AFPM: Alignment-based Frame Patch Modeling for Cross-Dataset EEG Decoding

Dongrui Wu; Siyang Li; Xiaoqing Chen

REVIEW 4 major objections 6 minor 46 references

AFPM: Alignment-based Frame Patch Modeling for Cross-Dataset EEG Decoding

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read AFPM claims that a calibration-free, alignment-based frame-patch Transformer can outperform target-domain-tuned baselines on cross-dataset EEG decoding.

desk verdict A worthwhile empirical study whose 'calibration-free' claim overstates what the method actually does; the framework is solid but the framing needs major revision. read the letter →

arxiv 2507.11911 v1 pith:L5ZZKMEI submitted 2025-07-16 cs.HC cs.IRcs.LG

Xiaoqing Chen , Siyang Li , Dongrui Wu This is my paper · ORCID

classification cs.HCcs.IRcs.LG

keywords EEGdecodingbrain-computerinterfacecross-datasetgeneralizationtransferlearningEuclideanalignmentfoundationmodelsmotorimageryevent-relatedpotentials

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

This paper proposes a two-stage pipeline for cross-dataset EEG decoding that needs no labeled data from the new user: first spatially align raw EEG by selecting task-relevant channels, whitening each domain's covariance with Euclidean alignment, and remapping to a fixed channel layout; then encode synchronized multi-channel temporal windows as patches for a Transformer. Pretrained on eight motor-imagery datasets and six event-related-potential datasets, it is evaluated on five unseen datasets and, without any target-domain fine-tuning or calibration, outperforms 17 baselines that were trained or calibrated on those target datasets, with gains up to 4.40% balanced accuracy on motor imagery and 3.78% Cohen's Kappa on ERP. The claim is that a BCI model can be pretrained across corpora and applied to a new user directly, making calibration-free deployment practical.

What carries the argument

The load-bearing object is the domain-specific whitening matrix $ar{R}^{-1/2}$ computed from unlabeled EEG samples of each new subject, combined with two standardization steps: channel selection by neurophysiological priors and channel mapping into a fixed template. This turns heterogeneous montages and non-stationary distributions into a shared input space, so a single Transformer can treat all datasets alike. The complementary object is Frame-Patch Encoding: instead of tokenizing each channel separately, it forms a patch from all selected channels over a short temporal window, then averages neighboring patches before the Transformer, preserving cross-channel spatiotemporal structure.

What would settle it

Take a held-out dataset whose channel montage shares no electrodes with the task's target channel set, and give AFPM zero unlabeled recordings from the target subject before inference. If the model cannot produce its whitening matrix and either fails to run or drops to chance accuracy, the claim that AFPM is calibration-free in the strict sense is refuted.

Watch

Extended reading notes

Core claim

The central claim is that cross-dataset EEG generalization can be achieved without target-domain calibration by making the input representation uniform before the network sees it. AFPM whitens each domain's spatial covariance via Euclidean alignment, $ ilde{X}_n = ar{R}^{-1/2}X_n'$, where $ar{R}$ is the mean covariance of unlabeled samples from that subject or session, after restricting channels to a task-relevant set (17 sensorimotor channels for motor imagery, 28 central-parietal channels for ERP) and mapping them into a canonical template. The whitened, mapped signals are cut into temporal frames that span all selected channels at once, and these frame patches are averaged and fed to a Transformer. On five held-out datasets, this label-free procedure beats 17 baselines that were trained or calibrated on the target datasets themselves, in some cases by large margins such as 0.7620 versus 0.7180 balanced accuracy on Weibo2014.

Load-bearing premise

The 'calibration-free' claim assumes that recording unlabeled EEG from the new user and using it to compute the Euclidean whitening matrix is not calibration; if a real BCI deployment forbids any data from the new user before inference, the central mechanism cannot run.

