Pith. sign in

REVIEW 3 major objections 5 minor 32 references

Sleep Stage Classification using Multimodal Embedding Fusion from EOG and PSM

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Fine-tuned ImageBind scores 74.5% on EEG-free sleep staging.

desk verdict A workmanlike ImageBind application to sleep staging with a new clinical dataset and decent baselines, but the headline 'significant' superiority over DeepSleepNet is a rounding-level tie with no significance testing. read the letter →

arxiv 2506.06912 v1 pith:SVSIBCSG submitted 2025-06-07 cs.CV

Olivier Papillon , Rafik Goubran , James Green , Julien Larivière-Chartier , Caitlin Higginson , Frank Knoefel , Rébecca Robillard This is my paper
classification cs.CV
keywords sleepstageclassificationelectrooculography(EOG)pressure-sensitivemat(PSM)multimodalembeddingfusionImageBindtransferlearningfine-tuninghomemonitoring
verification ladder T0 review T1 audit T2 compute T3 formal

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 sets out to show that five-stage sleep–wake classification can be done with two unobtrusive sensors, electrooculography (EOG) and a pressure-sensitive mat (PSM), instead of the EEG electrodes used in standard polysomnography. Its central move is to repurpose ImageBind, a pre-trained model originally built to bind audio and video, by encoding the two EOG channels as a stereo audio file and the mat's 18-by-8 pressure frames as video. On 85 clinical nights, fine-tuning this model on both modalities reaches 0.745 accuracy and 0.683 average per-stage F1, slightly above the best single-channel EOG baseline and well above PSM-only or bottleneck-fusion alternatives. The same model reaches 0.710 accuracy with EOG alone, and even without fine-tuning it reaches 0.690, which the paper reads as evidence that pre-trained multimodal embeddings transfer to a medical domain far from their original training. If this holds, home sleep monitoring may not need scalp electrodes or a sleep laboratory.

What carries the argument

The load-bearing mechanism is ImageBind's joint embedding space, which aligns audio and video into a common representation. EOG is prepared as 16 kHz dual-channel WAV audio and converted to a 128-mel spectrogram; PSM is supplied as low-resolution 18-by-8 pressure frames treated as video. Each modality is embedded into a 1024-dimensional vector, the two vectors are concatenated into 2048 dimensions, and a single linear layer maps that fused vector to the five sleep stages. Fine-tuning updates the ImageBind weights, while linear probing keeps them fixed, and the comparison between those two settings isolates what the pre-trained alignment itself contributes.

What would settle it

A reader could test the central claim by holding the training protocol fixed and swapping only the input encodings: feed the same EOG epochs as raw waveforms to a 1D convolutional network and the same PSM frames as flattened time series to a temporal model, then compare against the ImageBind pipeline at equal data splits; if the simpler encodings match or beat 0.745 accuracy, the multimodal embedding alignment is not the source of the gain. Alternatively, changing the mel-spectrogram resolution or the audio sampling rate should change accuracy if the audio analogy is load-bearing.

Watch

Extended reading notes

Core claim

The central claim is that fusing dual-channel EOG and PSM through ImageBind's shared embedding space improves five-stage sleep staging over each modality alone and over existing deep models. The paper reports that the fine-tuned fusion reaches 0.745 accuracy and 0.683 average per-stage F1 on 85 nights, compared with 0.743/0.682 for DeepSleepNet on single-channel EOG, 0.631/0.543 for MBT on PSM video plus single-channel EOG, and 0.399/0.164 for ViViT on PSM video alone. An ImageBind model using only dual-channel EOG reaches 0.710/0.636, and a linear-probed (frozen-weights) fusion reaches 0.690/0.614, which the authors take as evidence that the pre-trained embedding space adapts to this new domain with limited labelled data. To the paper's knowledge this is the first reported use of ImageBind to fuse PSM and EOG data for sleep stage classification.

Load-bearing premise

Everything rests on the assumption that turning 250–512 Hz EOG into 16 kHz audio spectrograms and 18-by-8 pressure frames into video preserves the sleep-stage information ImageBind needs; if that encoding loses discriminative content, the reported gains come from the preprocessing rather than from the sensors.

