Pith. sign in

REVIEW 3 major objections 3 minor 58 references

MSBraM: A Multi-scale Self-supervised Brain Foundation Model for Hierarchical EEG Dynamics Learning

T0 review · 3 major / 3 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read Explicitly modeling EEG at three temporal scales — via multi-scale discrete codes, parallel transformer branches, and curriculum masking — yields a self-supervised foundation model that outperforms single-scale pretrained baselines on 11 of

desk verdict MSBraM is a solid multi-scale extension of LaBraM with real empirical scope, but the headline generalization claim is undercut by pretraining/downstream overlap that the paper never acknowledges. read the letter →

arxiv 2607.21402 v1 pith:AGKUEC3M submitted 2026-07-23 cs.AI

Tao Zhou , Jing Han , Lingyu Shu , Zixing Zhang This is my paper
classification cs.AI
keywords EEGself-supervisedlearningfoundationmodelmulti-scalerepresentationvectorquantizationmaskedpredictiontransferbrain-computerinterface
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 fix a blind spot in EEG foundation models: they treat brain signals as a single temporal texture, whereas real EEG carries both millisecond events and second-scale rhythms. MSBraM is a two-stage pipeline that first vector-quantizes raw EEG into discrete codes at three temporal resolutions, then pretrains a multi-branch transformer to predict masked codes under a masking ratio that grows from 0.3 to 0.5. The authors report best performance on 11 of 12 public datasets spanning emotion, motor imagery, seizure, and regression tasks, with significant gains on seven. If the transfer numbers hold, the contribution is a concrete architectural recipe — multi-scale tokenization, aligned masking, and bidirectional fusion — that future EEG models can adopt. The main caveat is that the evaluation assumes downstream data are unseen; Appendix B places at least two downstream datasets inside the pretraining corpus.

What carries the argument

The carrying mechanism is the multi-scale neural tokenizer combined with the multi-branch MSBraMEncoder. Three scale-specific codebooks discretize the signal at fine, medium, and coarse temporal resolutions; three parallel transformer branches process those scales; a Bidirectional Feature Pyramid Network fuses the branches with learnable per-scale weights; and a spatially aligned curriculum mask forces the model to predict codes at all scales, with the masking ratio rising from 0.3 to 0.5 over training. The aligned masking prevents information leakage where fine-scale context would reveal coarse-scale answers, so the model must genuinely integrate local transients and long-range context.

What would settle it

Retrain MSBraM and the strongest baselines on the pretraining corpus with PhysioNet-MI and the error-related-potential source removed, then re-evaluate on those two downstream benchmarks; if the reported balanced-accuracy advantages shrink to noise, the generalization claim for those benchmarks collapses. A lighter check: trace subject or session identifiers across pretraining and downstream test splits to confirm whether any overlap exists.

Watch

Extended reading notes

Core claim

On the paper's own terms: MSBraM is a self-supervised brain foundation model that learns hierarchical EEG representations without labels. It claims that prior EEG pretraining approaches produce low-quality, homogeneous representations because they ignore multi-scale temporal dynamics, and that explicitly modeling these dynamics is critical for effective EEG foundation models. The model discretizes raw signals into codes drawn from three codebooks of 8,192 entries each, trains the encoder to reconstruct the Fourier spectrum at each scale, then pretrains through masked code prediction with a curriculum schedule. The authors claim state-of-the-art or highly competitive performance across 12 dat

Load-bearing premise

The transfer claim assumes downstream benchmarks are disjoint from the pretraining corpus, but Appendix B lists PhysioNet-MI and the error-related-potential source dataset among the 18 pretraining datasets, so at least two reported gains may reflect overlap rather than generalization.

Editorial extensions

If this is right

  • If multi-scale pretraining is the right inductive bias, then single-scale masked EEG modeling leaves a systematic gap that more data or larger models alone will not close.
  • The two-stage recipe — tokenizer, then masked multi-scale code prediction — transfers across classification and regression tasks, suggesting a shared discrete representation of EEG dynamics.
  • Curriculum masking from 0.3 to 0.5 outperforms fixed masking on three of four tested datasets, so the masking schedule is itself a meaningful design lever.
  • The gains are achieved at 2.9M parameters, indicating that explicit multi-scale structure can substitute for raw model scale in EEG foundation models.
  • A channel-wise spatial baseline still wins on one emotion dataset, implying that temporal multi-scale structure is necessary but not sufficient for all EEG tasks.