Editorial extensions

If this is right

A model pretrained on multiple EEG corpora can be applied to a new dataset with no labeled target data, while baselines need target-dataset training or calibration to compete.
The largest reported gains are on Weibo2014 motor imagery (4.40% balanced accuracy and 6.78% AUC-PR over the strongest baseline) and EPFLP300 ERP (3.78% Cohen's Kappa over the strongest baseline).
Ablations show channel selection is the most important component on motor imagery, channel mapping is second, Euclidean alignment third, and frame-patch encoding fourth, so the alignment module carries much of the cross-dataset transfer.
Performance rises as more training datasets are added, indicating the method benefits from EEG data scaling rather than saturating early.
At 269.28k parameters, AFPM is much smaller than the foundation-model baselines, suggesting calibration-free cross-dataset decoding does not require a huge pretrained model.

Reading between the lines

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

Strictly zero-shot deployment, where no EEG recording from the new user is allowed before inference, would break the Euclidean alignment step because $ar{R}$ in Equation (2) cannot be estimated; a truly calibration-free variant would need a normalization that does not depend on target-domain samples.
The same channel-selection-plus-frame-patch recipe could transfer to other EEG paradigms, such as steady-state visual evoked potentials or sleep staging, by swapping the priors-defined target channel set; this is a direct testable extension.
Because the paper shows that fine-tuning on a fraction of a subject's data further improves accuracy, AFPM is not only a zero-calibration method but also a strong initialization for subject-specific adaptation.
The fixed 17-channel MI and 28-channel ERP target sets imply that datasets with very few overlapping channels, such as BNCI2014-004 with three channels, contribute less usable information; a data-driven channel-selection rule would be needed for truly arbitrary montages.

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.

Desk Editor's Note

read the letter

The paper is a solid empirical study with an overstated central claim. AFPM combines three known pieces—neuro-prior channel selection, Euclidean alignment, and remapping to a fixed channel template—with a frame-patch encoding that packs all channels into temporal windows before a Transformer. The combination is sensible, and the evaluation is unusually thorough: 17 baselines, eight MI and six ERP training datasets, five held-out test datasets, ablations, parameter sensitivity, scaling behavior, and fine-tuning analysis. The headline result—outperforming calibrated baselines without any target labels—is credible and practically useful.

The soft spots are about framing, not the core method. First, 'calibration-free' is the wrong term. The Euclidean alignment step (Eqs. 2–3) estimates a reference covariance from the target subject's unlabeled recordings, so some target-user data collection is required before inference. The paper is really presenting a label-free or unsupervised-calibration framework, and that should be stated plainly. Second, the baselines are all trained or fine-tuned with target labels; the foundation models' true zero-shot performance is not reported. So the claim to be 'the first calibration-free cross-dataset framework' is not established against zero-shot versions of LaBraM, EEGPT, or CBraMod. Third, the hyperparameter analysis in Fig. 4 uses the test datasets to show performance under different P and h; the text says performance stayed consistent, but it does not disclose whether the final values were selected using the test set. That is a selection-on-test risk. The abstract's 3.58% ERP gain also doesn't match the tables (looks like 3.78% on Kappa and 2.98% on AUROC). No significance tests are reported, and the code is promised but not yet released.

None of these sink the paper. The framework is legitimate, the ablations support each component, and the scale of the study is a real plus. The main issues are the calibration-free terminology and the missing zero-shot baseline comparisons.

I'd send it to review with a request for major revision on those points. It's a useful contribution to the EEG transfer-learning literature, and a careful referee can help the authors turn it into a solid publication.

Referee Report

4 major / 6 minor

Summary. The paper proposes AFPM, a two-module framework for cross-dataset EEG decoding. Spatial Alignment (SA) selects task-relevant channels from brain-region priors, applies Euclidean alignment to reduce distribution shift, and remaps selected channels to a unified layout. Frame-Patch Encoding (FPE) constructs synchronized spatiotemporal patches from all selected channels and feeds them to a Transformer encoder. The authors pretrain AFPM on eight MI and six ERP datasets and evaluate it on five held-out datasets (three MI, two ERP) without any labeled target-domain fine-tuning, reporting gains over 17 baselines of up to 4.40% balanced accuracy on MI and 3.78% Cohen's Kappa on ERP. The central claim is that AFPM is the first calibration-free cross-dataset EEG decoding framework.