Editorial extensions

If this is right

  • A practical sleep monitor could run on a bed mat and two small eye electrodes, dropping the EEG cap and the laboratory setting.
  • Adding the mat to EOG yields a modest but real improvement (accuracy from 0.710 to 0.745, per-stage F1 from 0.636 to 0.683), so body-pressure information does add signal beyond eye movements.
  • Pre-trained non-medical models transfer to sleep staging with little labelled data: linear probing alone reaches 0.690 accuracy, which matters for clinical datasets that are small and expensive to annotate.
  • The remaining errors are mostly between adjacent stages (NREM1, NREM2, NREM3), so further gains will require better separation of those stages rather than better detection of Wake and REM.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A reader should not yet attribute the gain to multimodal fusion per se: the jump from linear probing (0.690) to fine-tuning (0.745) is larger than the jump from EOG-only to EOG-plus-PSM (0.710 to 0.745), so the EOG-as-audio encoder may be carrying most of the improvement.
  • The audio analogy can be tested directly: changing the mel-spectrogram parameters, the 16 kHz resampling, or replacing the spectrogram with a raw-waveform encoder should move the accuracy if the audio encoding is what makes ImageBind work.
  • The same recipe could be applied to other physiological pairs that can be rendered as audio and video, such as respiratory effort plus camera-based body movement, to see whether the audio-video binding generalizes beyond sleep staging.
  • For clinical use, per-stage recall will matter more than mean accuracy; the reported macro-F1 of 0.683 implies that the confused N1/N2/N3 pairs still limit downstream tasks like apnea staging.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The manuscript proposes a sleep stage classification pipeline that converts dual-channel EOG into 16 kHz WAV audio and 128-mel spectrograms, and PSM pressure data into low-resolution videos, then feeds both into ImageBind to obtain 1024-dimensional embeddings that are concatenated and classified by a linear layer. The authors evaluate on 85 clinical nights with patient-level five-fold cross-validation and report that fine-tuned ImageBind fusing EOG and PSM reaches 0.745 accuracy and 0.683 macro-F1, outperforming DeepSleepNet (0.743/0.682), MBT (0.631/0.543), and ViViT (0.399/0.164). They also report that linear probing without fine-tuning achieves 0.690/0.614 and that EOG-only fine-tuning reaches 0.710/0.636. The central claims are that fine-tuning ImageBind significantly improves performance and that fusing PSM with EOG provides an additional benefit.

Significance. If the headline comparison were statistically supported, the paper would make a useful contribution: it is the first reported application of ImageBind to PSM+EOG fusion, it uses a clinical dataset with a sensible patient-level split, and it demonstrates that a non-medical pretrained audio-video model can be adapted to very different physiological inputs with limited labels. The preprocessing is described in enough detail to be replicated, and the use of frozen embeddings as a baseline is methodologically clean and avoids circularity. However, the significance of the results is currently the main unresolved issue: the differences against DeepSleepNet are numerically negligible and no uncertainty quantification is provided, so the paper's stated contribution is not yet established.

major comments (3)
  1. [Section IV, Table I] The claim in the abstract and Section IV that fine-tuning ImageBind 'significantly improves' classification accuracy is not supported by the reported numbers. ImageBind (PSM + dual-channel EOG) achieves 0.745 accuracy and 0.683 macro-F1, while DeepSleepNet (single-channel EOG) achieves 0.743 and 0.682; the advantage is 0.002 accuracy and 0.001 F1, which is far smaller than the expected patient-level variability for a 5-class sleep staging task. No per-fold results, confidence intervals, or paired significance tests (e.g., McNemar's test or a patient-level Wilcoxon signed-rank test) are reported. The paper should either provide such statistical evidence or remove the word 'significantly' and reframe the comparison as a feasibility result.
  2. [Section IV, Table II and Section IV.A] The fusion benefit claim is also presented without uncertainty quantification. The text states that adding PSM data improves accuracy from 0.710 to 0.745 and F1 from 0.636 to 0.683, and that linear probing is substantially worse than fine-tuning (0.690/0.614 vs 0.745/0.683), but no error bars or significance tests are given. Because the differences are modest and the sample size is only 85 patients, these comparisons need the same per-fold reporting and paired tests as the main comparison. Without this, the conclusion that multimodal fusion with ImageBind improves sleep staging is not yet demonstrated.
  3. [Section III-E] The hyperparameters for all models were selected by choosing the configuration with the highest average five-way accuracy across the five folds of the same cross-validation that is used to report the final results. This means the reported numbers are optimistic estimates of generalization performance: the test folds have influenced model selection. The paper acknowledges in Section IV that repeated cross-validation during hyperparameter optimization 'can sometimes lead to overfitting' but does not quantify or correct for this. A nested cross-validation or a fixed held-out validation set should be used, or the authors should at least report the variance of the selected configuration across folds and show that the conclusions are robust to the hyperparameter choice.
minor comments (5)
  1. [Section III-B] The phrase 'thereby avoiding overfitting' should be 'avoiding patient-level data leakage'; a patient-level split prevents leakage of data from the same subject into both training and test sets, not overfitting in general.
  2. [Section III-D] There is a typo: 'single-chanel' should be 'single-channel'. Also, 'LOG' in the MBT description appears to be a typo for 'EOG' (the left EOG channel).
  3. [Section IV.A] The heading 'Comparaison' is French; it should be 'Comparison'.
  4. [Section III-C] The formatted 'W A V' with spaces is a rendering issue; use 'WAV' consistently throughout the text and figure captions.
  5. [Table I] The model name capitalization is inconsistent: 'ViVit' in Table I and the caption versus 'ViViT' in the text and references. Also, consider reporting per-class F1 or the class distribution, since macro-F1 alone can hide severe imbalance effects in sleep staging.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported accuracies are empirical measurements on held-out clinical folds compared against independently pretrained baselines.

