Pith. sign in

REVIEW 4 major objections 8 minor 3 cited by

CLEAN-MI: A Scalable and Efficient Pipeline for Constructing High-Quality Neurodata in Motor Imagery Paradigm

T0 review · 4 major / 8 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A four-stage preprocessing pipeline for motor-imagery EEG claims to raise leave-one-subject-out accuracy by filtering channels, screening subjects, and aligning distributions.

desk verdict Headline accuracy gains are mostly a selection-bias artifact and the alignment step is never actually tested, but the channel-template result survives a matched comparison and the paper is worth a serious revision. read the letter →

arxiv 2506.11830 v1 pith:IOYUW5NS submitted 2025-06-13 cs.CE cs.LG

Dingkun Liu , Zhu Chen , Dongrui Wu This is my paper · ORCID
classification cs.CEcs.LG
keywords motorimageryEEGbrain-computerinterfacedataconstructionpipelinechannelselectionsubjectscreeningEuclideanalignmentfoundationmodels
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

CLEAN-MI is a four-stage pipeline for turning raw, multi-source motor-imagery (MI) EEG recordings into standardized, higher-quality training data for brain-computer interface models. The paper's central claim is that combining 8–30 Hz band-pass filtering, a sensorimotor channel template, subject screening, and Euclidean alignment systematically removes irrelevant or low-quality data and reduces inter-subject variability, with consistent gains in classification performance. On Weibo2014, the full pipeline raises average leave-one-subject-out accuracy from 65.8% to 81.90%; on Cho2017, from 66.99% to 74.55%; on BNCI2015001, the channel template alone raises it from 73.74% to 75.68%. The motivation is that foundation models for MI-based BCIs need large, clean, homogeneous datasets, which public EEG collections do not currently provide.

What carries the argument

The load-bearing object is the pipeline itself, a fixed sequence of four transformations. A band-pass filter isolates the $\alpha$ (8–13 Hz) and $\beta$ (13–30 Hz) rhythms; a manually defined channel template keeps only electrodes over the FC, C, CP, and T regions that overlie or neighbour the sensorimotor cortex; subject screening trains a classifier per participant and drops anyone below 0.6 within-subject accuracy; and Euclidean alignment replaces each trial $X_i$ with $\bar{R}^{-1/2}X_i$, where $\bar{R}$ is the subject's mean covariance matrix, so every subject's second-order statistics become comparable. Time axes are also resampled and truncated to a fixed length, and a compact convolutional EEG classifier (EEGNet) is used for all screening and evaluation.

What would settle it

Recompute the leave-one-subject-out averages on Weibo2014 and Cho2017 while including the test accuracies of the subjects who were dropped by the 0.6 threshold, or sweep the threshold from 0.5 to 0.8; if the averages return to the raw-data levels, the subject-selection gains are an artifact of the exclusion rule.

Watch

Extended reading notes

Core claim

The paper argues that the obstacle to large-scale MI foundation models is not model capacity but data: EEG from different headsets and subjects is noisy, variably informative, and spatially incompatible. CLEAN-MI addresses this by retaining only the 8–30 Hz sensorimotor band, selecting EEG channels over frontal-central, central, centro-parietal, and temporal regions, excluding subjects whose within-subject classification accuracy falls below 0.6, and applying Euclidean alignment to whiten each subject's covariance. The evidence is leave-one-subject-out classification on three public benchmarks, where every processing step improves or maintains average accuracy and the combined pipeline gives the largest gains, while the channel template cuts computational cost by 50–70 percent.

Load-bearing premise

The reported gains from subject screening are computed only over the subjects who pass the 0.6 within-subject accuracy cutoff, so the pipeline's benefit assumes that excluding low-scoring subjects is a valid way to evaluate a data-cleaning step.

Editorial extensions

If this is right

  • On Weibo2014, channel template plus subject selection raises average leave-one-subject-out accuracy from 65.8% to 81.90%.
  • On Cho2017, the same combination raises average accuracy from 66.99% to 74.55%, while the channel template alone cuts computational cost by an estimated 50–70%.
  • On BNCI2015001, the channel template alone raises average accuracy from 73.74% to 75.68%.
  • Because the pipeline standardizes frequency band, channel set, trial length, and marginal distribution, its output is directly usable for pretraining MI foundation models.
  • Subject screening removes participants whose data are substantially degraded by inattention or artifacts, so the surviving set is a cleaner training pool.