Significance. If the results hold, AFPM is a practically valuable contribution: it demonstrates that pretraining on multiple EEG corpora with channel remapping and unlabeled-data alignment can transfer to new users and datasets without supervised calibration. The experiments are extensive (8+6 training datasets, 5 test datasets, 17 baselines including 5 foundation models), and the ablations isolate the contributions of channel selection, Euclidean alignment, channel mapping, and frame-patch encoding. The method is also parameter-efficient (269k parameters). However, the central practical claim depends on a narrow definition of calibration: the method still requires unlabeled target-subject recordings to compute the Euclidean alignment matrix, and the evaluation protocol does not fully establish the claimed advantage over zero-shot foundation models. These caveats, if addressed, would make the contribution solid; as written, the claim is overstated.

major comments (4)

[III.B.2, Eqs. (2)-(3)] The 'calibration-free' claim is load-bearing for the paper's practical-value argument, but it is not supported by the method as described. Eq. (2) computes the reference covariance Rbar from D channel-selected EEG samples 'collected at the same time period, with the same EEG cap, from the same subject,' i.e., from the target user, and Eq. (3) whitens with Rbar^{-1/2}. In a strict BCI deployment that forbids any recording from the new user before inference, Rbar is undefined and inference cannot proceed. The method is label-free, not zero-calibration. The abstract, introduction, and conclusion should be revised to state explicitly that AFPM requires unlabeled target-subject data for alignment, or the term 'calibration-free' should be replaced with 'without supervised calibration.'
[IV.F, Fig. 4] The hyperparameter analysis reports performance on the five test datasets (BNCI2014-001, Weibo2014, Zhou2016, EPFLP300, BNCI2015-003) for varying Transformer depth, number of averaging patches P, and shifting step h. If P and h were selected after observing these test-set results, the reported main results in Tables IV and V are optimistically biased by test-set selection. The paper must clarify whether P and h were fixed a priori from validation data or chosen from these test-set curves. If the latter, the evaluation protocol needs to be corrected, e.g., by nested cross-validation or by reporting results for a pre-specified default configuration.
[IV.B.4, Tables IV-V] The comparison against baselines is asymmetric in a way that affects the headline claim. AFPM benefits from unlabeled target-subject data through Euclidean alignment, whereas the baseline models are described as 'trained and calibrated using each dataset's original channel configurations and temporal durations, preserving the native data structure.' It is unclear whether Euclidean alignment was applied to the baseline features; if it was not, the comparison conflates the alignment contribution with the AFPM architecture. Additionally, the five foundation-model baselines (BENDR, BIOT, LaBraM, EEGPT, CBraMod) are fine-tuned on target labels, and their zero-shot performance is not reported. To substantiate the claim of being 'the first calibration-free cross-dataset EEG decoding framework,' the authors should report zero-shot results for these foundation models or otherwise show that no existing method can operate without target labels.
[IV.B.4] The evaluation protocol for AFPM itself is not fully specified. The paper states that AFPM 'was evaluated directly on downstream datasets using the pretrained model without fine-tuning' and that cross-subject three-fold cross-validation was used for all baseline methods. If AFPM also used target-subject unlabeled data for Euclidean alignment within each fold, this must be stated, because it defines the exact deployment conditions. If, instead, the alignment is computed on the full test set including all subjects, the evaluation is not strictly cross-subject. Please specify how the Euclidean alignment reference is estimated for the test subjects and whether this estimation is nested within the cross-validation folds.

minor comments (6)

[Table V] The header 'Coken's Kappa' is a typo for 'Cohen's Kappa'; please correct it.
[IV.C and References] The baseline model is referred to as 'BENDER' in the text and Table IV, but the reference list entry and related work use 'BENDR.' Please unify the naming.
[I, II.C] The related work and introduction refer to 'foundation models' but do not define what qualifies as one; consider adding a one-sentence definition or citing a standard reference.
[III.C.1, Eq. (5)] The notation in Eq. (5) uses d both as the frame stride and, implicitly, in the index (g−1)d; please ensure the stride and window length m are defined clearly before the equation, and clarify the relationship G = ⌈T'/d⌉ + 1 with the zero-padding described.
[Table I] BNCI2014-004 has only 3 channels, which is much fewer than the 17-channel MI target set after channel selection; this is worth commenting on in the text, since the effective input for that dataset is only 3 channels. The paper should state whether this affects training stability or whether it is benign due to the mapping into the template.
[IV.B.4] The sentence 'Unless otherwise specified in the ablation study, Euclidean alignment was applied to all datasets across all experiments' is ambiguous: it is not clear whether this includes the baseline models. Please state explicitly for which methods and which stages Euclidean alignment is used.

