Pith. sign in

REVIEW 4 major objections 5 minor 3 cited by

An EEG foundation model that learns one channel at a time can outperform models trained on full electrode montages, if pretrained at scale and used as a frozen feature extractor.

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 ↗

T0 review · deepseek-v4-flash

2026-08-04 15:48 UTC pith:LYWQLVK5

load-bearing objection A genuinely new single-channel EEG foundation model with a hard-to-ignore empirical win, but the montage-flexibility headline outruns the experiments. the 4 major comments →

arxiv 2509.17920 v2 pith:LYWQLVK5 submitted 2025-09-22 cs.LG

SingLEM: Single-Channel Large EEG Model

Jamiyan Sukhbaatar , Satoshi Imamura , Ibuki Inoue , Shoya Murakami , Kazi Mahmudul Hassan , Seungwoo Han , Ingon Chanpornpakdi , Toshihisa Tanaka This is my paper
classification cs.LG
keywords EEG foundation modelsingle-channel representation learningmasked autoencoderself-supervised learningmotor imagery classificationcognitive task classificationmontage flexibilityfrozen feature extractor
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.

This paper tries to establish that the right way to build an EEG foundation model is to ignore electrode layouts entirely: learn one representation per channel, at massive scale, and let downstream tasks recombine channels by simple concatenation. The authors propose SingLEM, a self-supervised masked-autoencoder model pretrained on 71 public EEG datasets (about 357,000 single-channel hours), and show that its frozen per-channel features, fed to an SVM, achieve the best accuracy and macro F1 on all six tested motor-imagery and cognitive tasks compared with four multi-channel foundation models and handcrafted frequency-domain baselines. They further argue that this design is montage-flexible and hardware-agnostic, and that per-electrode analysis produces spatially interpretable maps consistent with known neurophysiology. A sympathetic reader would care because it suggests a data-efficient path to general EEG tools that work across the heterogeneous montages found in practice.

Core claim

The authors claim that a single-channel self-supervised foundation model, SingLEM, pretrained on 71 public EEG datasets (about 357,000 single-channel hours), produces reusable per-electrode representations. When these representations are simply concatenated and classified by an SVM, SingLEM achieves the highest accuracy and macro F1 in this comparison across six leave-one-subject-out motor-imagery and cognitive tasks, beating four multi-channel foundation models and handcrafted frequency-domain baselines. The paper further claims that excluding the three datasets underlying the downstream tasks from pretraining changes results negligibly, and that per-channel accuracy maps localize to cortic

What carries the argument

The carrying mechanism is a three-stage single-channel encoder: a three-layer 1D CNN with kernels sized to different EEG frequency bands, a local feature-embedding transformer that contextualizes each one-second token with its neighbors, and a 12-layer global transformer that integrates long-range sequence structure. Pretraining uses an asymmetric masked autoencoder that reconstructs both masked and unmasked tokens plus a 13–50 Hz spectral loss; downstream classification simply concatenates per-channel token representations and feeds them to an SVM. An end-of-channel token prevents information from leaking across channels during training.

Load-bearing premise

The advantage over multi-channel models rests on the untested premise that concatenating independently learned single-channel representations preserves all task-relevant spatial information; every downstream evaluation uses the full montage, so no experiment actually varies, removes, or reorders channels.

What would settle it

Take any of the six datasets and evaluate SingLEM with channels excluded or reordered, or with a montage not present in the 71-dataset pretraining corpus; if the SVM accuracy on concatenated per-channel features drops sharply, or falls below the handcrafted frequency-domain baseline when a single channel is removed, the montage-flexibility claim fails. A second check is to evaluate on a dataset with only two or three electrodes and compare against the same baseline.

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

If this is right

  • SingLEM works as a frozen feature extractor: a simple SVM on concatenated per-channel representations is enough to outperform the compared baselines, removing the need for task-specific fine-tuning.
  • Because each channel is embedded independently, the same pretrained model can in principle serve datasets with different electrode layouts, low-channel devices, or missing channels without retraining.
  • Per-electrode accuracy maps provide an interpretability tool that can reveal the spatial signature of a task, including confounds such as visual-cue processing during motor imagery.
  • The near-identical performance of models pretrained with and without the downstream datasets indicates that the learned features transfer across EEG paradigms, recording hardware, and subject populations.

Where Pith is reading between the lines

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

  • The montage-flexibility claim is testable in a way the paper does not test: systematically drop, reorder, or reduce electrodes at test time. If performance degrades gracefully, independent-channel pretraining truly recovers spatial information; if it collapses, the late-fusion premise is the weak link.
  • Independent per-channel representations make the reference electrode a hidden conditioning variable; combining SingLEM with reference-invariant normalization could improve cross-dataset comparability beyond what the paper reports.
  • The same per-channel, per-token representation could be reused for sequence-level tasks such as sleep staging or seizure prediction, and for automatic channel selection, since it already produces an interpretable per-electrode relevance score.
  • The pretraining losses include a 13–50 Hz spectral term, so the representation may be biased toward preserving beta and gamma content; this could be tested by comparing performance on tasks dominated by slower oscillations.