Reading between the lines

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

  • The same four-stage scaffolding could be adapted to other BCI paradigms by replacing the 8–30 Hz band and the sensorimotor channel template with paradigm-appropriate choices; the paper motivates paradigm-specific foundation models but only instantiates the MI case.
  • The fixed 0.6 screening threshold is a natural hyperparameter to sweep: varying it would reveal the trade-off between the amount of data retained and the average accuracy of the surviving pool.
  • A direct stress test of the pipeline would be to pretrain one MI foundation model on all 18 datasets summarized in the paper with and without CLEAN-MI, and compare transfer performance; the paper reports per-dataset leave-one-subject-out results and lists cross-dataset pretraining as future work.
  • The within-subject accuracy screen could also serve as an inexpensive online calibration gate, flagging participants whose own data are hard to classify before long multi-session recordings are undertaken.
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

4 major / 8 minor

Summary. The paper proposes CLEAN-MI, a pipeline for cleaning and standardizing motor-imagery EEG datasets before pretraining foundation models. The pipeline consists of 8-30 Hz bandpass filtering, a fixed MI-relevant channel template, resampling and trial-length normalization, screening of 'expert' subjects based on within-subject accuracy with a 0.6 threshold, and Euclidean alignment (EA) of covariance matrices. The authors evaluate components on Weibo2014, Cho2017, and BNCI2015001 using EEGNet with leave-one-subject-out (LOSO) cross-validation, and report that subject selection plus channel template raises average LOSO accuracy from 65.8% to 81.90% on Weibo2014 and from 66.99% to 74.55% on Cho2017, along with reduced computational cost. They conclude that the pipeline consistently improves data quality and classification performance.

Significance. The paper addresses a real bottleneck—heterogeneous, noisy MI EEG data for foundation-model pretraining—and the proposed pipeline is straightforward and interpretable. The channel-template selection is grounded in sensorimotor physiology and yields a computational-efficiency gain (50-70% reduction on Cho2017) that is credible independent of the subject-screening issue. If the subject-screening benefit were established with a non-circular evaluation and the alignment component were actually tested, the pipeline could be a useful contribution to MI dataset construction. As it stands, however, the headline accuracy improvements are not a valid estimate of the pipeline's benefit on unseen subjects, so the significance of the claimed results is not yet demonstrated.