Circularity Check

1 steps flagged · score 4.0 of 10

Core cross-dataset derivation is label-free and not circular, but the reported gains are partly tuned on the test datasets and the 'calibration-free' label relies on a narrow definition.

fitted input called prediction [Section IV.F (Model Parameter Analysis), Fig. 4, and Tables III-V.]
"The performance of models on the five datasets under different Transformer depth, number of averaging patches P, and shifting step h is shown in Fig. 4. We chose the Number of averaging patches P MI ={5, 15, 25, 35, 45}, shifting steps hMI ={5, 10, 15, 20, 25} for the MI paradigm and P ERP ={3, 5, 7, 10, 15}, hERP ={1, 2, 3, 4, 5} for the ERP paradigm."

The five datasets in Fig. 4 are exactly the evaluation datasets whose results appear in Tables IV and V (BNCI2014-001, Weibo2014, Zhou2016, EPFLP300, BNCI2015-003). The final hyperparameters in Table III (MI P=25, h=5; ERP P=5, h=2) fall inside the sweeps whose test-set performance is plotted. Therefore the AFPM rows in Tables IV-V are not independent held-out evaluations of a pre-specified model: the P and h values were selected with feedback from the same datasets on which the method is then declared to outperform 17 baselines. The 'calibration-free gains' are partly an artifact of evaluating on the tuning set.

full rationale

The main supervised training/evaluation chain is not circular: AFPM is pretrained on source corpora and applied to target subjects without using target labels; no test label is fitted, and the self-citations to Euclidean alignment [27] and transfer-learning tutorials [11] are published, independently checkable methods rather than unverified self-support. The 'calibration-free' wording is an overstatement rather than a circular derivation, because Eqs. (2)-(3) require unlabeled per-subject recordings to estimate the whitening matrix, so the method is label-free but not zero-calibration in the strict deployment sense. The one genuinely circular element is the hyperparameter analysis of Section IV.F, where P and h are swept on the same five datasets later reported as test results; this is a mild but real form of test-set feedback. It does not invalidate the core architecture or the direction of the results, but it means the headline gains are not fully out-of-sample.

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

The ledger shows that AFPM does not introduce new physical or mathematical entities. The main design choices are the hand-selected channel sets and the hyperparameters P/h, some of which were selected using test-set feedback. The alignment step also assumes the availability of unlabeled target-domain data.

free parameters (4)

Averaging window length P (MI=25, ERP=5) = 25 / 5
Chosen via the model parameter analysis in Fig. 4, which reports performance on the test datasets; not a result of a principled derivation.
Shifting step h (MI=5, ERP=2) = 5 / 2
Selected via the same test-set-based parameter sweep.
Target channel set T_MI (17 channels) = 17-channel set
Hand-selected based on motor cortex priors; not optimized but a design choice that affects all results.
Target channel set T_ERP (28 channels) = 28-channel set
Hand-selected based on midline parietal/central priors.

assumptions (4)

domain assumption Euclidean alignment's mean covariance matrix is a valid whitening reference for each domain.
Eq. 2-3 assume that the arithmetic mean of sample covariances in a domain provides a stable, invertible transformation that reduces inter-domain shift.
domain assumption Task-relevant EEG activity is localized to the selected channels.
Section III-B1 presupposes that MI relies on sensorimotor cortex channels and ERP on midline parietal/central channels, so removing other channels improves signal-to-noise.
domain assumption A short temporal window over all channels contains enough discriminative information for a linear projection to classify the task.
Frame-Patch Encoding (Eq. 5) assumes the flattened multichannel window is a sufficient statistic for decoding.
domain assumption The pretraining datasets and test datasets are drawn from the same task distribution after alignment.
Cross-dataset generalization in Section IV assumes intersubject/interdataset variability is handled by alignment and scaling.

