Pith. sign in

REVIEW 2 major objections 33 references

TinyCNNDeep: Lightweight Attention-Based CNN for EEG Classification of Eye States and Sleep Deprivation

T0 review · 2 major / 0 minor · reviewed 2026-06-26 · grok-4.3

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

Pith's one-line read TinyCNNDeep classifies four EEG states of sleep condition and eye state at 83.69% subject-wise accuracy from five channels.

desk verdict TinyCNNDeep gets a big reported accuracy lift on the four-class EEG task but the five-channel image padding step looks like the weakest link. read the letter →

arxiv 2606.26506 v1 pith:C2PQBBBR submitted 2026-06-25 cs.HC

Thien Nhan Vo , Yen Nhi Tran Thi , Ngan Nguyen Xuan Phuong , Xuan-The Tran This is my paper · ORCID
classification cs.HC
keywords EEGclassificationsleepdeprivationeyestateconvolutionalneuralnetworkattentionmechanismlightweightmodelsleep-relatedfour-class
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 introduces TinyCNNDeep to jointly classify normal versus sleep-deprived states and eyes-open versus eyes-closed states from short EEG segments. It converts five-channel recordings into 224x224 grayscale images via per-channel normalization and padding, then applies a residual CNN with squeeze-and-excitation attention. On data from 35 subjects, the model reaches 83.69% mean accuracy, exceeding the best traditional baseline by 36 percentage points while other standard EEG networks perform near chance. The work shows that this image-based approach with attention yields an efficient solution for multiclass sleep-related EEG tasks under minimal electrode setups.

What carries the argument

TinyCNNDeep, a lightweight convolutional neural network that combines residual learning with a Squeeze-and-Excitation (SE) attention module applied to 224x224 grayscale images derived from five EEG channels.

What would settle it

Accuracy falling below 70% on an independent set of subjects recorded under different conditions or with altered electrode placements would show that the performance does not generalize beyond the original dataset.

Watch

Extended reading notes

Core claim

TinyCNNDeep attains a subject-wise mean accuracy of 83.69% on the four-class EEG classification task involving sleep condition and eye state, outperforming the strongest baseline by 36.03 percentage points while using only five channels and an image-based input representation with residual learning and channel attention.

Load-bearing premise

Converting short multi-channel EEG segments into 224x224 grayscale images through per-channel Z-score normalization, min-max scaling, and center padding allows 2D convolutions to jointly model inter-channel and temporal structure without critical information loss.

Editorial extensions

If this is right

  • The image-based representation paired with residual feature extraction and channel attention supplies an accurate and computationally efficient framework for multiclass sleep-related EEG classification under a minimal electrode configuration.
  • Per-subject analysis reveals substantial inter-subject variability in classification performance.
  • Confusion-matrix inspection shows that residual misclassifications concentrate between eyes-closed states across the two sleep conditions.
  • Three established EEG architectures operate near chance level on this four-class task.

Reading between the lines

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

  • The method's success suggests that standard 2D convolutions can capture both temporal dynamics and inter-channel relationships in EEG when the signal is rendered as an image.
  • The same conversion and architecture could be applied to other few-channel EEG problems such as fatigue detection or cognitive workload classification.
  • If the accuracy holds across sessions, the approach may support portable monitoring devices that require only five electrodes.
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

2 major / 0 minor

Summary. The manuscript introduces TinyCNNDeep, a lightweight CNN that integrates residual blocks with a Squeeze-and-Excitation attention module for four-class EEG classification (normal vs. sleep-deprived crossed with eyes-open vs. eyes-closed). Short segments from five channels (Fp1, Fp2, O1, Oz, O2) are converted to 224x224 grayscale images via per-channel Z-score normalization, min-max scaling, and center padding; the model is evaluated on a 35-subject dataset and reports a subject-wise mean accuracy of 83.69%, exceeding the strongest baseline (Random Forest on time-frequency features) by 36.03 points while standard EEG architectures perform near chance.