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 proposes SingLEM, a single-channel EEG foundation model based on a masked autoencoder. Each EEG channel is tokenized and processed independently by a CNN-and-transformer encoder; downstream, per-channel representations are concatenated and classified by an SVM. The model is pretrained on 71 public datasets (about 357,000 single-channel hours). It is evaluated as a frozen feature extractor on six leave-one-subject-out tasks (three motor-imagery, three cognitive) against BENDR, BIOT, LaBraM, CBraMod, and Fourier features. The reported results give SingLEM the highest mean accuracy and macro F1 on five tasks, with a near-tie on WG-2C, a large gap on most tasks, a leakage-controlled pretraining ablation, and a topographic single-channel analysis. The paper claims that this establishes single-channel self-supervised learning as a montage-flexible foundation for reusable EEG feature extraction.

Significance. If fully substantiated, the approach would be practically important: a single-channel pretrained encoder usable as a fixed feature extractor across heterogeneous montages would address a real limitation of current EEG foundation models. The strengths of the study are the large and diverse pretraining corpus, the leakage-controlled evaluation (the main model is pretrained after excluding the three source datasets), the strictly external downstream protocol, and the release of code and pretrained models. The evaluation is not circular: held-out subjects, held-out source datasets, and an SVM trained on features are used, with no downstream fitted constants reused as predictions. The topographic analysis is a useful post hoc validation. However, the central montage-flexibility claim is not directly tested by any experiment, and one reported baseline metric is implausible as written.

major comments (4)
  1. [§I, §II.C, §V.A, §VI.C] The central broader claim—that SingLEM is a montage-flexible, hardware-agnostic foundation model—is not supported by the experiments as reported. All six downstream tasks use the full montage of each dataset, and the SVM input dimension is montage-specific. No experiment varies channel count, removes channels, uses an unseen montage, or tests channel reordering. Table V shows that individual channels carry task information, but it does not establish that concatenating independently learned per-channel representations preserves the joint spatial structure that multi-channel models exploit. The paper itself concedes in §VI.C that the late-fusion strategy "does not explicitly capture spatial relationships." Please add experiments with channel subsets, unseen montages, or channel permutation, or explicitly narrow the claim to full-montage late fusion.
  2. [§V.A, Table III] No statistical significance testing is reported. On WG-2C, SingLEM (69.87±7.41) is only 0.51 percentage points above CBraMod (69.36±8.28), and the standard deviations overlap substantially across LOSO folds. Given the abstract's claim that SingLEM "consistently outperformed" the baselines, paired significance tests across subjects (e.g., Wilcoxon signed-rank or permutation tests) and effect sizes should be provided for all tables; otherwise the WG-2C claim should be softened.
  3. [Table II] The WBCIC-MI-2C row for Fourier(k=16) reports accuracy 70.25, F1 69.86, and Cohen's kappa 0.708. For a binary classifier, these values are mutually inconsistent: solving for the chance-agreement term gives p_e=(0.7025−0.708)/(1−0.708), which is negative and therefore impossible. This appears to be a computation or reporting error. Please audit all metrics in Tables II and III; an error in a baseline does not overturn the headline result, but it reduces confidence in the reported numbers.
  4. [Abstract vs. §V] The abstract states: "Additional classifier and subject-adapted analyses supported the robustness of its representations." No such analyses appear in Section V or VI: the downstream experiments use only the SVM, and no subject-adapted or alternate-classifier results are reported. Either add the missing experiments and results, or remove this sentence from the abstract.
minor comments (5)
  1. [§I] Typo: "Because of to its high temporal resolution" should read "Because of its high temporal resolution."
  2. [Table III] The header "Cohen'κ" is missing the 's' (should be "Cohen's κ").
  3. [§III.C] The end-of-channel (EOC) token is introduced to "prevent unintended mixing when concatenating multiple channels or recordings," but its use during pretraining and downstream feature extraction is not explained. Please clarify whether EOC tokens are used in both phases and how they are removed before late fusion.
  4. [Fig. 2] The topographic maps are min-max normalized per map, which can make the magnitude of channel differences across tasks difficult to compare. Please state this in the caption or use a common color scale for maps that are meant to be compared.
  5. [§VI.A] The contribution bullet claims SingLEM "substantially reduces computational costs," but no runtime, parameter count, or feature-extraction cost measurements are reported. A short efficiency comparison would support this claim.

Circularity Check

0 steps flagged

No significant circularity: evaluation is external (LOSO with source datasets excluded); conceded limitations are validity concerns, not circular derivation.

full rationale

The paper's central derivation chain is self-contained and externally benchmarked. SingLEM is pretrained on 71 public datasets, but the six downstream tasks are drawn from three source datasets that are excluded from the leakage-controlled model's pretraining corpus (68 datasets), and the ablation in Table IV shows that including those datasets changes results negligibly (e.g., Dreyer-MI-2C accuracy 75.27 vs. 75.33). Downstream evaluation uses strict leave-one-subject-out folds, with SVM hyperparameters tuned only on validation splits within the training folds, so no fitted constants are recycled as predictions. The single-channel topographic analysis (Table V, Fig. 2) is post hoc interpretation using prior anatomical knowledge and does not feed back into the reported accuracy/F1 comparisons. The main limitation explicitly acknowledged in Section VI.C — that the late-fusion strategy 'does not explicitly capture spatial relationships' — weakens the broader montage-flexibility claim (no unseen-montage or channel-dropout experiment is performed), but that is an unsupported-generalization concern, not circularity: the claim is not defined in terms of the result, and no load-bearing self-citation or imported uniqueness theorem is used. The findings therefore have independent content and are not forced by construction.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 1 invented entities