full rationale

The paper's central empirical claim (fine-tuned ImageBind with PSM and dual-channel EOG reaches 0.745 accuracy and 0.683 macro-F1) is measured on held-out patient folds from 85 clinical nights and compared against independently pretrained baselines (DeepSleepNet, ViViT, MBT) that are fine-tuned on the same data. ImageBind's embeddings come from an external pretraining corpus, not from sleep labels, so the frozen-embedding result is not defined in terms of the target labels. No predicted quantity is constructed by fitting a parameter to the same quantity and then reading that quantity off as a prediction; the only fitted parameters are classifier weights and hyperparameters, and the reported metrics are computed on test folds. The paper's self-citations (e.g., PSM reviews and prior PSM sleep/wake work) provide background and sensor context but are not used to justify the accuracy numbers. The absence of significance testing and the numerically tiny margin over DeepSleepNet are statistical-evidence concerns, not circularity, and the paper itself acknowledges that further validation on future patient data is needed. Therefore the derivation chain is self-contained with respect to the measured comparisons, and no circular step is exhibited.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The empirical comparison rests on assumptions about label quality, epoch-level stationarity, the validity of treating EOG as audio and PSM as video, and the transferability of ImageBind embeddings. These are standard domain assumptions for this type of study; none are ad hoc invented entities. The main hand-chosen quantities are the training hyperparameters, which were selected by sweeping on the same folds used for evaluation.

free parameters (1)
  • ImageBind fine-tuning hyperparameters = lr=1.4e-7, weight_decay=0.005, epochs=6, batch_size=12
    Selected by a hyperparameter sweep over the same five folds used for reporting, which is model selection on test data and affects the reported 0.745 accuracy.
assumptions (4)
  • domain assumption Sleep stage labels assigned by a sleep technologist are correct and each 30-second epoch is homogeneous.
    Section III-A and III-B explain that stages were annotated by a registered sleep technologist and that data were split into 30-second epochs under the assumption that stage is constant within an epoch.
  • domain assumption EOG sampled at 250 or 512 Hz can be upsampled to 16 kHz and converted to a 128-mel spectrogram without losing sleep-stage-relevant information.
    Section III-C describes this conversion as the only EOG representation; no control experiment tests alternative representations.
  • domain assumption The ImageBind joint embedding space, trained on natural image, audio, and video, transfers to pressure-video and EOG-audio inputs.
    Section III-C and Section IV state that this domain shift is expected to work; the paper tests it empirically but does not derive why.
  • domain assumption Hyperparameter optimization on the same five folds used for final evaluation gives unbiased accuracy estimates.
    Section IV acknowledges that repeated cross-validation can lead to overfitting, so this assumption is questionable.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Sleep Stage Classification using Multimodal Embedding Fusion from EOG and PSM." pith.science (2026). https://pith.science/paper/SVSIBCSG

@misc{pith2026250606912,
  author       = {Pith},
  title        = {Pith review of: Sleep Stage Classification using Multimodal Embedding Fusion from EOG and PSM},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SVSIBCSG}},
  note         = {Machine review of arXiv:2506.06912}
}
read the original abstract