Significance. If the central result holds after validation, the work would demonstrate that an image-based representation paired with residual+SE blocks can deliver high multiclass accuracy on sleep-related EEG using only five electrodes and a computationally light model. The subject-wise analysis and confusion-matrix breakdown of error patterns between eyes-closed conditions are useful additions. The large margin over both the RF baseline and established EEG CNNs would be noteworthy if shown to arise from the proposed representation rather than implementation differences.

major comments (2)
  1. [Methods (EEG-to-image conversion)] Methods (EEG-to-image conversion paragraph): converting five-channel segments to 224x224 grayscale images via per-channel Z-score, min-max scaling, and center padding produces images in which the actual signal occupies only a small fraction of the canvas. This risks the 2D convolutions learning padding boundaries or normalized amplitude patterns instead of physiologically meaningful temporal or cross-channel structure; the absence of any ablation that isolates the representation step leaves the 83.69% accuracy claim dependent on an untested assumption.
  2. [Results (performance tables and baseline comparisons)] Results (performance tables and baseline comparisons): the near-chance accuracies reported for EEGNet, ShallowConvNet, and DeepConvNet versus TinyCNNDeep's 83.69% require explicit documentation of hyperparameter search, data partitioning (subject-wise vs. pooled), and cross-validation procedure. Without these details it is impossible to determine whether the 36-point gap reflects a genuine representational advantage or differences in training protocol or data leakage.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment below and outline the revisions we will make to strengthen the work.

read point-by-point responses

  1. Referee: [Methods (EEG-to-image conversion)] Methods (EEG-to-image conversion paragraph): converting five-channel segments to 224x224 grayscale images via per-channel Z-score, min-max scaling, and center padding produces images in which the actual signal occupies only a small fraction of the canvas. This risks the 2D convolutions learning padding boundaries or normalized amplitude patterns instead of physiologically meaningful temporal or cross-channel structure; the absence of any ablation that isolates the representation step leaves the 83.69% accuracy claim dependent on an untested assumption.

    Authors: We acknowledge the referee's valid concern that the central padding leaves the EEG signal occupying only a limited central region of the 224x224 canvas. The conversion procedure (per-channel Z-score normalization followed by min-max scaling to [0,255] and center padding with zeros) was chosen to produce a fixed-size input compatible with standard 2D CNN architectures while preserving the original temporal resolution and amplitude relationships across the five channels. The padding is constant (zero-valued) and does not vary with the data, reducing the likelihood of the model learning spurious boundary patterns; the residual blocks and SE module are intended to focus on the central signal content. Nevertheless, because no ablation isolating the representation step was performed, we will add such an ablation study in the revised manuscript (comparing center-padded images against zero-padded full-canvas versions, resized non-padded images, and a 1D CNN baseline on the raw segments) to demonstrate that the reported accuracy is driven by the signal content rather than padding artifacts. revision: yes

  2. Referee: [Results (performance tables and baseline comparisons)] Results (performance tables and baseline comparisons): the near-chance accuracies reported for EEGNet, ShallowConvNet, and DeepConvNet versus TinyCNNDeep's 83.69% require explicit documentation of hyperparameter search, data partitioning (subject-wise vs. pooled), and cross-validation procedure. Without these details it is impossible to determine whether the 36-point gap reflects a genuine representational advantage or differences in training protocol or data leakage.

    Authors: We agree that the current manuscript lacks sufficient detail on the training protocols for the baseline models. All experiments, including those for EEGNet, ShallowConvNet, DeepConvNet, and the Random Forest baseline, were performed under a strict subject-wise partitioning (no subject appears in both training and test sets) with 5-fold cross-validation across the 35 subjects. Hyperparameters for the deep models were selected via grid search on a held-out validation split within each training fold, using the same search space and early-stopping criteria applied to TinyCNNDeep. In the revised manuscript we will add a dedicated subsection (and supplementary table) that explicitly documents the hyperparameter grids, the exact subject-wise split procedure, the number of folds, and the training settings (optimizer, learning rate schedule, batch size, and regularization) for every model. This documentation will allow readers to verify that the performance gap is not attributable to differences in training protocol or data leakage. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity in empirical ML pipeline

full rationale

The paper reports measured classification accuracies from training and evaluating a CNN on EEG segments converted to images, with direct comparison to external baselines (Random Forest, EEGNet, etc.). No mathematical derivation, first-principles prediction, or fitted parameter is presented as an output; the 83.69% accuracy is an empirical result on held-out subject-wise data rather than a quantity forced by definition or self-citation. The image-conversion step is an explicit preprocessing choice whose validity is tested by the performance gap, not assumed by construction. No self-citation load-bearing steps, uniqueness theorems, or ansatz smuggling appear in the provided text.