Reading between the lines

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

  • A decisive next experiment is to remove the overlapping pretraining datasets (PhysioNet-MI and the error-related-potential source from Appendix B) and re-measure downstream gains; the paper's appendix shows the overlap, so this is the cleanest check on the transfer claim.
  • If the multi-scale codes reflect real neural dynamics, one could test whether codebook usage per scale aligns with classical frequency bands or with event-level versus background states — a physiological validation the paper does not report.
  • The design suggests testable extensions: more than three scales, per-subject adaptive curriculum schedules, or applying the same multi-scale tokenizer to intracranial EEG, which the authors list as future work.
  • The competitive-but-not-winning results on fine-grained emotion suggest that adding explicit spatial or channel-wise inductive biases to the multi-scale temporal machinery might combine the strengths of both approaches.
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 / 3 minor

Summary. The paper proposes MSBraM, a multi-scale self-supervised EEG foundation model. The architecture consists of a multi-scale neural tokenizer with separate codebooks per temporal resolution, a multi-branch transformer encoder with BiFPN fusion, and a curriculum masking schedule that increases the masking ratio during pretraining. The model is pretrained on over 2,400 hours of EEG data following the LaBraM protocol and evaluated on 10 tasks across 12 public datasets, reporting balanced accuracy, kappa/F1, and correlation/RMSE. The authors report state-of-the-art or competitive results on 11 of 12 datasets, with ablations supporting the BiFPN fusion and curriculum masking components.

Significance. If the reported results survive scrutiny, the contribution is meaningful: explicitly modeling multi-scale temporal dynamics in a self-supervised EEG foundation model is a plausible route to better transferable representations. The paper also has concrete strengths: five-seed error bars are reported throughout; several baselines are publicly released checkpoints; the ablations compare architectural choices (BiFPN vs. HRNet/FPN, CurrMasking vs. FixedMasking); and the model is computationally modest (2.9M parameters, 1.16G FLOPs). The main weakness is that the central 'strong generalization and transferability' claim is not established for all downstream benchmarks because at least two downstream datasets overlap with the pretraining corpus.

major comments (3)
  1. [Section 4.1, Appendix B Table 9, Appendix C] The pretraining corpus in Table 9 explicitly includes PhysioNet-MI [29] (47.3h) and Inria BCI [21] (29.98h). Table 1 lists PhysioNet-MI as a downstream task and labels KaggleERN with the same reference [21]. No subject/recording de-duplication between pretraining and downstream splits is reported. Consequently, the gains highlighted for motor imagery and ERN in Section 4.2 and Tables 3 and 12 are not evidence of transfer to unseen data. The abstract's 'strong generalization and transferability' claim is therefore overstated for at least two of the twelve benchmarks. The authors must either demonstrate that no subject/recording overlap exists (e.g., by dataset version or subject ID audit) or re-run the evaluation on held-out data and report the corrected numbers.
  2. [Appendix B vs. Table 1 (TUH subsets)] The same overlap concern may extend to the TUH family: pretraining includes TUAR, TUEP, TUSZ, and TUSL (all from [23]), while TUEV and TUAB (also from [23]) are used as downstream tasks. Since these are subsets of the Temple University Hospital EEG corpus, there is a real risk that subjects or recordings appear in both pretraining and downstream evaluation. Without an explicit subject/record overlap audit for the TUH data, four additional downstream benchmarks are potentially contaminated. This must be resolved before the 11-of-12 claim can be taken at face value.
  3. [Section 4.3, Eq. (8), Table 5] The curriculum masking ablation compares CurrMasking (r0=0.3, rmax=0.5) only against FixedMasking at r=0.3. Because the curriculum schedule has a higher average masking ratio, the observed gains on TUEV, TUAB, and BCIC-2a could be partly due to the ratio change rather than to the curriculum property itself. A fair ablation should include fixed ratios of 0.4 and 0.5, or a fixed-ratio control with the same mean masking ratio as the schedule. As written, the ablation is not sufficient to isolate the benefit of the curriculum mechanism.
minor comments (3)
  1. [Section F / Table 14] The statistical significance description is inconsistent: the text says a two-sided z-test is used, while the Table 14 footnote says 'one-tailed student t-test.' Please reconcile and report the exact test and the full set of significance results.
  2. [Section 4.2 and Table 6] The claim of 'best performance across 11 datasets' is based on the aggregated score in Table 14, but several individual metrics are not best (e.g., FACED Balanced Accuracy, CHB-MIT AUCPR/AUROC). Please phrase the claim in terms of the chosen aggregate metric to avoid overgeneralization. Also, in the BiFPN ablation text, the comparison '0.5770 vs. 0.5751' is against HRNet, not FPN; the wording should be corrected.
  3. [General] Equation (6) has inconsistent sub/superscript placement (o_A,k vs o^{A,k}), and Eq. (7)'s mask expansion indexing should be checked for notational clarity. These are presentation issues but worth cleaning up.