The central claim rests mostly on engineering choices and domain assumptions rather than on a closed-form derivation. The free parameters are architecture choices made by hand; the domain assumptions concern per-channel fusion and dataset distribution; no new physical or externally falsifiable entity is introduced apart from an internal EOC token.

free parameters (6)
  • Mask ratio p = 0.5
    Fixed mask proportion in §III.E; no ablation across mask ratios.
  • Loss weights lambda_1, lambda_2, lambda_3 = not reported
    Equation (17) defines the total loss, but the weights are never specified or ablated.
  • Tokenization length and overlap = 128 samples / 32-sample overlap
    §III.C sets 1 s tokens with 250 ms overlap; downstream features depend on this resolution.
  • Embedding dimensions (d, demb, D, r) = 128, 32, 128, 16
    §III.F sets model capacity and the final 16-dimensional token representation; no dimension ablation is reported.
  • Feature embedding window width w = 5
    Local context window in §III.F; no ablation for w is reported.
  • Global scaling factor = 1e4
    §III.B multiplies all data by 10^4; this amplitude normalization affects tokenizer inputs.
axioms (5)
  • domain assumption Per-channel independent encoding plus late concatenation preserves task-relevant spatial information
    Stated in §II.C and used in §V.A; no missing-channel or unseen-montage experiment tests it.
  • domain assumption Masked autoencoding of raw single-channel EEG learns general-purpose representations transferable without fine-tuning
    §III.E and §V.C; evidenced only on six MI and cognitive tasks, all broadly represented in pretraining by task type.
  • domain assumption Excluding the three source datasets is sufficient leakage control for a neutral downstream evaluation
    §V.C compares out-of-distribution and in-distribution variants, but the corpus still contains many other MI and cognitive datasets, so downstream tasks are not truly outside the task distribution.
  • domain assumption 128 Hz resampling and 1-second tokenization retain task-relevant EEG information
    §III.B-C applies fixed resampling and tokenization; no resolution or token-size ablation is reported.
  • standard math Transformer, convolution, and Huber loss behave as specified in the cited literature
    Background from [21] and [31] is assumed; no formal verification is provided.
invented entities (1)
  • End-of-channel (EOC) token no independent evidence
    purpose: Separates concatenated channels and recordings during pretraining so token sequences do not mix across channels.
    Introduced in §III.C; an internal model marker with no external falsifiable handle.

pith-pipeline@v1.3.0-alltime-deepseek · 22565 in / 15905 out tokens · 126928 ms · 2026-08-04T15:48:14.457852+00:00 · methodology

0 comments
read the original abstract

Current deep learning models for electroencephalography (EEG) are often task-specific and depend on large labeled datasets, limiting their adaptability. Although EEG foundation models seek broader applicability, many still rely on predefined multi-channel inputs, electrode-layout assumptions, or model-specific channel handling. To address these limitations, we introduce the Single-Channel Large EEG Model (SingLEM), a self-supervised foundation model whose hybrid convolutional--Transformer encoder maps each channel independently to a reusable representation capturing local and long-range temporal structure. These representations can be used individually or concatenated through late fusion. We assembled 71 public EEG datasets comprising approximately 9,200 subjects and 357,000 single-channel hours. For leakage-controlled evaluation, downstream results were obtained with a model pretrained on 68 datasets after excluding the three source datasets underlying the six tasks. A model pretrained on all 71 datasets is provided for general reuse. Across six motor imagery and cognitive tasks under strict leave-one-subject-out (LOSO) evaluation, the leakage-controlled model with concatenated representations and a support vector machine (SVM) classifier achieved the best overall performance among the compared pretrained and classical feature-based methods. Additional classifier and subject-adapted analyses supported the robustness of its representations. These findings support single-channel self-supervised learning as a montage-flexible foundation for reusable EEG feature extraction and electrode-level spatial analysis. The source code and pretrained models are available at https://github.com/ttlabtuat/SingLEM.

Figures

Figures reproduced from arXiv: 2509.17920 by Ibuki Inoue, Ingon Chanpornpakdi, Jamiyan Sukhbaatar, Kazi Mahmudul Hassan, Satoshi Imamura, Seungwoo Han, Shoya Murakami, Toshihisa Tanaka.