how reviews work

0 comments

Cite this review

Pith. "Pith review of AFPM: Alignment-based Frame Patch Modeling for Cross-Dataset EEG Decoding." pith.science (2026). https://pith.science/paper/L5ZZKMEI

@misc{pith2026250711911,
  author       = {Pith},
  title        = {Pith review of: AFPM: Alignment-based Frame Patch Modeling for Cross-Dataset EEG Decoding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L5ZZKMEI}},
  note         = {Machine review of arXiv:2507.11911}
}

read the original abstract

Electroencephalogram (EEG) decoding models for brain-computer interfaces (BCIs) struggle with cross-dataset learning and generalization due to channel layout inconsistencies, non-stationary signal distributions, and limited neurophysiological prior integration. To address these issues, we propose a plug-and-play Alignment-Based Frame-Patch Modeling (AFPM) framework, which has two main components: 1) Spatial Alignment, which selects task-relevant channels based on brain-region priors, aligns EEG distributions across domains, and remaps the selected channels to a unified layout; and, 2) Frame-Patch Encoding, which models multi-dataset signals into unified spatiotemporal patches for EEG decoding. Compared to 17 state-of-the-art approaches that need dataset-specific tuning, the proposed calibration-free AFPM achieves performance gains of up to 4.40% on motor imagery and 3.58% on event-related potential tasks. To our knowledge, this is the first calibration-free cross-dataset EEG decoding framework, substantially enhancing the practicalness of BCIs in real-world applications.

Figures

Figures reproduced from arXiv: 2507.11911 by the authors.

**Figure 2.** Frame Patch Encoding and Transformer encoder. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 5.** It is observed that subject-specific fine-tuning can improve model performance in most cases. Although AFPM has shown good generalisation ability, because the distribution of EEG data varies greatly from subject to subject, subject-specific fine-tuning still significantly affects the EEG decoding performance of the target subject [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

Figures from the paper (4 more)

**Figure 3.** Figure 3: Results of ablation study on three MI datasets: BNCI2 [PITH_FULL_IMAGE:figures/full_fig_p008_3.png]

**Figure 4.** Figure 4: Influence of model hyper-parameters, i.e., Transfor [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

**Figure 5.** Figure 5: Subject-wise performance of AFPM before and after su [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

**Figure 6.** Figure 6: AFPM Performance on five datasets as the number of trai [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 35 canonical work pages

[1]

Brain leaks an d consumer neurotechnology,

M. Ienca, P . Haselager, and E. J. Emanuel, “Brain leaks an d consumer neurotechnology,” Nature Biotechnology, vol. 36, no. 9, pp. 805–810, 2018

2018
[2]

Brain-computer interfaces i n neurological rehabilitation,

J. J. Daly and J. R. Wolpaw, “Brain-computer interfaces i n neurological rehabilitation,” The Lancet Neurology , vol. 7, no. 11, pp. 1032–1043, 2008

2008
[3]

Active tactile explorat ion using a brain-machine-brain interface,

J. E. O’Doherty, M. A. Lebedev, P . J. Ifft, K. Z. Zhuang, S. Shokur, H. Bleuler, and M. A. L. Nicolelis, “Active tactile explorat ion using a brain-machine-brain interface,” Nature, vol. 479, no. 7372, pp. 228–231, 2011

2011
[4]

Reach and grasp by people with tetraplegia using a neurally controlled robotic arm,

L. R. Hochberg, D. Bacher, B. Jarosiewicz, N. Y . Masse, J. D. Simeral, J. V ogel, S. Haddadin, J. Liu, S. S. Cash, P . V an Der Smagt et al. , “Reach and grasp by people with tetraplegia using a neurally controlled robotic arm,” Nature, vol. 485, no. 7398, pp. 372–375, 2012

2012
[5]