Accurate sleep stage classification is essential for diagnosing sleep disorders, particularly in aging populations. While traditional polysomnography (PSG) relies on electroencephalography (EEG) as the gold standard, its complexity and need for specialized equipment make home-based sleep monitoring challenging. To address this limitation, we investigate the use of electrooculography (EOG) and pressure-sensitive mats (PSM) as less obtrusive alternatives for five-stage sleep-wake classification. This study introduces a novel approach that leverages ImageBind, a multimodal embedding deep learning model, to integrate PSM data with dual-channel EOG signals for sleep stage classification. Our method is the first reported approach that fuses PSM and EOG data for sleep stage classification with ImageBind. Our results demonstrate that fine-tuning ImageBind significantly improves classification accuracy, outperforming existing models based on single-channel EOG (DeepSleepNet), exclusively PSM data (ViViT), and other multimodal deep learning approaches (MBT). Notably, the model also achieved strong performance without fine-tuning, highlighting its adaptability to specific tasks with limited labeled data, making it particularly advantageous for medical applications. We evaluated our method using 85 nights of patient recordings from a sleep clinic. Our findings suggest that pre-trained multimodal embedding models, even those originally developed for non-medical domains, can be effectively adapted for sleep staging, with accuracies approaching systems that require complex EEG data.

Figures

Figures reproduced from arXiv: 2506.06912 by the authors.

Figure 1
Figure 1. Model Overview of the Multimodal Embedding Fusion Process for EOG and PSM [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Model overview of the input modalities for the three [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The Confusion Matrix for ImageBind using Two Input [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 31 canonical work pages

  1. [1]

    Bed-Based Health Monitoring Using Pressure Sensitive Tech- nology: A Review,

    M. Cohen-McFarlane, S. Bennett, B. Wallace, R. Goubran, and F. Knoe- fel, “Bed-Based Health Monitoring Using Pressure Sensitive Tech- nology: A Review,”IEEE Instrumentation & Measurement Magazine, vol. 24, no. 2, pp. 13–23, 2021

    work page 2021

  2. [2]

    Sleep Basics,

    C. Clinic, “Sleep Basics,” 2024. https://my.clevelandclinic.org/ 12148-sleep-basics

    work page 2024

  3. [3]

    An Attention-Based Deep Learning Approach for Sleep Stage Classification With Single-Channel EEG,

    E. Eldele, Z. Chen, C. Liu, M. Wu, C.-K. Kwoh, X. Li, and C. Guan, “An Attention-Based Deep Learning Approach for Sleep Stage Classification With Single-Channel EEG,”IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 29, pp. 809–818, 2021

    work page 2021

  4. [4]

    Machine Learning-Based Automatic Detection of Central Sleep Apnea Events From a Pressure Sensitive Mat,

    H. Azimi, P. Xi, M. Bouchard, R. Goubran, and F. Knoefel, “Machine Learning-Based Automatic Detection of Central Sleep Apnea Events From a Pressure Sensitive Mat,”IEEE Access, vol. 8, pp. 173428– 173439, 2020

    work page 2020

  5. [5]

    Real-time Neonatal Respiratory Rate Estimation using a Pressure-Sensitive Mat,

    A. Bekele, S. Nizami, Y . S. Dosso, C. Aubertin, K. Greenwood, J. Har- rold, and J. R. Green, “Real-time Neonatal Respiratory Rate Estimation using a Pressure-Sensitive Mat,” in2018 IEEE International Symposium on Medical Measurements and Applications (MeMeA), pp. 1–5, 2018

    work page 2018

  6. [6]

    Breathing signal combining for respiration rate estimation in smart beds,

    H. Azimi, S. Soleimani Gilakjani, M. Bouchard, S. Bennett, R. A. Goubran, and F. Knoefel, “Breathing signal combining for respiration rate estimation in smart beds,” in2017 IEEE International Symposium on Medical Measurements and Applications (MeMeA), pp. 303–307, 2017

    work page 2017

  7. [7]

    Foubert,Posture recognition and postural transition detection using bed-based pressure sensor arrays

    N. Foubert,Posture recognition and postural transition detection using bed-based pressure sensor arrays. Master of Applied Science, Carleton University, Ottawa, Ontario, 2010

    work page 2010

  8. [8]

    Lying and sitting posture recognition and transition detection using a pressure sensor array,

    N. Foubert, A. M. McKee, R. A. Goubran, and F. Knoefel, “Lying and sitting posture recognition and transition detection using a pressure sensor array,” in2012 IEEE International Symposium on Medical Measurements and Applications Proceedings, pp. 1–6, 2012

    work page 2012

Show all 32 references
  1. [9]