Figure 1
Figure 1. Figure 1: Architecture of SingLEM. The model comprises three main modules: the temporal encoder, the feature embedding module, and the global transformer [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Topographical accuracy maps for the Dreyer-MI-2C, WBCIC-MI-3C, WBCIC-MI-2C, N-back-2C, DSR-2C, and WG-2C datasets. Note: accuracy [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 3 Pith papers

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

  1. STEAM:ASpatio-TEmporal Alignment Mixture-of-Experts Model with Hierarchical Pre-training for EEG Decoding

    cs.CV 2026-08 conditional novelty 6.0

    STEAM, a dual-branch spatio-temporal mixture-of-experts EEG model with two-stage pre-training, reports the best average rank across seven downstream EEG datasets and fourteen evaluation settings.

  2. BandVQ: Band-Wise Vector-Quantized EEG Foundation Model

    cs.LG 2026-05 unverdicted novelty 6.0

    Band-wise independent VQ-VAE tokenizers per EEG frequency band plus masked Transformer pretraining on 9,200+ subjects yields top reported transfer on three cognitive tasks.

  3. EEG-FM-Compass: Progress, Benchmarking, and Future Directions for EEG Foundation Models

    cs.LG 2026-01 conditional novelty 6.0

    A unified benchmark of 12 EEG foundation models across 13 datasets finds specialists remain competitive and larger pre-trained models do not consistently improve downstream decoding.

Reference graph

Works this paper leans on

108 extracted references · 2 linked inside Pith · cited by 3 Pith papers

  1. [1]

    Brain–computer interfaces using sensorimotor rhythms: Current state and future perspectives,

    H. Yuanet al., “Brain–computer interfaces using sensorimotor rhythms: Current state and future perspectives,”IEEE Trans. Biomed. Eng., vol. 61, no. 5, pp. 1425–1435, 2014

    2014

  2. [2]

    DeepSleepNet: A model for automatic sleep stage scoring based on raw single-channel EEG,

    A. Suprataket al., “DeepSleepNet: A model for automatic sleep stage scoring based on raw single-channel EEG,”IEEE Trans. Neural Syst. Rehabil. Eng., vol. 25, no. 11, pp. 1998–2008, 2017

    1998

  3. [3]

    Automated detection of abnormalities from an EEG recording of epilepsy patients with a compact convolutional neural network,

    T. Shojiet al., “Automated detection of abnormalities from an EEG recording of epilepsy patients with a compact convolutional neural network,”Biomed. Signal Process. Control, vol. 70, p. 103013, 2021

    2021

  4. [4]

    EEG-based seizure onset detection of frontal and temporal lobe epilepsies using 1DCNN,

    X. Wanget al., “EEG-based seizure onset detection of frontal and temporal lobe epilepsies using 1DCNN,”IEEE Trans. Neural Syst. Rehabil. Eng., vol. 33, pp. 2263–2272, 2025

    2025

  5. [5]

    Real-time epileptic seizure prediction method with spatio-temporal information transfer learning,

    K. Menget al., “Real-time epileptic seizure prediction method with spatio-temporal information transfer learning,”IEEE J. Biomed. Health Inform., vol. 29, no. 3, pp. 2222–2232, 2025

    2025

  6. [6]

    Electroencephalography machine-learning features and methods for early diagnosis and classification of Parkin- son’s disease (2013–2023): A review,

    M. Emad-Ud-Dinet al., “Electroencephalography machine-learning features and methods for early diagnosis and classification of Parkin- son’s disease (2013–2023): A review,”IEEE Sensors J., vol. 25, no. 12, pp. 21 017–21 032, 2025

    2013

  7. [7]

    Machine-learning-based approaches for post- traumatic stress disorder diagnosis using video and EEG sensors: A review,

    A. Othmaniet al., “Machine-learning-based approaches for post- traumatic stress disorder diagnosis using video and EEG sensors: A review,”IEEE Sensors J., vol. 23, no. 20, pp. 24 135–24 151, 2023

    2023

  8. [8]

    Emotions recognition using EEG signals: A survey,

    S. M. Alarc ˜aoet al., “Emotions recognition using EEG signals: A survey,”IEEE Trans. Affect. Comput., vol. 10, no. 3, pp. 374–393, 2019

    2019

  9. [9]

    EmT: A novel transformer for generalized cross-subject EEG emotion recognition,

    Y . Dinget al., “EmT: A novel transformer for generalized cross-subject EEG emotion recognition,”IEEE Trans. Neural Netw. Learn. Syst., vol. 36, no. 6, pp. 10 381–10 393, 2025

    2025

  10. [10]

    EEG-based cross-subject emotion recognition using sparse Bayesian learning with enhanced covariance alignment,

    W. Wanget al., “EEG-based cross-subject emotion recognition using sparse Bayesian learning with enhanced covariance alignment,”IEEE Trans. Affect. Comput., vol. 16, no. 2, pp. 1190–1204, 2025

    2025

  11. [11]

    A review of classification algorithms for EEG-based brain–computer interfaces: a 10 year update,

    F. Lotteet al., “A review of classification algorithms for EEG-based brain–computer interfaces: a 10 year update,”J. Neural Eng., vol. 15, no. 3, p. 031005, 2018

    2018

  12. [13]

    Unsupervised domain adaptation with synchronized self-training for cross-domain motor imagery recognition,

    P. Chenet al., “Unsupervised domain adaptation with synchronized self-training for cross-domain motor imagery recognition,”IEEE J. Biomed. Health Inform., vol. 29, no. 5, pp. 3664–3677, 2025