Assumptions & free parameters 3 free parameters · 2 assumptions · 0 invented entities

Abstract-only review limits visibility into exact hyperparameters; typical CNN training choices (learning rate, batch size, layer counts) and the specific channel selection and image conversion steps function as free parameters. Domain assumptions about data representativeness and information preservation in the image transform are required but untested in the provided text.

free parameters (3)
  • selection of five EEG channels (Fp1, Fp2, O1, Oz, O2)
    Chosen as physiologically relevant; no justification or ablation shown in abstract.
  • 224x224 image size and center-padding rule
    Arbitrary but fixed choice to enable standard 2D CNN input.
  • per-channel Z-score then min-max scaling
    Specific normalization sequence chosen by authors.
assumptions (2)
  • domain assumption The 35-subject dataset is representative for generalization of the four-class accuracy claim.
    Subject-wise mean is presented as the key metric.
  • domain assumption Image conversion preserves inter-channel and temporal information needed for the classification task.
    Central premise enabling the 2D CNN approach.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TinyCNNDeep: Lightweight Attention-Based CNN for EEG Classification of Eye States and Sleep Deprivation." pith.science (2026). https://pith.science/paper/C2PQBBBR

@misc{pith2026260626506,
  author       = {Pith},
  title        = {Pith review of: TinyCNNDeep: Lightweight Attention-Based CNN for EEG Classification of Eye States and Sleep Deprivation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C2PQBBBR}},
  note         = {Machine review of arXiv:2606.26506}
}
read the original abstract

Sleep deprivation impairs vigilance and cognitive function, yet jointly identifying the sleep condition (normal vs deprived) and the eye state (open vs closed) from electroencephalography (EEG) remains underexplored. We address this four-class problem with TinyCNNDeep, a lightweight convolutional neural network that combines residual learning with a Squeeze-and-Excitation (SE) attention module. We convert short multi-channel EEG segments from five physiologically relevant channels (Fp1, Fp2, O1, Oz, O2) into 224x224 grayscale images through per-channel Z-score normalization, min-max scaling, and center padding, enabling 2D convolutions to jointly model inter-channel and temporal structure. On a 35-subject dataset recorded under normal-sleep and sleep-deprivation sessions, TinyCNNDeep attains a subject-wise mean accuracy of 83.69%, outperforming the strongest baseline (Random Forest with combined time-frequency features, 47.66%) by 36.03 percentage points, while three established EEG architectures (EEGNet, ShallowConvNet, DeepConvNet) operate near chance. Per-subject analysis quantifies inter-subject variability, and confusion-matrix inspection shows that residual misclassifications concentrate between eyes-closed states across sleep conditions. These results indicate that an image-based EEG representation paired with residual feature extraction and channel attention provides an accurate and computationally efficient framework for multiclass sleep-related EEG classification under a minimal electrode configuration.

Figures

Figures reproduced from arXiv: 2606.26506 by the authors.

Figure 2
Figure 2. It is designed as a lightweight yet effective model for [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 1
Figure 1. Overview of the proposed EEG data processing pipeline. The pipeline consists of six stages: (1) raw dataset acquisition from 71 subjects with [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Detailed architecture of the proposed TinyCNNDeep model. The model receives [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figures from the paper (2 more)
Figure 3
Figure 3. Figure 3: Confusion matrix for sub-14 (accuracy: 99.76%), representing a high [PITH_FULL_IMAGE:figures/full_fig_p006_3.png]
Figure 4
Figure 4. Figure 4: Confusion matrix for sub-5 (accuracy: 71.22%), illustrating inter-class [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 3 canonical work pages

  1. [1]

    M. P. Walker,Why We Sleep: Unlocking the Power of Sleep and Dreams. Scribner, 2017

    2017

  2. [2]

    The memory function of sleep,

    S. Diekelmann and J. Born, “The memory function of sleep,”Nature Reviews Neuroscience, vol. 11, no. 2, pp. 114–126, 2010

    2010

  3. [3]

    The global problem of insufficient sleep and its serious public health implications,

    V . K. Chattu, M. D. Manzar, S. Kumary, D. Burber, E. Acquah-Gyan, and S. R. Pandi-Perumal, “The global problem of insufficient sleep and its serious public health implications,”Healthcare, vol. 7, no. 1, p. 1, 2019

    2019

  4. [4]

    A meta-analysis of the impact of short-term sleep deprivation on cognitive variables,