major comments (4)
  1. [Tables IV and V] Tables IV and V compare averages computed over different subject sets. In Table IV, the Raw Data row reports 65.8% over all 10 subjects, while the Subject Selection row reports 77.41% over only the six subjects retained after the 0.6 threshold; the parenthetical 74.75 is the raw-data average on those same six subjects. The headline improvement from 65.8% to 81.90% therefore includes an 8.95-point increase that is purely an arithmetic consequence of excluding S2, S3, S4, and S9. On the matched population, subject selection alone improves from 74.75% to 77.41% (+2.66 points), and the channel template accounts for most of the remaining gain (80.08% before subject selection on the same six subjects). Table V shows the same pattern: the raw-data parenthetical 72.99 is close to the Subject Selection average 73.64, so the claimed screening benefit is about 0.65 points, not the 66.99-to-73.64 or 66.99-to-74.55 gap suggested by the unmatched comparisons. The paper must report averages over the full population (or over a held-out set of unscreened subjects) and must present the matched baselines as the primary comparison.
  2. [Section V-A] The expert-subject selection is circular with respect to the LOSO evaluation used to validate it. The screening rule trains on each subject's own labels to decide whether to exclude that subject, and the LOSO evaluation then tests on the retained subjects. This does not estimate performance on the stated deployment scenario—training a foundation model that will face new, unscreened subjects—because low-accuracy subjects are removed from the test set by the same rule that is being evaluated. A correct protocol would apply screening only to the training subjects (e.g., using a calibration split) and evaluate on all held-out subjects, including those that would have been screened out; alternatively, the paper should report per-subject LOSO accuracy for excluded subjects. Without this, the 'consistent improvements' claim in Section V and the abstract is not supported.
  3. [Section III-F and Tables III-V] The marginal distribution alignment component described in Section III-F, Eqs. (1)-(2), is never evaluated in the experiments. Tables III-V contain rows for raw data, channel template, subject selection, and subject selection plus channel template, but no row adds Euclidean alignment (EA). The abstract and Section VII state that CLEAN-MI integrates marginal distribution alignment, so the experimental section must include an ablation row with EA (and ideally EA alone and EA combined with the other components). As written, the full CLEAN-MI pipeline as claimed has not been tested.
  4. [Section III-E and Tables IV-V] The 0.6 threshold in Section III-E is a free parameter, and no sensitivity analysis is provided. Since the number of excluded subjects (and therefore the reported averages) is a direct function of this threshold, the reader cannot tell whether the improvements are robust or an artifact of the specific cutoff. The paper should vary the threshold (e.g., 0.50, 0.55, 0.60, 0.65, 0.70) and report LOSO accuracy on a fixed evaluation population. Additionally, the statistical testing behind the p-values in Tables IV and V is unspecified (test type, number of runs, seeds), and Table V's footnote uses p<0.1 rather than p<0.05, so the significance claims should be re-stated with a described test and corrected for multiple comparisons.
minor comments (8)
  1. [Section III-C] The Weibo2014 channel template is described as containing 35 channels, but the enumerated list contains 36 channel names; please correct the count or remove one channel.
  2. [Table III] Table III's caption says 'using raw data and subject selection,' but the rows are labeled Raw Data and Channel Template; clarify whether subject selection was applied before the Channel Template row and why S7 is marked '—' in that row.
  3. [Section I] Section I states 'Section VI discusses the future work. Finally, Section VI draws conclusions,' but the manuscript has two separate sections (VI and VII); renumber the references.
  4. [Section IV] The sentence 'This section introduces 18 MI benchmark datasets their SOTA approaches' is grammatically incomplete and should be rewritten.
  5. [Section IV-4 and Table II] BNCI2014002 is described as having 13 participants in Section IV-4 but Table II lists 14 subjects; align these numbers.
  6. [References] Reference [2] (Guo et al.) is not cited in the text, and reference [3] duplicates reference [27] (both are Lawhern et al., EEGNet); remove or cite appropriately.
  7. [Figure 6] Figure 6 appears to contain corrupted path strings in the panel labels; please regenerate the figure.
  8. [Abstract] The abstract claims 'consistent improvements in data quality and classification performance,' but data quality is never directly measured (only classification accuracy as a proxy); consider stating this explicitly.

Circularity Check

1 steps flagged · score 6.0 of 10

Subject-screening gains are computed only on subjects retained by the 0.6 within-subject accuracy threshold, making part of the headline accuracy improvement a consequence of the exclusion rule.

  1. self definitional [Section V-A, Subject Selection; Tables IV and V]
    "Subjects with classification accuracies below threshold (typically set to 0.6) were excluded from further training."

    The same within-subject accuracy signal used to define expert subjects also determines who remains in the LOSO evaluation. In Table IV, the Raw Data average 65.8 is over all ten Weibo2014 subjects, while the parenthetical 74.75 is the average over the six subjects kept after excluding S2, S3, S4, and S9; the Subject Selection (77.41) and Subject Selection + Channel Template (81.90) rows are computed only on those six subjects. Table V repeats this pattern (66.99 vs. the parenthetical 72.99). The headline gains, 65.8 to 81.90 and 66.99 to 74.55, therefore compare a full-population baseline with a selected-subset result, so the subject-selection improvement is partly forced by the exclusion rule rather than being an independent empirical finding on a common test population.