A high-performance neuroprosthesis for speech decoding a nd avatar control,

S. L. Metzger, K. T. Littlejohn, A. B. Silva, D. A. Moses, M . P . Seaton, R. Wang, M. E. Dougherty, J. R. Liu, P . Wu, M. A. Berger, I. Zhuravleva, A. Tu-Chan, K. Ganguly, G. K. Anumanchipalli, and E. F. Chang , “A high-performance neuroprosthesis for speech decoding a nd avatar control,” Nature, vol. 7976, no. 620, pp. 1037–1046, 2023

2023
[6]

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

work page 2012
[7]

Optimizing spatial ﬁlters for robust eeg single-trial ana lysis,

B. Blankertz, R. Tomioka, S. Lemm, M. Kawanabe, and K.-r. Muller, “Optimizing spatial ﬁlters for robust eeg single-trial ana lysis,” IEEE Signal Processing Magazine , vol. 25, no. 1, pp. 41–56, 2008

work page 2008
[8]

xDAW N algorithm to enhance evoked potentials: Application to brain–comput er interface,

B. Rivet*, A. Souloumiac, V . Attina, and G. Gibert, “xDAW N algorithm to enhance evoked potentials: Application to brain–comput er interface,” IEEE Trans. on Biomedical Engineering , vol. 56, no. 8, pp. 2035–2043, 2009

work page 2009

Show all 46 references

[9]

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
[10]

EEG Conformer: Con volutional transformer for EEG decoding and visualization,

Y . Song, Q. Zheng, B. Liu, and X. Gao, “EEG Conformer: Con volutional transformer for EEG decoding and visualization,” IEEE Trans. on Neural Systems and Rehabilitation Engineering , vol. 31, pp. 710–719, 2023

2023
[11]

Transfer learning for moto r imagery based brain-computer interfaces: A tutorial,

D. Wu, X. Jiang, and R. Peng, “Transfer learning for moto r imagery based brain-computer interfaces: A tutorial,” Neural Networks, vol. 153, pp. 235–253, 2022

2022
[12]

Tftl: A task-free transfer learning strategy for eeg-base d cross-subject and cross-dataset motor imagery bci,

Y . Wang, J. Wang, W. Wang, J. Su, C. Bunterngchit, and Z.- G. Hou, “Tftl: A task-free transfer learning strategy for eeg-base d cross-subject and cross-dataset motor imagery bci,” IEEE Trans. on Biomedical Engineering, vol. 72, no. 2, pp. 810–821, 2025

2025
[13]

Language models are few-shot learners,

T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P . D hariwal, A. Neelakantan, P . Shyam, G. Sastry, A. Askell et al. , “Language models are few-shot learners,” in Proc. Advances in Neural Information Processing Systems, Red Hook, NY , 2020, pp. 1877–1

2020
[14]

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

D. Kostas, S. Aroca-Ouellette, and F. Rudzicz, “BENDR: Using trans- formers and a contrastive self-supervised learning task to learn from massive amounts of EEG data,” Frontiers in Human Neuroscience , vol. 15, p. 653659, 2021

2021
[15]

BIOT: Biosignal Tra nsformer for cross-data learning in the wild,

C. Y ang, M. B. Westover, and J. Sun, “BIOT: Biosignal Tra nsformer for cross-data learning in the wild,” in Proc. Advances in Neural Information Processing Systems, vol. 36, Dec. 2023, pp. 78 240–78 260

2023
[16]

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

W. Jiang, L. Zhao, and B. liang Lu, “Large brain model for learning generic representations with tremendous EEG data in BCI,” i n Proc. Int’l Conf. on Learning Representations , May 2024. 10

2024
[17]

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,” in Proc. Advances in Neural Information Processing Systems , vol. 37, V ancouver, Canada, 2024, pp. 39 249–39 280

2024
[18]

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,” in Proc. Int’l Conf. on Learning Representations , 2025

2025
[19]

EEG- based discrimination between imagination of right and left hand move- ment,

G. Pfurtscheller, C. Neuper, D. Flotzinger, and M. Preg enzer, “EEG- based discrimination between imagination of right and left hand move- ment,” Electroencephalography and Clinical Neurophysiology, vol. 103, no. 6, pp. 642–651, 1997

1997
[20]

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
[21]