Circularity Check

2 steps flagged · score 4.0 of 10

Pretraining corpus overlaps two downstream benchmarks (PhysioNet-MI and KaggleERN/Inria BCI), making part of the claimed generalization self-referential.

  1. other [Appendix B Table 9 (Pretraining Datasets); Table 1 and Appendix C (PhysioNet-MI downstream); Section 4.2 and Table 3]
    "Table 9: "PhysioNet-MI[29] 160Hz 64 47.3h"; Table 1: "PhysioNet-MI[29] 160 64 9,837 4 4-class"; Appendix C: "The split is subject-wise: subjects 1-69 for training, 70-88 for validation, and 89-109 for testing.""

    The pretraining corpus described in Table 9 includes PhysioNet-MI (47.3 h) as one of the 18 pretraining datasets, while Table 1 lists PhysioNet-MI as a downstream motor-imagery benchmark. Appendix C sets the downstream test split to subjects 89-109. The paper reports no exclusion of these test subjects/recordings from the pretraining corpus, so the model was pretrained on the same dataset it is later said to 'generalize' to. Consequently, the PhysioNet-MI results in Table 3 and their contribution to the 'strong generalization' claim in Section 4.2 are not evidence on unseen data; they reduce to inputs already seen during pretraining.

  2. other [Appendix B Table 9 (Pretraining Datasets); Table 1, Appendix C, and Table 12 (KaggleERN downstream); Section E.2]
    "Table 9: "Inria BCI[21] 600Hz 56 29.98h"; Table 1: "KaggleERN[21] 200 56 8,840 2 2-class"; Appendix C: "It provides an official split of 16 subjects for training and 10 for testing. ... The held-out test set (10 subjects) is used for final evaluation.""

    The pretraining corpus includes Inria BCI from reference [21] (29.98 h), and the downstream KaggleERN benchmark also cites reference [21], i.e., the same underlying data source. The downstream evaluation holds out 10 subjects from that same source for testing. Because the paper does not report removing those held-out subjects from the pretraining corpus, the ERN results in Table 12 and Section E.2 are self-referential: the 'held-out' test data are part of the pretraining input unless explicitly excluded, which the paper never states.

full rationale

MSBraM's internal derivation is not circular. Equation (6) is an FFT-reconstruction VQ objective and Eq. (9) is a masked-code cross-entropy loss; both are self-supervised and neither is defined in terms of downstream accuracies. The ablations in Tables 5-6 compare architectural choices on the same fixed data, and the baselines are external/previously released models, so those comparisons are legitimate. The only self-citation in the paper (reference [54], Tyee, cited for epilepsy detection) is not load-bearing. The circularity concern is confined to the evaluation claim. Table 9 places PhysioNet-MI and Inria BCI in the pretraining corpus, while Table 1 evaluates PhysioNet-MI and KaggleERN (same [21] source) downstream; Appendix C's subject-wise splits do not carve out the pretraining data, and no subject/record de-duplication is reported. Thus, for those benchmarks, the claimed generalization is at least partly self-referential: the model was pretrained on data later used as 'held-out' test data. A similar risk extends to TUEV/TUAB because four TUH subsets (TUAR/TUEP/TUSZ/TUSL) are pretraining datasets and TUEV/TUAB come from the same TUH corpus [23], again without an overlap audit. Section 4.4 lists limitations but does not acknowledge this overlap. Since 8-10 of the 12 downstream benchmarks remain genuinely external to the pretraining corpus and the architecture itself is not derived from its own results, the circularity is partial rather than global.