full rationale

The channel template, frequency filtering, and Euclidean alignment components are not circular: channel-template comparisons on matched subjects (e.g., 80.08 vs. 74.75 on the six retained Weibo2014 subjects) are valid measurements against public data, and EA is a standard external method. The circularity is localized to the expert-subject-selection evaluation: subjects are excluded using their own labels, and the reported accuracy averages are then computed only over the retained subjects, with no accuracy reported for excluded subjects and no threshold sensitivity analysis. Because the central claim of consistent improvements relies on the subject-selection rows, the headline numbers are partially self-fulfilling. The paper therefore has partial circularity (score 6), not full circularity, since the channel-template results and the matched-subject comparisons retain independent content.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new entities are introduced; 'expert subjects' is a labeling of existing subjects. The central claimed improvement rests on a hand-chosen threshold, manually selected channel lists, and an untested alignment step.

free parameters (3)
  • Subject selection accuracy threshold = 0.6
    Subjects whose within-subject EEGNet accuracy is below 0.6 are excluded (Section V-A). The reported improvements depend directly on this hand-chosen threshold; no sensitivity analysis is provided.
  • Weibo2014 channel template (35 channels) = FT7, FC5, FC3, FC1, FCZ, FC2, FC4, FC6, FT8, T7, C5, C3, C1, Cz, C2, C4, C6, T8, TP7, CP5, CP3, CP1, CPz, CP2, CP4…
    Channel list selected by hand from spatial priors (Section V-A); affects all Weibo2014 results.
  • Cho2017 channel template (38 channels) = FT7, FC5, FC3, FC1, FCZ, FC2, FC4, FC6, FT8, T7, C5, C3, C1, Cz, C2, C4, C6, T8, TP7, CP5, CP3, CP1, CPz, CP2, CP4…
    Channel list selected by hand from spatial priors (Section V-A); affects all Cho2017 results.
assumptions (4)
  • domain assumption ERD/ERS modulations in the alpha (8-13 Hz) and beta (13-30 Hz) bands over sensorimotor cortex are the primary informative signals for motor imagery
    Invoked in Sections III-B and III-C to justify bandpass filtering and the FC/C/CP/T channel template.
  • ad hoc to paper Within-subject classification accuracy on the subject's own data is a valid measure of EEG data quality, and 0.6 is the correct cutoff for excluding noisy subjects
    Introduced in Section III-E and used in Section V-A; the cutoff is not derived from data or theory.
  • ad hoc to paper Leave-one-subject-out accuracy computed only over the retained expert subjects is the correct metric for the pipeline's benefit
    Section V-A and Tables III-V; this evaluates the pipeline only on the filtered set, so the subject-selection gain is by construction.
  • domain assumption Euclidean alignment reduces inter-subject distribution shift
    Section III-F cites prior work [4]; this component is described but never tested in the experiments.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CLEAN-MI: A Scalable and Efficient Pipeline for Constructing High-Quality Neurodata in Motor Imagery Paradigm." pith.science (2026). https://pith.science/paper/IOYUW5NS

@misc{pith2026250611830,
  author       = {Pith},
  title        = {Pith review of: CLEAN-MI: A Scalable and Efficient Pipeline for Constructing High-Quality Neurodata in Motor Imagery Paradigm},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IOYUW5NS}},
  note         = {Machine review of arXiv:2506.11830}
}
read the original abstract

The construction of large-scale, high-quality datasets is a fundamental prerequisite for developing robust and generalizable foundation models in motor imagery (MI)-based brain-computer interfaces (BCIs). However, EEG signals collected from different subjects and devices are often plagued by low signal-to-noise ratio, heterogeneity in electrode configurations, and substantial inter-subject variability, posing significant challenges for effective model training. In this paper, we propose CLEAN-MI, a scalable and systematic data construction pipeline for constructing large-scale, efficient, and accurate neurodata in the MI paradigm. CLEAN-MI integrates frequency band filtering, channel template selection, subject screening, and marginal distribution alignment to systematically filter out irrelevant or low-quality data and standardize multi-source EEG datasets. We demonstrate the effectiveness of CLEAN-MI on multiple public MI datasets, achieving consistent improvements in data quality and classification performance.

Figures

Figures reproduced from arXiv: 2506.11830 by the authors.