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

J. Jing, W. Ge, S. Hong, M. B. Fernandes, Z. Lin, C. Y ang, S . An, A. F. Struck, A. Herlopian, I. Karakis et al. , “Development of expert-level classiﬁcation of seizures and rhythmic and periodic patter ns during EEG interpretation,” Neurology, vol. 100, no. 17, pp. e1750–e1762, 2023

2023
[22]

Self-super vised eeg representation learning for automatic sleep staging,

C. Y ang, D. Xiao, M. B. Westover, and J. Sun, “Self-super vised eeg representation learning for automatic sleep staging,” 202 3. [Online]. Available: https://arxiv.org/abs/2110.15278

arXiv
[23]

IFNet: An interactive freq uency con- volutional neural network for enhancing motor imagery deco ding from EEG,

J. Wang, L. Y ao, and Y . Wang, “IFNet: An interactive freq uency con- volutional neural network for enhancing motor imagery deco ding from EEG,” IEEE Trans. on Neural Systems and Rehabilitation Engineeri ng, vol. 31, pp. 1900–1911, 2023

1900
[24]

FBCNet: A multi-view convolu tional neural network for brain-computer interface,

R. Mane, E. Chew, K. Chua, K. K. Ang, N. Robinson, A. P . Vinod, S.-W. Lee, and C. Guan, “FBCNet: A multi-view convolu tional neural network for brain-computer interface,” 2021. [Onli ne]. Available: https://arxiv.org/abs/2104.01233

2021 arXiv
[25]

Transformer-based spatial-temporal feature learning for EEG decoding,

Y . Song, X. Jia, L. Y ang, and L. Xie, “Transformer-based spatial-temporal feature learning for EEG decoding,” arXiv preprint arXiv:2106.11170, 2021

2021 arXiv
[26]

EEG-Deformer: A dense convolutional transformer for brai n-computer interfaces,

Y . Ding, Y . Li, H. Sun, R. Liu, C. Tong, C. Liu, X. Zhou, and C. Guan, “EEG-Deformer: A dense convolutional transformer for brai n-computer interfaces,” IEEE Journal of Biomedical and Health Informatics , vol. 29, no. 3, pp. 1909–1918, 2025

1909
[27]

Transfer learning for brain-computer i nterfaces: A Euclidean space data alignment approach,

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

2019
[28]

A comprehensive survey on transfer learning,

F. Zhuang, Z. Qi, K. Duan, D. Xi, Y . Zhu, H. Zhu, H. Xiong, a nd Q. He, “A comprehensive survey on transfer learning,” Proceedings of the IEEE , vol. 109, no. 1, pp. 43–76, 2021

2021
[29]

Br ant: Foundation model for intracranial neural signal,

D. Zhang, Z. Y uan, Y . Y ANG, J. Chen, J. Wang, and Y . Li, “Br ant: Foundation model for intracranial neural signal,” in Proc. Advances in Neural Information Processing Systems , vol. 36, Pittsburgh, PA, 2023, pp. 26 304–26 321

2023
[30]

Ph ysionet: an nih research resource for complex signals,

M. Costa, G. B. Moody, I. Henry, and A. L. Goldberger, “Ph ysionet: an nih research resource for complex signals,” Journal of Electrocardiology, vol. 36, pp. 139–144, 2003

2003
[31]

A large eeg database with users’ proﬁle information for motor image ry brain- computer interface research,

P . Dreyer, A. Roc, L. Pillette, S. Rimbert, and F. Lotte, “A large eeg database with users’ proﬁle information for motor image ry brain- computer interface research,” Scientiﬁc data, vol. 10, no. 1, p. 580, 2023

2023
[32]

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. 98, 2021

2021
[33]

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
[34]

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, 01 2019

2019
[35]

M3CV: A multi-subject, multi-session, and mu lti-task database for EEG-based biometrics challenge,

G. Huang, Z. Hu, W. Chen, S. Zhang, Z. Liang, L. Li, L. Zhan g, and Z. Zhang, “M3CV: A multi-subject, multi-session, and mu lti-task database for EEG-based biometrics challenge,” Neuroimage, vol. 264, p. 119666, 2022

2022
[36]