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

No new physical entities are introduced. The main free parameters are hyperparameters of the masking schedule and the codebook configuration; the key domain assumption is the validity of the LaBraM pretraining protocol and the disjointness of pretraining and downstream data, the latter of which is violated.

free parameters (2)
  • Curriculum masking schedule (r0, rmax, t0, tmax) = 0.3, 0.5, 5, 35
    Hand-chosen hyperparameters that control the masking difficulty ramp; the gains of CurrMasking over FixedMasking depend on these values (Table 8).
  • Number of scales and codebook sizes = 3 scales; 3 codebooks of 8192×64
    Architecture choices adopted without a principled selection procedure; codebook size per scale is fixed at 8192 (Table 7).
assumptions (3)
  • domain assumption LaBraM architecture and pretraining protocol are a valid base
    Section 3 states "Following the LaBraM architecture" and reuses LaBraM's curated 2,400h corpus; any failure of the base assumption transfers to MSBraM.
  • domain assumption Downstream datasets are disjoint from the pretraining corpus
    Violated: PhysioNet-MI and Inria BCI/KaggleERN appear in both Table 9 (pretraining) and Table 1 (downstream).
  • domain assumption FFT-spectrum reconstruction is a useful pretext for EEG SSL
    Borrowed from LaBraM; the tokenizer loss (Eq. 6) regresses magnitude and phase of the FFT, but the paper does not demonstrate that this target is preferable to time-domain reconstruction.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MSBraM: A Multi-scale Self-supervised Brain Foundation Model for Hierarchical EEG Dynamics Learning." pith.science (2026). https://pith.science/paper/AGKUEC3M

@misc{pith2026260721402,
  author       = {Pith},
  title        = {Pith review of: MSBraM: A Multi-scale Self-supervised Brain Foundation Model for Hierarchical EEG Dynamics Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AGKUEC3M}},
  note         = {Machine review of arXiv:2607.21402}
}
read the original abstract

Self-supervised foundation models have recently shown strong potential for electroencephalogram (EEG)-based analysis. However, existing approaches struggle to capture the inherently multi-scale temporal structure of EEG signals, where local neural patterns and long-range dependencies jointly encode task-relevant information. This limitation hampers cross-scale representation learning and generalization across diverse downstream tasks. To address this challenge, we propose MSBraM, a Multi-Scale self-supervised Brain foundation Model designed to learn hierarchical EEG representations. MSBraM follows a two-stage pretraining framework. First, a multi-scale neural tokenizer discretizes raw EEG signals into semantic codes at different temporal resolutions via vector-quantized reconstruction. Second, the model is pretrained to predict masked codes using a curriculum multi-scale masking strategy, progressively integrating fine-grained local patterns with global temporal context. We pretrain MSBraM on over 2,400 hours of EEG data and evaluate it across 10 downstream tasks on 12 public datasets. Extensive experiments show that MSBraM achieves superior performance on other state-of-the-art pretrained models, demonstrating strong generalization and transferability. These results indicate that explicitly modeling multi-scale temporal dynamics is critical for effective EEG foundation models.

Figures

Figures reproduced from arXiv: 2607.21402 by the authors.

Figure 1
Figure 1. The pretraining pipeline of MSBraM. It contains two stages: (a) Multi-scale Neural Tokenizer Training, i.e., the multi-scale tokenizer learns scale-specific codebooks by reconstructing the Fourier spectrum, discretizing signals into tokens across temporal resolutions; and (b) Multi-scale Masked EEG Modeling Training, i.e., the encoder is pre-trained via masked prediction, driven by our curriculum multi-scale masking… view at source ↗
Figure 2
Figure 2. Architecture of MSBraMEncoder. It comprises several stacked stages, with each stage utilizing a downsample module to produce a coarser scale, a multi-branch Transformer for represen￾tation modeling, and a fusion module (FuseLayer) for cross-scale fusion. spatial positional embeddings {𝑒 𝑐 1 , 𝑒𝑐 2 , ..., 𝑒𝑐 𝐶 }, where 𝑇𝑚𝑎𝑥 > 𝑁 de￾notes the maximum sequence length. For each patch located at temporal index 𝑖 and elect… view at source ↗
Figure 3
Figure 3. Radar plots comparing MSBraM (red) with other latest foundation models across 10 tasks and 12 datasets in terms of bal￾anced accuracy for classification tasks and Pearson’s correlation for regression tasks. where 𝑟0 and 𝑟𝑚𝑎𝑥 are the initial and maximum masking ratios, while 𝑡0 is the warm-up epoch, and 𝑡𝑚𝑎𝑥 is the total training epoch. Loss The objective of multi-scale masked EEG modeling is to predict discrete patc… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