    J. Lim and D. F. Dinges, “A meta-analysis of the impact of short-term sleep deprivation on cognitive variables,”Psychological Bulletin, vol. 136, no. 3, pp. 375–389, 2010

    2010

  5. [5]

    Effects of sleep deprivation on cognition,

    W. D. Killgore, “Effects of sleep deprivation on cognition,”Progress in Brain Research, vol. 185, pp. 105–129, 2010

    2010

  6. [6]

    R. B. Berry, R. Brooks, C. E. Gamaldo, S. M. Harding, R. M. Lloyd, C. L. Marcus, and B. V . Vaughn,The AASM Manual for the Scoring of Sleep and Associated Events. American Academy of Sleep Medicine, 2017

    2017

  7. [7]

    The two- process model of sleep regulation: A reappraisal,

    A. A. Borb ´ely, S. Daan, A. Wirz-Justice, and T. Deboer, “The two- process model of sleep regulation: A reappraisal,”Journal of Sleep Research, vol. 25, no. 2, pp. 131–143, 2016

    2016

  8. [8]

    EEG differences between eyes-closed and eyes-open resting conditions,

    R. J. Barry, A. R. Clarke, S. J. Johnstone, C. A. Magee, and J. A. Rushby, “EEG differences between eyes-closed and eyes-open resting conditions,”Clinical Neurophysiology, vol. 118, no. 12, pp. 2765–2773, 2007

    2007

Show all 33 references
  1. [9]

    Effect of total sleep deprivation on reaction time and waking EEG activity in man,

    I. Lorenzo, J. Ramos, C. Arce, M. A. Guevara, and M. Corsi-Cabrera, “Effect of total sleep deprivation on reaction time and waking EEG activity in man,”Sleep, vol. 18, no. 5, pp. 346–354, 1995

    1995

  2. [10]

    Alpha rhythm changes after sleep deprivation: Periodic and aperiodic contributions,

    K. Hoedlmoseret al., “Alpha rhythm changes after sleep deprivation: Periodic and aperiodic contributions,”Journal of Neurophysiology, vol. 131, pp. 520–532, 2024

    2024

  3. [11]

    Aperiodic and periodic EEG changes after sleep deprivation,

    T. Cassimet al., “Aperiodic and periodic EEG changes after sleep deprivation,”Journal of Sleep Research, vol. 33, no. 4, p. e14102, 2024

    2024

  4. [12]

    A comparative review on sleep stage classification methods in patients and healthy individuals,

    R. Boostani, F. Karimzadeh, and M. Nami, “A comparative review on sleep stage classification methods in patients and healthy individuals,” Computer Methods and Programs in Biomedicine, vol. 140, pp. 77–91, 2017

    2017

  5. [13]

    Sleep stage classification using EEG signal analysis: A compre- hensive survey and new investigation,

    K. A. Aboalayon, M. Faezipour, W. S. Almuhammadi, and S. Mosleh- pour, “Sleep stage classification using EEG signal analysis: A compre- hensive survey and new investigation,”Entropy, vol. 18, no. 9, p. 272, 2016. PREPRINT, 2026 9

    2016

  6. [14]

    Deep learning for electroen- cephalogram (EEG) classification tasks: A review,

    A. Craik, Y . He, D. Bhatt, and T. Bhatt, “Deep learning for electroen- cephalogram (EEG) classification tasks: A review,”Journal of Neural Engineering, vol. 16, no. 3, p. 031001, 2019

    2019

  7. [15]

    EEGNet: A compact convolutional neural network for EEG-based brain-computer interfaces,

    V . J. Lawhern, A. J. Solon, N. R. Waytowich, S. M. Gordon, C. P. Hung, and B. J. Lance, “EEGNet: A compact convolutional neural network for EEG-based brain-computer interfaces,”Journal of Neural Engineering, vol. 15, no. 5, p. 056013, 2018

    2018

  8. [16]

    Deep learning with convolutional neural networks for EEG decoding and visualization,

    R. T. Schirrmeister, J. T. Springenberg, L. D. J. Fiederer, M. Glasstetter, K. Eggensperger, M. Tangermann, F. Hutter, W. Burgard, and T. Ball, “Deep learning with convolutional neural networks for EEG decoding and visualization,”Human Brain Mapping, vol. 38, no. 11, pp. 5391–...