Figure 1
Figure 1. A closed-loop MI-based BCI system. Following the aforementioned pipeline, numerous special￾arXiv:2506.11830v1 [cs.CE] 13 Jun 2025 [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the CLEAN-MI pipeline. EEG signals colle [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. MI paradigm data collection process. Notably, EEG headsets, sampling rates, and trial durations vary across datasets depending on the recording hardware and experimental protocol; therefore, harmonizing these parame￾ters during the preprocessing stage is indispensable. B. Frequency Filtering The corresponding frequencies and their effects on behavior are summarized in Table I. Among these, the α (8–13 Hz) and β (13–… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Scalp topographies of SMR power changes during [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Schematic of the EEG headset electrode positions. [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 6
Figure 6. Figure 6: t-SNE visualization of the data in BNCI2014004. (a) Before EA; (b) After EA. Different colors represent trials from different subjects. 7) Cho2017 [15]: Cho2017 dataset includes EEG data from 52 subjects (19 females, mean age 24.8 ± 3.86 years) performing MI tasks for …

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

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

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

    cs.LG 2026-01 conditional novelty 6.0 of 10

    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.

  2. MIRepNet: A Pipeline and Foundation Model for EEG-Based Motor Imagery Classification

    cs.CV 2025-07 conditional novelty 4.0 of 10

    An EEG foundation model pretrained exclusively on motor-imagery data, using masked-reconstruction plus classification, reports state-of-the-art few-shot decoding on five MI benchmarks.

  3. The 1st International Workshop on Disentangled Representation Learning for Controllable Generation (DRL4Real): Methods and Results

    cs.LG 2025-08 unverdicted

    A workshop summary describing nine papers on disentangled representation learning applied to controllable generation, editing, video, 3D, EEG, and compression.

Reference graph

Works this paper leans on

41 extracted references · 30 canonical work pages · cited by 3 Pith papers

  1. [1]

    Brain computer i nterfaces, a review,

    L. F. Nicolas-Alonso and J. Gomez-Gil, “Brain computer i nterfaces, a review,” Sensors, vol. 12, no. 2, pp. 1211–1279, 2012

    2012

  2. [2]

    On calibr ation of modern neural networks,

    C. Guo, G. Pleiss, Y . Sun, and K. Q. Weinberger, “On calibr ation of modern neural networks,” in Int’l Conf. on Machine Learning , (Sydney, Austraia), pp. 1321–1330, Aug. 2017

    work page 2017

  3. [3]

    EEGNet: A compact convolutional neural net work 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 net work for EEG-based brain-computer interfaces,” Journal of Neural Engineering , vol. 15, no. 5, p. 056013, 2018

    2018

  4. [4]

    Transfer learning for brain-computer in terfaces: A euclidean space data alignment approach,

    H. He and D. Wu, “Transfer learning for brain-computer in terfaces: A euclidean space data alignment approach,” IEEE Trans. on Biomedical Engineering, vol. 67, no. 2, pp. 399–410, 2019

    work page 2019

  5. [5]

    Pr e-stimulus sensorimotor rhythms influence brain–computer interface c lassification performance,

    C. L. Maeder, C. Sannelli, S. Haufe, and B. Blankertz, “Pr e-stimulus sensorimotor rhythms influence brain–computer interface c lassification performance,” IEEE Trans. on neural systems and rehabilitation engi- neering, vol. 20, no. 5, pp. 653–662, 2012

    work page 2012

  6. [6]

    Filtering technique s for channel selection in motor imagery EEG applications: a survey,

    M. Z. Baig, N. Aslam, and H. P . Shum, “Filtering technique s for channel selection in motor imagery EEG applications: a survey,” Artificial Intelligence Review, vol. 53, no. 2, pp. 1207–1232, 2020

    work page 2020

  7. [7]

    Cross-dataset transfer learning for motor ima gery signal classification via multi-task learning and pre-training,

    Y . Xie, K. Wang, J. Meng, J. Y ue, L. Meng, W. Yi, T.-P . Jung, M. Xu, and D. Ming, “Cross-dataset transfer learning for motor ima gery signal classification via multi-task learning and pre-training,” Journal of Neural Engineering, vol. 20, no. 5, p. 056037, 2023

    work page 2023

  8. [8]