58 extracted references · 1 canonical work pages

  1. [21]

    Perrin Margaux, Maby Emmanuel, Daligault Sébastien, Bertrand Olivier, and Mattout Jérémie. 2012. Objective and subjective evaluation of online error correction during P300-based spelling.Advances in Human-Computer Interaction 2012, 1 (2012), 578295

    2012

  2. [29]

    Gerwin Schalk, Dennis J McFarland, Thilo Hinterberger, Niels Birbaumer, and Jonathan R Wolpaw. 2004. BCI2000: a general-purpose brain-computer interface (BCI) system.IEEE Transactions on Biomedical Engineering51, 6 (2004), 1034– 1043

    2004

  3. [23]

    Iyad Obeid and Joseph Picone. 2016. The temple university hospital EEG data corpus.Frontiers in Neuroscience10 (2016), 196

    2016

  4. [1]

    Benjamin Blankertz, Guido Dornhege, Matthias Krauledat, Klaus-Robert Müller, and Gabriel Curio. 2007. The non-invasive Berlin brain–computer interface: fast acquisition of effective performance in untrained subjects.NeuroImage37, 2 (2007), 539–550

    2007

  5. [2]

    Jingjing Chen, Xiaobin Wang, Chen Huang, Xin Hu, Xinke Shen, and Dan Zhang

  6. [3]

    Paolo Detti, Giampaolo Vatti, and Garazi Zabalo Manrique de Lara. 2020. EEG synchronization analysis for seizure prediction: A study on data of noninvasive recordings.Processes8, 7 (2020), 846

    2020

  7. [4]

    Navid Mohammadi Foumani, Geoffrey Mackellar, Soheila Ghane, Saad Irtza, Nam Nguyen, and Mahsa Salehi. 2024. EEG2Rep: Enhancing self-supervised EEG representation through informative masked inputs. InProc. ACM SIGKDD Conference on Knowledge Discovery (KDD). 5544–5555

    2024

  8. [5]

    Golnaz Ghiasi, Tsung-Yi Lin, and Quoc V. Le. 2019. NAS-FPN: Learning scalable feature pyramid architecture for object detection. InProc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 7036–7045

    2019

Show all 58 references
  1. [6]

    Yongtian He, Trieu Phat Luu, Kevin Nathan, Sho Nakagome, and Jose L Contreras- Vidal. 2018. A mobile brain-body imaging dataset recorded during treadmill walking with a brain-computer interface.Scientific Data5, 1 (2018), 1–10

    2018

  2. [7]

    Weibang Jiang, Liming Zhao, and Baoliang Lu. 2024. Large brain model for learning generic representations with tremendous EEG data in BCI. InProc. International Conference on Learning Representations (ICLR)

    2024

  3. [8]

    Jin Jing, Wendong Ge, Shenda Hong, Marta Bento Fernandes, Zhen Lin, Chaoqi Yang, Sungtae An, Aaron F Struck, Aline Herlopian, Ioannis Karakis, et al. 2023. Development of expert-level classification of seizures and rhythmic and periodic patterns during EEG interpretation.Neuro...