    In- Bed Mobility Monitoring Using Pressure Sensors,

    S. Bennett, Z. Ren, R. Goubran, K. Rockwood, and F. Knoefel, “In- Bed Mobility Monitoring Using Pressure Sensors,”IEEE Transactions on Instrumentation and Measurement, vol. 64, no. 8, pp. 2110–2120, 2015

    2015

  2. [10]

    Detection of False Alarms in the NICU Using Pressure Sensitive Mat,

    D. G. Kyrollos, K. Greenwood, J. Harrold, and J. R. Green, “Detection of False Alarms in the NICU Using Pressure Sensitive Mat,” in2021 IEEE Sensors Applications Symposium (SAS), pp. 1–5, 2021

    2021

  3. [11]

    Unobtru- sive Sleep Stage Identification Using a Pressure-Sensitive Bed Sheet,

    L. Samy, M.-C. Huang, J. J. Liu, W. Xu, and M. Sarrafzadeh, “Unobtru- sive Sleep Stage Identification Using a Pressure-Sensitive Bed Sheet,” IEEE Sensors Journal, vol. 14, no. 7, pp. 2092–2101, 2014

    2014

  4. [12]

    Sleep-Wake and Body Position Classification with Deep Learning using Pressure Sensor Mat Measurements,

    C. Green, M. Bouchard, R. Goubran, R. Robillard, C. Higginson, E. Lee, and F. Knoefel, “Sleep-Wake and Body Position Classification with Deep Learning using Pressure Sensor Mat Measurements,” in2023 IEEE International Symposium on Medical Measurements and Applications (MeMeA),...

    2023

  5. [13]

    Adult polysomnography,

    L. Smolley, “Adult polysomnography,” inEncyclopedia of Sleep and Circadian Rhythms (Second Edition)(C. A. Kushida, ed.), pp. 474–477, Oxford: Academic Press, second edition ed., 2023

    2023

  6. [14]

    Real-time eye tracking for the assessment of driver fatigue,

    J. Xu, J. Min, and J. Hu, “Real-time eye tracking for the assessment of driver fatigue,”Healthcare Technology Letters, vol. 5, pp. 54–58, Jan. 2018

    2018

  7. [15]

    Development of an EOG-Based Automatic Sleep-Monitoring Eye Mask,

    S.-F. Liang, C.-E. Kuo, Y .-C. Lee, W.-C. Lin, Y .-C. Liu, P.-Y . Chen, F.-Y . Cherng, and F. Shaw, “Development of an EOG-Based Automatic Sleep-Monitoring Eye Mask,”IEEE Transactions on Instrumentation and Measurement, vol. 64, pp. 1–1, 11 2015

    2015

  8. [16]

    EOGNET: A Novel Deep Learning Model for Sleep Stage Classification Based on Single- Channel EOG Signal,

    J. Fan, C. Sun, M. Long, C. Chen, and W. Chen, “EOGNET: A Novel Deep Learning Model for Sleep Stage Classification Based on Single- Channel EOG Signal,”Frontiers in Neuroscience, vol. 15, 2021

    2021

  9. [17]

    Sleep stage classification using Single-Channel EOG,

    M. M. Rahman, M. I. H. Bhuiyan, and A. R. Hassan, “Sleep stage classification using Single-Channel EOG,”Computers in Biology and Medicine, vol. 102, pp. 211–220, 2018

    2018

  10. [18]

    Enhancing Healthcare with EOG: A Novel Approach to Sleep Stage Classification,

    S. Maiti, S. K. Sharma, and R. S. Bapi, “Enhancing Healthcare with EOG: A Novel Approach to Sleep Stage Classification,” in2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 2305–2309, 2024

    2024

  11. [19]

    Full-coverage unobtrusive health monitoring of elders at homes,

    Z. Huang, J. Li, and Z. He, “Full-coverage unobtrusive health monitoring of elders at homes,”Internet of Things, vol. 26, p. 101182, 2024

    2024

  12. [20]

    DeepSleepNet: A Model for Automatic Sleep Stage Scoring Based on Raw Single-Channel EEG,

    A. Supratak, H. Dong, C. Wu, and Y . Guo, “DeepSleepNet: A Model for Automatic Sleep Stage Scoring Based on Raw Single-Channel EEG,” IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 25, p. 1998–2008, Nov. 2017

    1998

  13. [21]

    A review of deep learning-based information fusion techniques for multimodal medical image classification,