    Commande robuste d’un effecteur par une i nterface cerveau–machine eeg asynchrone,

    B. Alexandre, “Commande robuste d’un effecteur par une i nterface cerveau–machine eeg asynchrone,” Universit´ e de Grenoble, 2006

    work page 2006

Show all 41 references
  1. [9]

    Review of the BCI competition IV,

    M. Tangermann, K.-R. M¨ uller, A. Aertsen, N. Birbaumer, C. Braun, C. Brunner, R. Leeb, C. Mehring, K. J. Miller, G. R. M¨ uller-P utz, et al., “Review of the BCI competition IV,” Frontiers in neuroscience, vol. 6, p. 55, 2012

    2012

  2. [10]

    Brain–computer communication: motiva tion, aim, and impact of exploring a virtual apartment,

    R. Leeb, F. Lee, C. Keinrath, R. Scherer, H. Bischof, and G. Pfurtscheller, “Brain–computer communication: motiva tion, aim, and impact of exploring a virtual apartment,” IEEE Trans. on Neural Systems and Rehabilitation Engineering , vol. 15, no. 4, pp. 473–482, 2007

    2007

  3. [11]

    Random forests in non–invasive sensorimotor rhythm brain-comput er interfaces: a practical and convenient non-linear classifier,

    D. Steyrl, R. Scherer, J. Faller, and G. R. M¨ uller-Putz , “Random forests in non–invasive sensorimotor rhythm brain-comput er interfaces: a practical and convenient non-linear classifier,” Biomedical Engineer- ing/Biomedizinische Technik, vol. 61, no. 1, pp. 77–86, 2016

    2016

  4. [12]

    Autocalibration and recurrent adaptation: Towards a plug and play online erd-BCI,

    J. Faller, C. Vidaurre, T. Solis-Escalante, C. Neuper, and R. Scherer, “Autocalibration and recurrent adaptation: Towards a plug and play online erd-BCI,” IEEE Trans. on Neural Systems and Rehabilitation Engineering, vol. 20, no. 3, pp. 313–319, 2012

    2012

  5. [13]

    Individually adapted imagery impr oves brain– computer interface performance in end-users with disabili ty,

    R. Scherer, J. Faller, E. V . Friedrich, E. Opisso, U. Cos ta, A. K¨ ubler, and G. R. M¨ uller-Putz, “Individually adapted imagery impr oves brain– computer interface performance in end-users with disabili ty,” PloS One, vol. 10, no. 5, p. e0123727, 2015

    2015

  6. [14]

    MOABB: trustworthy algor ithm bench- marking for BCIs,

    V . Jayaram and A. Barachant, “MOABB: trustworthy algor ithm bench- marking for BCIs,” Journal of Neural Engineering , vol. 15, no. 6, p. 066011, 2018

    2018

  7. [15]

    EEG dataset s for motor imagery brain–computer interface,

    H. Cho, M. Ahn, S. Ahn, M. Kwon, and S. C. Jun, “EEG dataset s for motor imagery brain–computer interface,” GigaScience, vol. 6, no. 7, p. gix034, 2017

    2017

  8. [16]

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

    M.-H. Lee, O.-Y . Kwon, Y .-J. Kim, H.-K. Kim, Y .-E. Lee, J. Williamson, S. Fazli, and S.-W. Lee, “EEG dataset and openbmi toolbox for three BCI paradigms: An investigation into BCI illiteracy,” GigaScience, vol. 8, no. 5, p. giz002, 2019

    2019

  9. [17]

    Beam- forming in noninvasive brain–computer interfaces,

    M. Grosse-Wentrup, C. Liefhold, K. Gramann, and M. Buss , “Beam- forming in noninvasive brain–computer interfaces,” IEEE Trans. on Biomedical Engineering , vol. 56, no. 4, pp. 1209–1219, 2009

    2009

  10. [18]

    Upper limb movements can be decoded from the time-domain of low–freque ncy EEG,

    P . Ofner, A. Schwarz, J. Pereira, and G. R. M¨ uller-Putz , “Upper limb movements can be decoded from the time-domain of low–freque ncy EEG,” PloS One , vol. 12, no. 8, p. e0182578, 2017. 10

    2017

  11. [19]

    Physiobank, physiotoolkit, and physionet: comp onents of a new research resource for complex physiologic signals,

    A. L. Goldberger, L. A. Amaral, L. Glass, J. M. Hausdorff , P . C. Ivanov, R. G. Mark, J. E. Mietus, G. B. Moody, C.-K. Peng, and H . E. Stanley, “Physiobank, physiotoolkit, and physionet: comp onents of a new research resource for complex physiologic signals,” Circulation, vo...