    2025

  13. [14]

    Fine-grained spatial-frequency-time framework for motor imagery brain–computer interface,

    G. Liuet al., “Fine-grained spatial-frequency-time framework for motor imagery brain–computer interface,”IEEE J. Biomed. Health Inform., vol. 29, no. 6, pp. 4121–4133, 2025

    2025

  14. [15]

    Leveraging peripheral visual stimuli for enhanced SSVEP-based BCIs in fast calibration scenario,

    X. Heet al., “Leveraging peripheral visual stimuli for enhanced SSVEP-based BCIs in fast calibration scenario,”IEEE Sensors J., vol. 25, no. 10, pp. 17 683–17 695, 2025

    2025

  15. [16]

    A transformer-based deep neural network model for SSVEP classification,

    J. Chenet al., “A transformer-based deep neural network model for SSVEP classification,”Neural Netw., vol. 164, pp. 521–534, 2023

    2023

  16. [17]

    Advancing sleep disorder diagnostics: A transformer- based EEG model for sleep stage classification and OSA prediction,

    C. Wanet al., “Advancing sleep disorder diagnostics: A transformer- based EEG model for sleep stage classification and OSA prediction,” IEEE J. Biomed. Health Inform., vol. 29, no. 2, pp. 878–886, 2025

    2025

  17. [18]

    SleepSatelightFTC: A lightweight and interpretable deep learning model for single-channel EEG-based sleep stage classifica- tion,

    A. Itoet al., “SleepSatelightFTC: A lightweight and interpretable deep learning model for single-channel EEG-based sleep stage classifica- tion,”IEEE Access, vol. 13, pp. 46 263–46 272, 2025

    2025

  18. [19]

    Characteristics of physical environments that en- hance learning: A systematic review of EEG-based empirical studies,

    Y .-K. Honget al., “Characteristics of physical environments that en- hance learning: A systematic review of EEG-based empirical studies,” J. Environ. Psychol., vol. 102, p. 102525, 2025

    2025

  19. [20]

    Enhanced cognitive load detection in air traffic control operators using EEG and a hybrid deep learning approach,

    Y . Zhouet al., “Enhanced cognitive load detection in air traffic control operators using EEG and a hybrid deep learning approach,”IEEE Access, vol. 13, pp. 12 127–12 137, 2025

    2025

  20. [21]

    BERT: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlinet al., “BERT: Pre-training of deep bidirectional transformers for language understanding,” inProc. Conf. North Amer. Chapter Assoc. Comput. Linguistics (NAACL-HLT), 2019, pp. 4171–4186

    2019

  21. [22]

    GPT-4 technical report,

    J. Achiamet al., “GPT-4 technical report,”arXiv preprint arXiv:2303.08774, 2023

    Pith/arXiv arXiv 2023

  22. [23]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiyet al., “An image is worth 16x16 words: Transformers for image recognition at scale,”arXiv preprint arXiv:2010.11929, 2020

    Pith/arXiv arXiv 2010

  23. [24]

    Large AI models in health informatics: Applications, challenges, and the future,

    J. Qiuet al., “Large AI models in health informatics: Applications, challenges, and the future,”IEEE J. Biomed. Health Inform., vol. 27, no. 12, pp. 6074–6087, 2023

    2023

  24. [25]

    Foundation model for advancing healthcare: Challenges, opportunities and future directions,

    Y . Heet al., “Foundation model for advancing healthcare: Challenges, opportunities and future directions,”IEEE Rev. Biomed. Eng., vol. 18, pp. 172–191, 2025

    2025

  25. [26]

    A comprehensive survey of foundation models in medicine,

    W. Khanet al., “A comprehensive survey of foundation models in medicine,”IEEE Rev. Biomed. Eng., pp. 1–22, 2025

    2025

  26. [27]

    BENDR: Using transformers and a contrastive self- supervised learning task to learn from massive amounts of EEG data,

    D. Kostaset al., “BENDR: Using transformers and a contrastive self- supervised learning task to learn from massive amounts of EEG data,” Front. Hum. Neurosci., vol. 15, p. 653659, 2021

    2021

  27. [28]

    Biot: Biosignal transformer for cross-data learning in the wild,

    C. Yanget al., “Biot: Biosignal transformer for cross-data learning in the wild,”Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), vol. 36, pp. 78 240–78 260, 2023

    2023

  28. [29]

    Large brain model for learning generic representations with tremendous EEG data in BCI,

    W. Jianget al., “Large brain model for learning generic representations with tremendous EEG data in BCI,” inProc. Int. Conf. Learn. 11 TABLE V ACCURACY RESULTS FOR INDIVIDUAL CHANNELS ACROSS THE FOLLOWING SIX DATASETS:DREYER-MI-2C(D1),WBCIC-MI-3C(D2),WBCIC-MI-2C (D3),N-BACK-2C(D4),DSR-2C(D5),AND WG-2C(D6). ADASH(–)INDICATES THAT THE CHANNEL WAS NOT PRESEN...

