work page doi:10.1016/j.neuron.2017.02.004 2017

  28. [28]

    Redline, S.et al.The Effects of Age, Sex, Eth- nicity, and Sleep-Disordered Breathing on Sleep Architecture.Arch. Intern. Med.164, 406, DOI: https://doi.org/10.1001/archinte.164.4.406 (2004)

    work page doi:10.1001/archinte.164.4.406 2004

  29. [29]

    & Chellappa, S

    Lok, R., Qian, J. & Chellappa, S. L. Sex differences in sleep, circadian rhythms, and metabolism: Impli- cations for precision medicine.Sleep Med. Rev.75, 101926, DOI: https://doi.org/10.1016/j.smrv.2024. 101926 (2024)

    work page doi:10.1016/j.smrv.2024 2024

  30. [30]

    Bonnet, M. H. & Arand, D. L. Clinical effects of sleep fragmentation versus sleep deprivation.Sleep Med. Rev.7, 297–310, DOI: https://doi.org/10.105 3/smrv.2001.0245 (2003)

    arXiv 2001

  31. [31]

    G., Scott, M

    Norman, R. G., Scott, M. A., Ayappa, I., Walsleben, J. A. & Rapoport, D. M. Sleep Continuity Measured By Survival Curve Analysis.Sleep29, 1625–1631, DOI: https://doi.org/10.1093/sleep/29.12.1625 (2006)

    work page doi:10.1093/sleep/29.12.1625 2006

  32. [32]

    R., Ulfenborg, B

    Stahlschmidt, S. R., Ulfenborg, B. & Synnergren, J. Multimodal deep learning for biomedical data fusion: a review.Briefings Bioinform.23, DOI: https://doi.org/10.1093/BIB/BBAB569 (2022)

    doi:10.1093/bib/bbab569 2022

  33. [33]

    CoRRabs/2409.12454, DOI: https://doi.org/10.485 50/ARXIV.2409.12454 (2024)

    Shi, E.et al.FoME: A Foundation Model for EEG using Adaptive Temporal-Lateral Attention Scaling. CoRRabs/2409.12454, DOI: https://doi.org/10.485 50/ARXIV.2409.12454 (2024). 15

    Pith/arXiv arXiv doi:10.48550/arxiv.2409.12454 2024

  34. [34]

    NAPS: Attention-Based Fusion of Heterogeneous Physiological Signals

    Rossi, A. D.et al.NAP: Attention-Based Late Fusion for Automatic Sleep Staging, DOI: https: //doi.org/10.48550/ARXIV.2511.03488 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2511.03488 2025

  35. [35]

    Brain Sci.11, 1274, DOI: https://doi.org/10.3390/ brainsci11101274 (2021)

    Qian, X.et al.A Review of Methods for Sleep Arousal Detection Using Polysomnographic Signals. Brain Sci.11, 1274, DOI: https://doi.org/10.3390/ brainsci11101274 (2021)

    2021

  36. [36]

    Cesari, M.et al.Sleep modelled as a continuous and dynamic process predicts healthy ageing better than traditional sleep scoring.Sleep Med.77, 136–146, DOI: https://doi.org/10.1016/j.sleep.2020.11.033 (2021)

    work page doi:10.1016/j.sleep.2020.11.033 2020

  37. [37]

    & Mikkelsen, K

    Phan, H. & Mikkelsen, K. Automatic sleep staging of EEG signals: Recent development, challenges, and future directions.Physiol. Meas.43, 04TR01, DOI: https://doi.org/10.1088/1361-6579/ac6049 (2022)

    doi:10.1088/1361-6579/ac6049 2022

  38. [38]

    van Gorp, H.et al.Certainty about uncertainty in sleep staging: a theoretical framework.Sleep45, DOI: https://doi.org/10.1093/sleep/zsac134 (2022)

    work page doi:10.1093/sleep/zsac134 2022

  39. [39]

    Rep.15, DOI: https: //doi.org/10.1038/s41598-025-06019-4 (2025)

    Bechny, M.et al.Beyond accuracy: a framework for evaluating algorithmic bias and performance, applied to automated sleep scoring.Sci. Rep.15, DOI: https: //doi.org/10.1038/s41598-025-06019-4 (2025)

    work page doi:10.1038/s41598-025-06019-4 2025

  40. [40]

    Chambon, S., Galtier, M. N. & Gramfort, A. Domain adaptation with optimal transport improves EEG sleep stage classifiers. In2018 International Work- shop on Pattern Recognition in Neuroimaging, PRNI 2018, Singapore, Singapore, June 12-14, 2018, 1–4, DOI: https://doi.org/10.1109/PRNI.2018.8423957 (IEEE, 2018)

    arXiv 2018

  41. [41]

    Phan, H.et al.Towards More Accurate Automatic Sleep Staging via Deep Transfer Learning.IEEE Trans. Biomed. Eng.68, 1787–1798, DOI: https: //doi.org/10.1109/TBME.2020.3020381 (2021)

    arXiv 2020

  42. [42]