    2000

  12. [20]

    Bci2000: a general–purpose brain-computer inter face (BCI) system,

    G. Schalk, D. J. McFarland, T. Hinterberger, N. Birbaum er, and J. R. Wolpaw, “Bci2000: a general–purpose brain-computer inter face (BCI) system,” IEEE Trans. on Biomedical Engineering , vol. 51, no. 6, pp. 1034–1043, 2004

    2004

  13. [21]

    Deep learning with convolutional neural networks for EEG d ecoding and visualization,

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

    2017

  14. [22]

    Open access dataset for EEG+ nirs single- trial classification,

    J. Shin, A. von L ¨ uhmann, B. Blankertz, D.-W. Kim, J. Jeo ng, H.-J. Hwang, and K.-R. M¨ uller, “Open access dataset for EEG+ nirs single- trial classification,” IEEE Trans. on Neural Systems and Rehabilitation Engineering, vol. 25, no. 10, pp. 1735–1745, 2016

    2016

  15. [23]

    Evaluation of EEG oscillatory patterns and cognitive proc ess during simple and compound limb motor imagery,

    W. Yi, S. Qiu, K. Wang, H. Qi, L. Zhang, P . Zhou, F. He, and D . Ming, “Evaluation of EEG oscillatory patterns and cognitive proc ess during simple and compound limb motor imagery,” PloS One , vol. 9, no. 12, p. e114853, 2014

    2014

  16. [24]

    A fully automa ted trial selection method for optimization of motor imagery ba sed brain– computer interface,

    B. Zhou, X. Wu, Z. Lv, L. Zhang, and X. Guo, “A fully automa ted trial selection method for optimization of motor imagery ba sed brain– computer interface,” PloS One , vol. 11, no. 9, p. e0162657, 2016

    2016

  17. [25]

    Continuous sensor imotor rhythm based brain computer interface learning in a large populati on,

    J. R. Stieger, S. A. Engel, and B. He, “Continuous sensor imotor rhythm based brain computer interface learning in a large populati on,” Scientific Data, vol. 8, no. 1, p. 98, 2021

    2021

  18. [26]

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

    H. Liu, P . Wei, H. Wang, X. Lv, W. Duan, M. Li, Y . Zhao, Q. Wa ng, X. Chen, G. Shi, et al. , “An EEG motor imagery dataset for brain computer interface in acute stroke patients,” Scientific Data , vol. 11, no. 1, p. 131, 2024

    2024

  19. [27]

    EEGNet: a compact convolutional neural net work for EEG–based brain–computer interfces,

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

    2018

  20. [28]

    GPT-4 technical report,

    J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L . Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkat, et al. , “GPT-4 technical report,” arXiv preprint arXiv:2303.08774 , 2023

    2023 arXiv

  21. [29]

    The llama 3 herd of models,

    A. Dubey, A. Jauhri, A. Pandey, A. Kadian, A. Al-Dahle, A . Letman, A. Mathur, A. Schelten, A. Y ang, A. Fan, et al. , “The llama 3 herd of models,” arXiv preprint arXiv:2407.21783 , 2024

    2024 arXiv

  22. [30]

    Qwen technical report,

    J. Bai, S. Bai, Y . Chu, Z. Cui, K. Dang, X. Deng, Y . Fan, W. G e, Y . Han, F. Huang, et al. , “Qwen technical report,” arXiv preprint arXiv:2309.16609, 2023

    2023 arXiv

  23. [31]

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

    W.-B. Jiang, L.-M. Zhao, and B.-L. Lu, “Large brain mode l for learning generic representations with tremendous EEG data in BCI,” i n Proc. Int’l Conf. on Learning Representations , (Vienna, Austria), May. 2024