    2023

  4. [9]

    2019.Brain invaders calibration-less P300-based BCI with modulation of flash duration dataset (bi2015a)

    Louis Korczowski, Martine Cederhout, Anton Andreev, Grégoire Cattan, Pedro Luiz Coelho Rodrigues, Violette Gautheret, and Marco Congedo. 2019.Brain invaders calibration-less P300-based BCI with modulation of flash duration dataset (bi2015a). Ph. D. Dissertation. GIPSA-lab

    2019

  5. [10]

    Demetres Kostas, Stephane Aroca-Ouellette, and Frank Rudzicz. 2021. BENDR: Using transformers and a contrastive self-supervised learning task to learn from massive amounts of EEG data.Frontiers in Human Neuroscience15 (2021), 653659

    2021

  6. [11]

    Vernon J Lawhern, Amelia J Solon, Nicholas R Waytowich, Stephen M Gordon, Chou P Hung, and Brent J Lance. 2018. EEGNet: A compact convolutional neural network for EEG-based brain–computer interfaces.Journal of Neural Engineering 15, 5 (2018), 056013

    2018

  7. [12]

    Hongli Li, Man Ding, Ronghua Zhang, and Chunbo Xiu. 2022. Motor imagery EEG classification algorithm based on CNN-LSTM feature fusion network.Biomedical Signal Processing and Control72 (2022), 103342

    2022

  8. [13]

    Girshick, Kaiming He, Bharath Hariharan, and Serge J

    Tsung-Yi Lin, Piotr Dollár, Ross B. Girshick, Kaiming He, Bharath Hariharan, and Serge J. Belongie. 2017. Feature pyramid networks for object detection. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 936–944

    2017

  9. [14]

    Chenyu Liu, Xinliang Zhou, Jiaping Xiao, Zhengri Zhu, Liming Zhai, Ziyu Jia, and Yang Liu. 2024. VSGT: Variational spatial and gaussian temporal graph models for EEG-based emotion recognition. InProc. International Joint Conference on Artificial Intelligence (IJCAI). 3078–3086

    2024

  10. [15]

    Qile Liu, Weishan Ye, Lingli Zhang, and Zhen Liang. 2025. EEG-SCMM: Soft contrastive masked modeling for cross-corpus EEG-based emotion recognition. InProc. ACM International Conference on Multimedia (ACM MM). 5834–5842

    2025

  11. [16]

    Shu Liu, Lu Qi, Haifang Qin, Jianping Shi, and Jiaya Jia. 2018. Path aggregation network for instance segmentation. InProc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 8759–8768

    2018

  12. [17]

    Wei Liu, Jie-Lin Qiu, Wei-Long Zheng, and Bao-Liang Lu. 2021. Comparing recognition performance and robustness of multimodal deep learning models for multimodal emotion recognition.IEEE Transactions on Cognitive and Devel- opmental Systems14, 2 (2021), 715–729. MSBraM: A Multi...

    2021

  13. [18]

    Wei Liu, Wei-Long Zheng, Ziyi Li, Si-Yuan Wu, Lu Gan, and Bao-Liang Lu. 2022. Identifying similarities and differences in emotion recognition with EEG and eye movements among chinese, german, and french people.Journal of Neural Engineering19, 2 (2022), 026012

    2022

  14. [19]

    Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. 2022. A convnet for the 2020s. InProc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 11976–11986

    2022

  15. [20]

    Matthew D Luciw, Ewa Jarocka, and Benoni B Edin. 2014. Multi-channel EEG recordings during 3,936 grasp and lift trials with varying weight and friction. Scientific Data1, 1 (2014), 1–11

    2014

  16. [22]

    Wajid Mumtaz. 2016. MDD Patients and Healthy Controls EEG Data (New). https://figshare.com/articles/dataset/EEG_Data_New/4244171. doi:10.6084/m9. figshare.4244171.v2

    2016

  17. [24]

    Yue Pan, Cunbo Li, Peiyang Li, Fali Li, Feng Wan, Dezhong Yao, Zehong Cao, and Peng Xu. 2025. Real-Time EEG Emotion Recognition from Dynamic Mixed Spatiotemporal Graph Learning. InProc. ACM International Conference on Multi- media (ACM MM). 5697–5706

    2025

  18. [25]

    Wei Yan Peh, Yuanyuan Yao, and Justin Dauwels. 2022. Transformer convolu- tional neural networks for automated artifact detection in scalp EEG. InProc. Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC). 3599–3602

    2022

  19. [26]