    2024

  29. [30]

    CBraMod: A criss-cross brain foundation model for EEG decoding,

    J. Wanget al., “CBraMod: A criss-cross brain foundation model for EEG decoding,” inProc. Int. Conf. Learn. Represent. (ICLR), 2025. [Online]. Available: https://openreview.net/forum?id=NPNUHgHF2w

    2025

  30. [31]

    Masked autoencoders are scalable vision learners,

    K. Heet al., “Masked autoencoders are scalable vision learners,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2022, pp. 16 000–16 009

    2022

  31. [32]

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

    R. T. Schirrmeisteret al., “Deep learning with convolutional neural networks for EEG decoding and visualization,”Hum. Brain Mapp., vol. 38, no. 11, pp. 5391–5420, 2017

    2017

  32. [33]

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

    V . J. Lawhernet al., “EEGNet: a compact convolutional neural network for EEG-based brain–computer interfaces,”J. Neural Eng., vol. 15, no. 5, p. 056013, 2018

    2018

  33. [34]

    Development of expert-level classification of seizures and rhythmic and periodic patterns during EEG interpretation,

    J. Jinget al., “Development of expert-level classification of seizures and rhythmic and periodic patterns during EEG interpretation,”Neurology, vol. 100, no. 17, pp. e1750–e1762, 2023

    2023

  34. [35]

    EEG conformer: Convolutional transformer for EEG decoding and visualization,

    Y . Songet al., “EEG conformer: Convolutional transformer for EEG decoding and visualization,”IEEE Trans. Neural Syst. Rehabil. Eng., vol. 31, pp. 710–719, 2023

    2023

  35. [36]

    EEG-deformer: A dense convolutional transformer for brain-computer interfaces,

    Y . Dinget al., “EEG-deformer: A dense convolutional transformer for brain-computer interfaces,”IEEE J. Biomed. Health Inform., vol. 29, no. 3, pp. 1909–1918, 2025

    1909

  36. [37]

    Deep learning-based electroencephalography analysis: a systematic review,

    Y . Royet al., “Deep learning-based electroencephalography analysis: a systematic review,”J. Neural Eng., vol. 16, no. 5, p. 051001, 2019

    2019

  37. [38]

    EEG datasets for motor imagery brain–computer interface,

    H. Choet al., “EEG datasets for motor imagery brain–computer interface,”GigaScience, vol. 6, no. 7, p. gix034, 2017

    2017

  38. [39]

    BCI2000: a general-purpose brain-computer interface (BCI) system,

    G. Schalket al., “BCI2000: a general-purpose brain-computer interface (BCI) system,”IEEE Trans. Biomed. Eng., vol. 51, no. 6, pp. 1034– 1043, 2004

    2004

  39. [40]

    Open access dataset for EEG+NIRS single-trial classi- fication,

    J. Shinet al., “Open access dataset for EEG+NIRS single-trial classi- fication,”IEEE Trans. Neural Syst. Rehabil. Eng., vol. 25, no. 10, pp. 1735–1745, 2017

    2017

  40. [41]

    A large electroencephalographic motor imagery dataset for electroencephalographic brain computer interfaces,

    M. Kayaet al., “A large electroencephalographic motor imagery dataset for electroencephalographic brain computer interfaces,”Sci. Data, vol. 5, no. 1, pp. 1–16, 2018

    2018

  41. [42]

    EEG dataset and OpenBMI toolbox for three BCI paradigms: An investigation into BCI illiteracy,

    M.-H. Leeet al., “EEG dataset and OpenBMI toolbox for three BCI paradigms: An investigation into BCI illiteracy,”GigaScience, vol. 8, no. 5, p. giz002, 2019

    2019

  42. [43]

    Motor imagery under distraction—an open access BCI dataset,

    S. Brandlet al., “Motor imagery under distraction—an open access BCI dataset,”Front. Neurosci., vol. 14, p. 566147, 2020

    2020

  43. [44]

    Open access dataset integrating EEG and fNIRS during Stroop tasks,

    Z. Chenet al., “Open access dataset integrating EEG and fNIRS during Stroop tasks,”Sci. Data, vol. 10, no. 1, p. 618, 2023

    2023

  44. [45]

    ChineseEEG: A Chinese linguistic corpora EEG dataset for semantic alignment and neural decoding,

    X. Mouet al., “ChineseEEG: A Chinese linguistic corpora EEG dataset for semantic alignment and neural decoding,”Sci. Data, vol. 11, no. 1, p. 550, 2024

    2024

  45. [46]

    Resting-state EEG data before and after cognitive activity across the adult lifespan and a 5-year follow-up,

    S. Getzmannet al., “Resting-state EEG data before and after cognitive activity across the adult lifespan and a 5-year follow-up,”Sci. Data, vol. 11, no. 1, p. 988, 2024

    2024

  46. [47]

    EEG and fNIRS datasets based on Stroop task during two weeks of high-altitude exposure in new immigrants,

    X. Jiet al., “EEG and fNIRS datasets based on Stroop task during two weeks of high-altitude exposure in new immigrants,”Sci. Data, vol. 11, no. 1, p. 350, 2024

    2024

  47. [48]