    2024

  24. [32]

    EEGPT: Pret rained transformer for universal and reliable representation of E EG signals,

    G. Wang, W. Liu, Y . He, C. Xu, L. Ma, and H. Li, “EEGPT: Pret rained transformer for universal and reliable representation of E EG signals,” (V ancouver, Canada), pp. 39249–39280, Dec. 2025

    2025

  25. [33]

    Spatial distilla tion based distribution alignment (SDDA) for cross-headset EEG class ification,

    D. Liu, S. Li, Z. Wang, W. Li, and D. Wu, “Spatial distilla tion based distribution alignment (SDDA) for cross-headset EEG class ification,” arXiv preprint arXiv:2503.05349 , 2025

    2025 arXiv

  26. [34]

    Neurophysiological pr edictor of SMR-based BCI performance,

    B. Blankertz, C. Sannelli, S. Halder, E. M. Hammer, A. K¨ ubler, K.-R. M¨ uller, G. Curio, and T. Dickhaus, “Neurophysiological pr edictor of SMR-based BCI performance,” Neuroimage, vol. 51, no. 4, pp. 1303– 1309, 2010

    2010

  27. [35]

    Swi tching EEG headsets made easy: Reducing offline calibration effort using active wighted adaptation regularization,

    D. Wu, V . J. Lawhern, W. D. Hairston, and B. J. Lance, “Swi tching EEG headsets made easy: Reducing offline calibration effort using active wighted adaptation regularization,” IEEE Trans. on Neural Systems and Rehabilitation Engineering , vol. 24, no. 11, pp. 1125–1137, 2016

    2016

  28. [36]

    Enha ncing transfer performance across datasets for brain-computer i nterfaces using a combination of alignment strategies and adaptive batch no rmalization,

    L. Xu, M. Xu, Z. Ma, K. Wang, T.-P . Jung, and D. Ming, “Enha ncing transfer performance across datasets for brain-computer i nterfaces using a combination of alignment strategies and adaptive batch no rmalization,” Journal of Neural Engineering , vol. 18, no. 4, p. 0460e5, 2021

    2021

  29. [37]

    Manifold embedded knowledge transf er for brain-computer interfaces,

    W. Zhang and D. Wu, “Manifold embedded knowledge transf er for brain-computer interfaces,” IEEE Trans. on Neural Systems and Reha- bilitation Engineering , vol. 28, no. 5, pp. 1117–1127, 2020

    2020

  30. [38]

    A cross-dataset adaptive domain selection transfer learn ing framework for motor imagery-based brain-computer interfaces,

    J. Jin, G. Bai, R. Xu, K. Qin, H. Sun, X. Wang, and A. Cichoc ki, “A cross-dataset adaptive domain selection transfer learn ing framework for motor imagery-based brain-computer interfaces,” Journal of Neural Engineering, vol. 21, no. 3, p. 036057, 2024

    2024

  31. [39]

    CBraMoD: A criss-cross brain foundation model for EEG deco ding,

    J. Wang, S. Zhao, Z. Luo, Y . Zhou, H. Jiang, S. Li, T. Li, an d G. Pan, “CBraMoD: A criss-cross brain foundation model for EEG deco ding,” arXiv preprint arXiv:2412.07236 , 2024

    2024

  32. [40]

    EEGfo rmer: A transformer–based brain activity classification method u sing EEG signal,

    Z. Wan, M. Li, S. Liu, J. Huang, H. Tan, and W. Duan, “EEGfo rmer: A transformer–based brain activity classification method u sing EEG signal,” Frontiers in Neuroscience, vol. 17, p. 1148855, 2023

    2023

  33. [41]

    NeuroLM: A uni versal multi-task foundation model for bridging the gap between la nguage and EEG signals,

    W.-B. Jiang, Y . Wang, B.-L. Lu, and D. Li, “NeuroLM: A uni versal multi-task foundation model for bridging the gap between la nguage and EEG signals,” arXiv preprint arXiv:2409.00101 , 2024

    2024 arXiv

Pith tools

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