    J.et al.Validation of the sleep EEG headband ZMax.bioRxivDOI: https://doi.org/10.1 101/2023.08.18.553744 (2023)

    Esfahani, M. J.et al.Validation of the sleep EEG headband ZMax.bioRxivDOI: https://doi.org/10.1 101/2023.08.18.553744 (2023)

    2023

  43. [43]

    & Menon, C

    Markov, K., Elgendi, M. & Menon, C. Evaluating the performance of wearable EEG sleep monitoring devices: a meta-analysis approach.npj Biomed. Innov.2, DOI: https://doi.org/10.1038/s44385-025 -00034-w (2025)

    work page doi:10.1038/s44385-025 2025

  44. [44]

    P.et al.Gastric Banding Surgery versus Continuous Positive Airway Pressure for Obstructive Sleep Apnea: A Randomized Controlled Trial.Am

    Bakker, J. P.et al.Gastric Banding Surgery versus Continuous Positive Airway Pressure for Obstructive Sleep Apnea: A Randomized Controlled Trial.Am. J. Resp. Crit. Care197, 1080–1083, DOI: https: //doi.org/10.1164/rccm.201708-1637le (2018)

    doi:10.1164/rccm.201708-1637le 2018

  45. [45]

    Zhang, G.-Q.et al.The National Sleep Research Resource: Towards a Sleep Data Commons.J. Am. Med. Inform. Assn.25, 1351–1358, DOI: https://doi. org/10.1093/jamia/ocy064 (2018)

    doi:10.1093/jamia/ocy064 2018

  46. [46]

    L.et al.Prevalence and risk factors for sleep-disordered breathing in 8- to 11-year-old chil- dren: Association with race and prematurity.J

    Rosen, C. L.et al.Prevalence and risk factors for sleep-disordered breathing in 8- to 11-year-old chil- dren: Association with race and prematurity.J. Pediatr.142, 383–389, DOI: https://doi.org/10.106 7/mpd.2003.28 (2003)

    2003

  47. [47]

    Redline, S.et al.The Familial Aggregation of Ob- structive Sleep Apnea.Am. J. Resp. Crit. Care151, 682–687, DOI: https://doi.org/10.1164/ajrccm/151 .3_pt_1.682 (1995)

    work page doi:10.1164/ajrccm/151 1995

  48. [48]

    L.et al.A Randomized Trial of Adeno- tonsillectomy for Childhood Sleep Apnea.New Engl

    Marcus, C. L.et al.A Randomized Trial of Adeno- tonsillectomy for Childhood Sleep Apnea.New Engl. J. Med.368, 2366–2376, DOI: https://doi.org/10.1 056/nejmoa1215881 (2013)

    2013

  49. [49]

    Perslev, M.et al.DCSM Sleep Staging Dataset, DOI: https://doi.org/10.17894/UCPH.282D3C1E-9 B98-4C1E-886E-704AFDFA9179 (2021)

    work page doi:10.17894/ucph.282d3c1e-9 2021

  50. [50]

    Rosen, C. L.et al.A Multisite Randomized Trial of Portable Sleep Studies and Positive Airway Pressure Autotitration Versus Laboratory-Based Polysomnog- raphy for the Diagnosis and Treatment of Obstruc- tive Sleep Apnea: The HomePAP Study.Sleep35, 757–767, DOI: https://doi.org/10.5665/sleep.1870 (2012)

    doi:10.5665/sleep.1870 2012

  51. [51]

    Chen, X.et al.Racial/Ethnic Differences in Sleep Disturbances: The Multi-Ethnic Study of Atheroscle- rosis (MESA).SleepDOI: https://doi.org/10.5665/ sleep.4732 (2015)

    2015

  52. [52]

    Blackwell, T.et al.Associations Between Sleep Architecture and Sleep-Disordered Breathing and Cognition in Older Community-Dwelling Men: The Osteoporotic Fractures in Men Sleep Study.J. Am. Geriatr. Soc.59, 2217–2225, DOI: https://doi.org/ 10.1111/j.1532-5415.2011.03731.x (2011)

    arXiv 2011

  53. [53]

    M.et al.You Snooze, You Win: The PhysioNet/Computing in Cardiology Challenge

    Ghassemi, M. M.et al.You Snooze, You Win: The PhysioNet/Computing in Cardiology Challenge

  54. [54]

    Goldberger, A. L.et al.PhysioBank, PhysioToolkit, and PhysioNet: Components of a New Research Re- source for Complex Physiologic Signals.Circulation 101, DOI: https://doi.org/10.1161/01.cir.101.23.e215 (2000)

    doi:10.1161/01.cir.101.23.e215 2000

  55. [55]

    H., Tuk, B., Kamphuisen, H

    Kemp, B., Zwinderman, A. H., Tuk, B., Kamphuisen, H. A. C. & Oberye, J. J. L. Analysis of a sleep- dependent neuronal feedback loop: The slow-wave microcontinuity of the EEG.IEEE Trans. Biomed. Eng.47, 1185–1194, DOI: https://doi.org/10.1109/ 10.867928 (2000)

    2000

  56. [56]