    Y . Li, M. E. H. Daho, P.-H. Conze, R. Zeghlache, H. L. Boit ´e, R. Tadayoni, B. Cochener, M. Lamard, and G. Quellec, “A review of deep learning-based information fusion techniques for multimodal medical image classification,” 2024

    2024

  14. [22]

    Imagebind: One Embedding Space To Bind Them All,

    R. Girdhar, A. El-Nouby, Z. Liu, M. Singh, K. V . Alwala, A. Joulin, and I. Misra, “Imagebind: One Embedding Space To Bind Them All,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 15180–15190, 2023

    2023

  15. [23]

    Attention Bottlenecks for Multimodal Fusion,

    A. Nagrani, S. Yang, A. Arnab, A. Jansen, C. Schmid, and C. Sun, “Attention Bottlenecks for Multimodal Fusion,” inAdvances in Neu- ral Information Processing Systems(M. Ranzato, A. Beygelzimer, Y . Dauphin, P. Liang, and J. W. Vaughan, eds.), vol. 34, pp. 14200– 14213, Curran ...

    2021

  16. [24]

    Semi-Supervised Disease Classification Based on Limited Medical Image Data,

    Y . Zhang, C. Li, L. Zhaoxia, and M. Li, “Semi-Supervised Disease Classification Based on Limited Medical Image Data,”IEEE Journal of Biomedical and Health Informatics, vol. PP, pp. 1–12, 01 2024

    2024

  17. [25]

    Advancing Automated Sleep Stage Identification Through Multimodal PSG-Based Multiview Analysis,

    Y . Lin, Z. Chen, L. Ruan, H. Luo, A. Pumir, and J. Xu, “Advancing Automated Sleep Stage Identification Through Multimodal PSG-Based Multiview Analysis,” in2024 IEEE International Symposium on Medical Measurements and Applications (MeMeA), pp. 1–6, 2024

    2024

  18. [26]

    Analysis of a sleep-dependent neuronal feedback loop: the slow- wave microcontinuity of the EEG,

    B. Kemp, A. Zwinderman, B. Tuk, H. Kamphuisen, and J. Oberye, “Analysis of a sleep-dependent neuronal feedback loop: the slow- wave microcontinuity of the EEG,”IEEE Transactions on Biomedical Engineering, vol. 47, no. 9, pp. 1185–1194, 2000

    2000

  19. [27]

    ViViT: A Video Vision Transformer,

    A. Arnab, M. Dehghani, G. Heigold, C. Sun, M. Lu ˇci´c, and C. Schmid, “ViViT: A Video Vision Transformer,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp. 6836–6846, October 2021

    2021

  20. [28]

    The Kinetics Human Action Video Dataset,

    W. Kay, J. Carreira, K. Simonyan, B. Zhang, C. Hillier, S. Vijaya- narasimhan, F. Viola, T. Green, T. Back, P. Natsev,et al., “The Kinetics Human Action Video Dataset,”arXiv preprint arXiv:1705.06950, 2017

    2017 arXiv

  21. [29]

    Aasm scoring manual updates for 2017 (version 2.4),

    R. Berry, R. Brooks, S. Harding, R. Lloyd, S. Quan, M. Troester, and B. Vaughn, “Aasm scoring manual updates for 2017 (version 2.4),” Journal of clinical sleep medicine : JCSM : official publication of the American Academy of Sleep Medicine, vol. 13, 04 2017

    2017

  22. [30]

    Imagenet-21k pretraining for the masses,

    T. Ridnik, E. Ben-Baruch, A. Noy, and L. Zelnik-Manor, “Imagenet-21k pretraining for the masses,”

  23. [31]

    Audio Set: An ontology and human-labeled dataset for audio events,

    J. F. Gemmeke, D. P. W. Ellis, D. Freedman, A. Jansen, W. Lawrence, R. C. Moore, M. Plakal, and M. Ritter, “Audio Set: An ontology and human-labeled dataset for audio events,” inProc. IEEE ICASSP 2017, (New Orleans, LA), 2017

    2017

  24. [32]

    PhysioBank, PhysioToolkit, and PhysioNet: Components of a New Research Resource for Complex Physiologic Signals,

    A. L. Goldberger, L. A. N. Amaral, L. Glass, J. M. Hausdorff, P. C. Ivanov, R. G. Mark, J. E. Mietus, G. B. Moody, C.-K. Peng, and H. E. Stanley, “PhysioBank, PhysioToolkit, and PhysioNet: Components of a New Research Resource for Complex Physiologic Signals,”Circulation, vol....

    2000

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.