    Le petit prince hong kong (lpphk): Naturalistic fMRI and EEG data from older Cantonese speakers,

    M. Momenianet al., “Le petit prince hong kong (lpphk): Naturalistic fMRI and EEG data from older Cantonese speakers,”Sci. Data, vol. 11, no. 1, p. 992, 2024

    2024

  48. [49]

    A resting-state EEG dataset for sleep deprivation,

    C. Xianget al., “A resting-state EEG dataset for sleep deprivation,” Sci. Data, vol. 11, no. 1, p. 427, 2024

    2024

  49. [50]

    A mind-brain-body dataset of MRI, EEG, cognition, emotion, and peripheral physiology in young and old adults,

    A. Babayanet al., “A mind-brain-body dataset of MRI, EEG, cognition, emotion, and peripheral physiology in young and old adults,”Sci. Data, vol. 6, no. 1, pp. 1–21, 2019

    2019

  50. [51]

    The two decades brainclinics research archive for insights in neurophysiology (TDBRAIN) database,

    H. van Dijket al., “The two decades brainclinics research archive for insights in neurophysiology (TDBRAIN) database,”Sci. Data, vol. 9, no. 1, p. 333, 2022

    2022

  51. [52]

    An EEG & eye-tracking dataset of ALS patients & healthy people during eye-tracking-based spelling system usage,

    T. D. Ngoet al., “An EEG & eye-tracking dataset of ALS patients & healthy people during eye-tracking-based spelling system usage,”Sci. Data, vol. 11, no. 1, p. 664, 2024

    2024

  52. [53]

    Human EEG recordings for 1,854 concepts presented in rapid serial visual presentation streams,

    T. Grootswagerset al., “Human EEG recordings for 1,854 concepts presented in rapid serial visual presentation streams,”Sci. Data, vol. 9, no. 1, p. 3, 2022

    2022

  53. [54]

    PEARL-Neuro database: EEG, fMRI, health and lifestyle data of middle-aged people at risk of dementia,

    P. Dzianoket al., “PEARL-Neuro database: EEG, fMRI, health and lifestyle data of middle-aged people at risk of dementia,”Sci. Data, vol. 11, no. 1, p. 276, 2024

    2024

  54. [55]

    A large EEG dataset for studying cross-session variability in motor imagery brain-computer interface,

    J. Maet al., “A large EEG dataset for studying cross-session variability in motor imagery brain-computer interface,”Sci. Data, vol. 9, no. 1, p. 531, 2022

    2022

  55. [56]

    An EEG dataset of neural signatures in a competitive two-player game encouraging deceptive behavior,

    Y . Chenet al., “An EEG dataset of neural signatures in a competitive two-player game encouraging deceptive behavior,”Sci. Data, vol. 11, no. 1, p. 389, 2024

    2024

  56. [57]

    Multi-channel EEG recordings during a sustained- attention driving task,

    Z. Caoet al., “Multi-channel EEG recordings during a sustained- attention driving task,”Sci. Data, vol. 6, no. 1, p. 19, 2019

    2019

  57. [58]

    An open-access dataset of naturalistic viewing using simultaneous EEG-fMRI,

    Q. K. Telesfordet al., “An open-access dataset of naturalistic viewing using simultaneous EEG-fMRI,”Sci. Data, vol. 10, no. 1, p. 554, 2023

    2023

  58. [59]

    Multi-channel EEG recording during motor imagery of different joints from the same limb,

    X. Maet al., “Multi-channel EEG recording during motor imagery of different joints from the same limb,”Sci. Data, vol. 7, no. 1, p. 191, 2020

    2020

  59. [60]

    A multi-day and high-quality EEG dataset for motor imagery brain-computer interface,

    B. Yanget al., “A multi-day and high-quality EEG dataset for motor imagery brain-computer interface,”Sci. Data, vol. 12, no. 1, p. 488, 2025

    2025

  60. [61]

    A resource for assessing dynamic binary choices in the adult brain using EEG and mouse-tracking,

    K. Chenet al., “A resource for assessing dynamic binary choices in the adult brain using EEG and mouse-tracking,”Sci. Data, vol. 9, no. 1, p. 416, 2022

    2022

  61. [62]

    A large EEG database with users’ profile informa- tion for motor imagery brain-computer interface research,

    P. Dreyeret al., “A large EEG database with users’ profile informa- tion for motor imagery brain-computer interface research,”Sci. Data, vol. 10, no. 1, p. 580, 2023

    2023

  62. [63]

    A test-retest resting, and cognitive state EEG dataset during multiple subject-driven states,

    Y . Wanget al., “A test-retest resting, and cognitive state EEG dataset during multiple subject-driven states,”Sci. Data, vol. 9, no. 1, p. 566, 2022. 12

    2022

  63. [64]

    A large finer-grained affective computing EEG dataset,

    J. Chenet al., “A large finer-grained affective computing EEG dataset,” Sci. Data, vol. 10, no. 1, p. 740, 2023

    2023

  64. [65]

    Simultaneous acquisition of EEG and NIRS during cognitive tasks for an open access dataset,