    F.et al.The Sleep Heart Health Study: Design, Rationale, and Methods.Sleep20, 1077– 1085, DOI: https://doi.org/10.1093/sleep/20.12.1077 (1997)

    Quan, S. F.et al.The Sleep Heart Health Study: Design, Rationale, and Methods.Sleep20, 1077– 1085, DOI: https://doi.org/10.1093/sleep/20.12.1077 (1997)

    doi:10.1093/sleep/20.12.1077 1997

  57. [57]

    P.et al.Sleep-Disordered Breathing and Cognition in Older Women.J

    Spira, A. P.et al.Sleep-Disordered Breathing and Cognition in Older Women.J. Am. Geriatr. Soc.56, 16 45–50, DOI: https://doi.org/10.1111/j.1532-5415.20 07.01506.x (2007)

    work page doi:10.1111/j.1532-5415.20 2007

  58. [58]

    Guillot, A., Sauvet, F., During, E. H. & Thorey, V. Dreem Open Datasets: Multi-Scored Sleep Datasets to Compare Human and Automated Sleep Staging. IEEE T. Neur. Sys. Reh.28, 1955–1965, DOI: https: //doi.org/10.1109/tnsre.2020.3011181 (2020)

    arXiv 1955

  59. [59]

    Khalighi, S., Sousa, T., dos Santos, J. M. & Nunes, U. ISRUC-Sleep: A comprehensive public dataset for sleep researchers.Comput. Methods Programs Biomed.124, 180–192, DOI: https://doi.org/10.101 6/J.CMPB.2015.10.013 (2016)

    2015

  60. [60]

    & Nielsen, T

    O’Reilly, C., Gosselin, N., Carrier, J. & Nielsen, T. Montreal Archive of Sleep Studies: An open- access resource for instrument benchmarking and exploratory research.J. Sleep Res.23, 628–635, DOI: https://doi.org/10.1111/jsr.12169 (2014)

    doi:10.1111/jsr.12169 2014

  61. [61]

    Vincent’s University Hospital / University College Dublin Sleep Apnea Database, DOI: https://doi.org/10.13026/C26C7D (2004)

    McNicholas, W.et al.St. Vincent’s University Hospital / University College Dublin Sleep Apnea Database, DOI: https://doi.org/10.13026/C26C7D (2004)

    work page doi:10.13026/c26c7d 2004

  62. [62]

    Perslev, M.et al.U-Sleep Data Repository, DOI: https://doi.org/10.17894/UCPH.0D1554E9-D86 B-4E08-B3C2-632B730CD362 (2021)

    work page doi:10.17894/ucph.0d1554e9-d86 2021

  63. [63]

    A Manual of Standard- ized Terminology, Techniques and Scoring System for Sleep Stages of Human Subjects(Public Health Service, U.S

    Rechtschaffen, A.et al. A Manual of Standard- ized Terminology, Techniques and Scoring System for Sleep Stages of Human Subjects(Public Health Service, U.S. Government Printing Office, Washing- ton, D.C., 1968)

    1968

  64. [64]

    In Guyon, I.et al.(eds.)Annu

    Vaswani, A.et al.Attention is All you Need. In Guyon, I.et al.(eds.)Annu. Conf. Neural Informa- tion Processing Systems, NeurIPS, 4–9 December, 5998–6008 (Curran Associates Inc., Red Hook, NY, USA, 2017)

    2017

  65. [65]

    J., Kale, S

    Reddi, S. J., Kale, S. & Kumar, S. On the Con- vergence of Adam and Beyond. In6th Int. Conf. Learning Representations, ICLR, April 30 – May 3 (OpenReview.net, 2018)

    2018

  66. [66]

    Medicine6, 33, DOI: https: //doi.org/10.1038/S41746-023-00784-0 (2023)

    Fiorillo, L.et al.U-Sleep’s resilience to AASM guidelines.npj Digit. Medicine6, 33, DOI: https: //doi.org/10.1038/S41746-023-00784-0 (2023)

    doi:10.1038/s41746-023-00784-0 2023

  67. [67]

    Classification assessment methods

    Tharwat, A. Classification assessment methods. Appl. Comput. Inform.17, 168–192, DOI: https: //doi.org/10.1016/j.aci.2018.08.003 (2021)

    doi:10.1016/j.aci.2018.08.003 2018

  68. [68]

    Outcomes of Sleep Disorders in Older Men,

    Pedregosa, F.et al.Scikit-learn: Machine Learning in Python.J. Mach. Learn. Res.12, 2825–2830 (2011). Acknowledgements We are grateful to M. Reißel and V. Sander for provid- ing us with computing resources. The Apnea, Bariatric surgery, and CPAP study (ABC Study) was supported by National Institutes of Health grants R01HL106410 and K24HL127307. Philips Re...

    2011

  69. [2018]

    InComputing in Cardiology, CinC 2018, Maas- tricht, The Netherlands, September 23-26, 2018, 1–4, DOI: https://doi.org/10.22489/CINC.2018.049 (www.cinc.org, 2018)

    doi:10.22489/cinc.2018.049 2018

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.