    2017

  9. [17]

    Benchmarking deep learning architectures for EEG-based brain-computer interfaces,

    L. Hernandez-Perezet al., “Benchmarking deep learning architectures for EEG-based brain-computer interfaces,”IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 32, pp. 1245–1256, 2024

    2024

  10. [18]

    Squeeze-and-excitation networks,

    J. Hu, L. Shen, and G. Sun, “Squeeze-and-excitation networks,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018, pp. 7132–7141

    2018

  11. [19]

    Multi-scale CNN with squeeze-and-excitation block for EEG-based motor imagery classification,

    M. Liet al., “Multi-scale CNN with squeeze-and-excitation block for EEG-based motor imagery classification,”Frontiers in Neuroscience, vol. 18, p. 1354869, 2024

    2024

  12. [20]

    Attention-enhanced convolutional neural networks for EEG-based brain-computer interfaces,

    Y . Zhanget al., “Attention-enhanced convolutional neural networks for EEG-based brain-computer interfaces,”IEEE Transactions on Biomedi- cal Engineering, vol. 72, no. 3, pp. 612–623, 2025

    2025

  13. [21]

    Learning representa- tions from EEG with deep recurrent-convolutional neural networks,

    P. Bashivan, I. Rish, M. Yeasin, and N. Codella, “Learning representa- tions from EEG with deep recurrent-convolutional neural networks,” in International Conference on Learning Representations (ICLR), 2016

    2016

  14. [22]

    EEG-to-image: A survey on deep learning approaches for visual EEG representation,

    Y . Wanget al., “EEG-to-image: A survey on deep learning approaches for visual EEG representation,”Neural Networks, vol. 174, p. 106234, 2024

    2024

  15. [23]

    Deep learning for sleep deprivation detection from EEG signals: A systematic review,

    H. Alsolaiet al., “Deep learning for sleep deprivation detection from EEG signals: A systematic review,”Computers in Biology and Medicine, vol. 171, p. 108120, 2024

    2024

  16. [24]

    Lightweight deep learning for sleep staging: A compre- hensive survey,

    H. Phanet al., “Lightweight deep learning for sleep staging: A compre- hensive survey,”IEEE Reviews in Biomedical Engineering, vol. 17, pp. 88–105, 2024

    2024

  17. [25]

    Transfer learning for cross-subject EEG classification: A review,

    D. Wuet al., “Transfer learning for cross-subject EEG classification: A review,”IEEE Transactions on Cognitive and Developmental Systems, vol. 16, no. 2, pp. 456–471, 2024

    2024

  18. [26]

    Inter- and intra-subject variability in EEG: A systematic survey,

    X.-T. Tran, T.-N. V o, S.-T. Vu, T.-T. Tran, M.-D. Nguyen, T. Do, and C.-T. Lin, “Inter- and intra-subject variability in EEG: A systematic survey,”arXiv preprint arXiv:2602.01019, 2026

    2026

  19. [27]

    EEG-SSM: Leveraging state-space model for dementia detection,

    X.-T. Tran, L. Le, Q. T. Nguyen, T. Do, and C.-T. Lin, “EEG-SSM: Leveraging state-space model for dementia detection,”arXiv preprint arXiv:2407.17801, 2024

    2024

  20. [28]

    EEG-Titans: Long-horizon seizure fore- casting via dual-branch attention and neural memory,

    T.-D. Pham and X.-T. Tran, “EEG-Titans: Long-horizon seizure fore- casting via dual-branch attention and neural memory,”arXiv preprint arXiv:2601.13748, 2026

    2026

  21. [29]

    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, no. 11, pp. 1998–2008, 2017

    1998

  22. [30]

    U-Sleep: Resilient high-frequency sleep staging,

    M. Perslev, S. Darkner, L. Kempfner, M. Nikolic, P. J. Jennum, and C. Igel, “U-Sleep: Resilient high-frequency sleep staging,”npj Digital Medicine, vol. 4, no. 1, p. 72, 2021

    2021

  23. [31]

    Batch normalization: Accelerating deep network training by reducing internal covariate shift,

    S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep network training by reducing internal covariate shift,” inProceedings of the International Conference on Machine Learning (ICML), 2015, pp. 448–456

    2015

  24. [32]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 770–778

    2016

  25. [33]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” International Conference on Learning Representations (ICLR), 2015

    2015

Pith tools

Reviewed June 26, 2026 · model on record in the stance chip above.