    J. Shinet al., “Simultaneous acquisition of EEG and NIRS during cognitive tasks for an open access dataset,”Sci. Data, vol. 5, no. 1, pp. 1–16, 2018

    2018

  65. [66]

    Open multi-session and multi-task EEG cognitive dataset for passive brain-computer interface applications,

    M. F. Hinsset al., “Open multi-session and multi-task EEG cognitive dataset for passive brain-computer interface applications,”Sci. Data, vol. 10, no. 1, p. 85, 2023

    2023

  66. [67]

    Dataset of concurrent EEG, ECG, and behavior with multiple doses of transcranial electrical stimulation,

    N. Gebodhet al., “Dataset of concurrent EEG, ECG, and behavior with multiple doses of transcranial electrical stimulation,”Sci. Data, vol. 8, no. 1, p. 274, 2021

    2021

  67. [68]

    Thinking out loud, an open-access EEG-based BCI dataset for inner speech recognition,

    N. Nietoet al., “Thinking out loud, an open-access EEG-based BCI dataset for inner speech recognition,”Sci. Data, vol. 9, no. 1, p. 52, 2022

    2022

  68. [69]

    EEG dataset for RSVP and P300 speller brain-computer interfaces,

    K. Wonet al., “EEG dataset for RSVP and P300 speller brain-computer interfaces,”Sci. Data, vol. 9, no. 1, p. 388, 2022

    2022

  69. [70]

    ZuCo, a simultaneous EEG and eye-tracking resource for natural sentence reading,

    N. Hollensteinet al., “ZuCo, a simultaneous EEG and eye-tracking resource for natural sentence reading,”Sci. Data, vol. 5, no. 1, pp. 1–13, 2018

    2018

  70. [71]

    Lower limb motor imagery EEG dataset based on the multi-paradigm and longitudinal-training of stroke patients,

    Y . Liuet al., “Lower limb motor imagery EEG dataset based on the multi-paradigm and longitudinal-training of stroke patients,”Sci. Data, vol. 12, no. 1, p. 314, 2025

    2025

  71. [72]

    High-density EEG mobile brain/body imaging data recorded during a challenging auditory gait pacing task,

    J. Wagneret al., “High-density EEG mobile brain/body imaging data recorded during a challenging auditory gait pacing task,”Sci. Data, vol. 6, no. 1, p. 211, 2019

    2019

  72. [73]

    SAM 40: Dataset of 40 subject EEG recordings to monitor the induced-stress while performing Stroop color-word test, arithmetic task, and mirror image recognition task,

    R. Ghoshet al., “SAM 40: Dataset of 40 subject EEG recordings to monitor the induced-stress while performing Stroop color-word test, arithmetic task, and mirror image recognition task,”Data Brief, vol. 40, p. 107772, 2022

    2022

  73. [74]

    The Cuban Human Brain Mapping Project, a young and middle age population-based EEG, MRI, and cognition dataset,

    P. A. Valdes-Sosaet al., “The Cuban Human Brain Mapping Project, a young and middle age population-based EEG, MRI, and cognition dataset,”Sci. Data, vol. 8, no. 1, p. 45, 2021

    2021

  74. [75]

    HD-EEG for tracking sub-second brain dynamics during cognitive tasks,

    A. Mheichet al., “HD-EEG for tracking sub-second brain dynamics during cognitive tasks,”Sci. Data, vol. 8, no. 1, p. 32, 2021

    2021

  75. [76]

    An EEG motor imagery dataset for brain computer interface in acute stroke patients,

    H. Liuet al., “An EEG motor imagery dataset for brain computer interface in acute stroke patients,”Sci. Data, vol. 11, no. 1, p. 131, 2024

    2024

  76. [77]

    EEG dataset for the recognition of different emotions induced in voice-user interaction,

    G.-Y . Choiet al., “EEG dataset for the recognition of different emotions induced in voice-user interaction,”Sci. Data, vol. 11, no. 1, p. 1084, 2024

    2024

  77. [78]

    ANPHY-Sleep: an open sleep database from healthy adults using high-density scalp electroencephalogram,

    X. Weiet al., “ANPHY-Sleep: an open sleep database from healthy adults using high-density scalp electroencephalogram,”Sci. Data, vol. 11, no. 1, p. 896, 2024

    2024

  78. [79]

    An open dataset for human SSVEPs in the frequency range of 1-60 Hz,

    M. Guet al., “An open dataset for human SSVEPs in the frequency range of 1-60 Hz,”Sci. Data, vol. 11, no. 1, p. 196, 2024

    2024

  79. [80]

    A wavelet-based technique to predict treatment outcome for major depressive disorder,

    W. Mumtazet al., “A wavelet-based technique to predict treatment outcome for major depressive disorder,”PLoS ONE, vol. 12, no. 2, p. e0171409, 2017

    2017

  80. [81]

    High-density scalp electroencephalogram dataset during sensorimotor rhythm-based brain-computer interfacing,

    S. Iwamaet al., “High-density scalp electroencephalogram dataset during sensorimotor rhythm-based brain-computer interfacing,”Sci. Data, vol. 10, no. 1, p. 385, 2023

    2023

Showing first 80 references.