    Huy Phan, Fernando Andreotti, Navin Cooray, Oliver Y Chén, and Maarten De Vos. 2019. SeqSleepNet: end-to-end hierarchical recurrent neural network for sequence-to-sequence automatic sleep staging.IEEE Transactions on Neural Systems and Rehabilitation Engineering27, 3 (2019), 400–410

    2019

  20. [27]

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. 2015. U-Net: Convolutional networks for biomedical image segmentation. InProc. Medical Image Computing and Computer-Assisted Intervention (MICCAI). 234–241

    2015

  21. [28]

    Arman Savran, Koray Ciftci, Guillame Chanel, Javier Cruz Mota, Luong Hong Viet, Bülent Sankur, Lale Akarun, Alice Caplier, and Michele Rombaut. 2006. Emotion detection in the loop from brain signals and facial images. InProc. eINTERFACE. 69–80

    2006

  22. [30]

    2009.Application of machine learning to epileptic seizure onset detection and treatment

    Ali Hossam Shoeb. 2009.Application of machine learning to epileptic seizure onset detection and treatment. Ph. D. Dissertation. Massachusetts Institute of Technology

    2009

  23. [31]

    Yonghao Song, Xueyu Jia, Lie Yang, and Longhan Xie. 2021. Transformer- based spatial-temporal feature learning for EEG decoding.arXiv preprint arXiv:2106.11170(2021)

    2021 arXiv

  24. [32]

    Yonghao Song, Qingqing Zheng, Bingchuan Liu, and Xiaorong Gao. 2022. EEG conformer: Convolutional transformer for EEG decoding and visualization.IEEE Transactions on Neural Systems and Rehabilitation Engineering31 (2022), 710–719

    2022

  25. [33]

    Mingxing Tan, Ruoming Pang, and Quoc V. Le. 2020. EfficientDet: Scalable and efficient object detection. InProc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 10778–10787

    2020

  26. [34]

    Michael Tangermann, Klaus-Robert Müller, Ad Aertsen, Niels Birbaumer, Christoph Braun, Clemens Brunner, Robert Leeb, Carsten Mehring, Kai J Miller, Gernot R Müller-Putz, et al. 2012. Review of the BCI competition IV.Frontiers in Neuroscience6 (2012), 55

    2012

  27. [35]

    Mastaneh Torkamani-Azar, Sumeyra Demir Kanik, Serap Aydin, and Mujdat Cetin. 2020. Prediction of reaction time and vigilance variability from spatio- spectral features of resting-state EEG in a long sustained attention task.IEEE Journal of Biomedical and Health Informatics24, ...

    2020

  28. [36]

    Logan Trujillo. 2020. Raw EEG Data. https://doi.org/10.18738/T8/SS2NHB. doi:10.18738/T8/SS2NHB

    2020 doi

  29. [37]

    Logan T Trujillo, Candice T Stanfield, and Ruben D Vela. 2017. The effect of electroencephalogram (EEG) reference choice on information-theoretic measures of the complexity and integration of EEG signals.Frontiers in Neuroscience11 (2017), 425

    2017

  30. [38]

    Christopher Wang, Vighnesh Subramaniam, Adam Uri Yaari, Gabriel Kreiman, Boris Katz, Ignacio Cases, and Andrei Barbu. 2023. BrainBERT: Self-supervised representation learning for intracranial recordings. InProc. International Confer- ence on Learning Representations (ICLR)

    2023

  31. [39]

    Guangyu Wang, Wenchao Liu, Yuhong He, Cong Xu, Lin Ma, and Haifeng Li. 2024. EEGPT: Pretrained transformer for universal and reliable representation of EEG signals. InProc. Advances in Neural Information Processing Systems (NeurIPS)

    2024

  32. [40]

    Jian Wang, Xiang Long, Guowei Chen, Zewu Wu, Zeyu Chen, and Errui Ding

  33. [41]

    Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, Wenyu Liu, and Bin Xiao. 2021. Deep high-resolution representation learning for visual recognition. IEEE Transactions on Pattern Analysis and Machine I...

    2021

  34. [42]

    Jiquan Wang, Sha Zhao, Zhiling Luo, Yangxuan Zhou, Haiteng Jiang, Shijian Li, Tao Li, and Gang Pan. 2025. CBraMod: A criss-cross brain foundation model for EEG decoding. InProc. International Conference on Learning Representations (ICLR)

    2025

  35. [43]

    Brandon Westover, and Jimeng Sun

    Chaoqi Yang, M. Brandon Westover, and Jimeng Sun. 2023. BIOT: Biosignal transformer for cross-data learning in the wild. InProc. Advances in Neural Information Processing Systems (NeurIPS), Vol. 36. 78240–78260

    2023

  36. [44]

    Brandon Westover, and Jimeng Sun

    Chaoqi Yang, Cao (Danica) Xiao, M. Brandon Westover, and Jimeng Sun. 2021. Self-supervised EEG Representation Learning for Automatic Sleep Staging.arXiv preprint arXiv:2110.15278(2021)

    2021 arXiv

  37. [45]

    Ke Yi, Yansen Wang, Kan Ren, and Dongsheng Li. 2023. Learning topology- agnostic EEG representations with geometry-aware modeling. InProc. Advances in Neural Information Processing Systems (NeurIPS)

    2023

  38. [46]

    Yuhui Yuan, Rao Fu, Lang Huang, Weihong Lin, Chao Zhang, Xilin Chen, and Jingdong Wang. 2021. HRFormer: High-resolution transformer for dense pre- diction. InProc. Advances in Neural Information Processing Systems (NeurIPS). 7281–7293

    2021

  39. [47]

    Zhizhang Yuan, Fanqi Shen, Meng Li, Yuguo Yu, Chenhao Tan, and Yang Yang

  40. [48]

    Daoze Zhang, Zhizhang Yuan, Junru Chen, Kerui Chen, and Yang Yang. 2024. Brant-X: A Unified Physiological Signal Alignment Framework. InProc. ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD). 4155–4166

    2024

  41. [49]

    Daoze Zhang, Zhizhang Yuan, Yang Yang, Junru Chen, Jingjing Wang, and Yafeng Li. 2023. Brant: Foundation model for intracranial neural signal. InProc. Advances in Neural Information Processing Systems (NeurIPS). 26304–26321

    2023

  42. [50]

    Gang Zhang, Ziyi Li, Chufeng Tang, Jianmin Li, and Xiaolin Hu. 2025. CEDNet: A cascade encoder–decoder network for dense prediction.Pattern Recognition 158 (2025), 111072

    2025

  43. [51]

    Wei-Long Zheng, Wei Liu, Yifei Lu, Bao-Liang Lu, and Andrzej Cichocki. 2018. EmotionMeter: A multimodal framework for recognizing human emotions.IEEE Transactions on Cybernetics(2018), 1–13

    2018

  44. [52]

    Wei-Long Zheng and Bao-Liang Lu. 2015. Investigating Critical Frequency Bands and Channels for EEG-based Emotion Recognition with Deep Neural Networks. IEEE Transactions on Autonomous Mental Development7, 3 (2015), 162–175

    2015

  45. [53]

    Wei-Long Zheng and Bao-Liang Lu. 2017. A multimodal approach to estimating vigilance using EEG and forehead EOG.Journal of Neural Engineering14, 2 (2017), 026017

    2017

  46. [54]

    Tao Zhou, Lingyu Shu, Zixing Zhang, and Jing Han. 2025. Tyee: A Unified, Modular, and Fully-Integrated Configurable Toolkit for Intelligent Physiological Health Care. InProc. ACM International Conference on Multimedia (ACM MM). 13628–13631

    2025

  47. [55]

    Igor Zyma, Sergii Tukaev, Ivan Seleznov, Ken Kiyono, Anton Popov, Mariia Chernykh, and Oleksii Shpenkov. 2019. Electroencephalograms during mental arithmetic task performance.Data4, 1 (2019), 14. Tao Zhou, Jing Han, Lingyu Shu, Zixing Zhang. Module Settings Parameters PatchEnc...

    2019

  48. [2022]

    UHRNet: Delving into improving semantic representation of high resolu- tion network for dense prediction.arXiv preprint arXiv:2210.07140(2022)

    2022 arXiv

  49. [2023]

    A large finer-grained affective computing EEG dataset.Scientific Data10, 1 (2023), 740

    2023

  50. [2024]

    Brainwave: A brain signal foundation model for clinical applications.arXiv preprint arXiv:2402.10251(2024)

    2024 arXiv

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.