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 Transactions on Neural Systems and Rehabilitation Engineering , vol. 15, no. 4, pp. 473–482, 2007

2007
[37]

Review of the BCI competition IV,

M. Tangermann, K.-R. Müller, A. Aertsen, N. Birbaumer, C. Braun, C. Brunner, R. Leeb, C. Mehring, K. J. Miller, G. Mueller-Put z et al. , “Review of the BCI competition IV,” Frontiers in Neuroscience, vol. 6, p. 55, 2012

2012
[38]

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
[39]

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, pp. 1–20, 2016

2016
[40]

Brain invaders adaptive versus non-adaptive p 300 brain- computer interface dataset,

E. V aineau, A. Barachant, A. Andreev, P . C. Rodrigues, G . Cattan, and M. Congedo, “Brain invaders adaptive versus non-adaptive p 300 brain- computer interface dataset,” arXiv preprint arXiv:1904.09111 , 2019

1904 arXiv
[41]

Brain invaders ca libration- less p300-based bci using dry eeg electrodes dataset (bi201 4a),

L. Korczowski, E. Ostaschenko, A. Andreev, G. Cattan, P . L. C. Rodrigues, V . Gautheret, and M. Congedo, “Brain invaders ca libration- less p300-based bci using dry eeg electrodes dataset (bi201 4a),” Ph.D. dissertation, GIPSA-lab, 2019

2019
[42]

Brain invaders cali bration-less p300-based bci with modulation of ﬂash duration dataset (bi 2015a),

L. Korczowski, M. Cederhout, A. Andreev, G. Cattan, P . L . C. Ro- drigues, V . Gautheret, and M. Congedo, “Brain invaders cali bration-less p300-based bci with modulation of ﬂash duration dataset (bi 2015a),” Ph.D. dissertation, GIPSA-lab, 2019

2019
[43]

How many people are able to co ntrol a P300-based brain–computer interface (bci)?

C. Guger, S. Daban, E. W. Sellers, C. Holzner, G. Krausz, R. Carabalona, F. Gramatica, and G. Edlinger, “How many people are able to co ntrol a P300-based brain–computer interface (bci)?” Neuroscience Letters, vol. 462, pp. 94–98, 2009

2009
[44]

An efﬁcient P300-based brain–computer interface for disabled subject s,

U. Hoffmann, J.-M. V esin, T. Ebrahimi, and K. Diserens, “An efﬁcient P300-based brain–computer interface for disabled subject s,” Journal of Neuroscience Methods, vol. 167, no. 1, pp. 115–125, 2008

2008
[45]

Motor imagery eeg cl assiﬁcation algorithm based on cnn-lstm feature fusion network,

H. Li, M. Ding, R. Zhang, and C. Xiu, “Motor imagery eeg cl assiﬁcation algorithm based on cnn-lstm feature fusion network,” Biomedical Signal Processing and Control , vol. 72, p. 103342, 2022

2022
[46]

Transformer convolut ional neural networks for automated artifact detection in scalp eeg,

W. Y . Peh, Y . Y ao, and J. Dauwels, “Transformer convolut ional neural networks for automated artifact detection in scalp eeg,” in Proc. IEEE Engineering in Medicine and Biology Society , Glasgow, United King- dom, Jul 2022, pp. 3599–3602

2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Brain leaks an d consumer neurotechnology,

M. Ienca, P . Haselager, and E. J. Emanuel, “Brain leaks an d consumer neurotechnology,” Nature Biotechnology, vol. 36, no. 9, pp. 805–810, 2018

2018

[2] [2]

Brain-computer interfaces i n neurological rehabilitation,

J. J. Daly and J. R. Wolpaw, “Brain-computer interfaces i n neurological rehabilitation,” The Lancet Neurology , vol. 7, no. 11, pp. 1032–1043, 2008

2008

[3] [3]

Active tactile explorat ion using a brain-machine-brain interface,

J. E. O’Doherty, M. A. Lebedev, P . J. Ifft, K. Z. Zhuang, S. Shokur, H. Bleuler, and M. A. L. Nicolelis, “Active tactile explorat ion using a brain-machine-brain interface,” Nature, vol. 479, no. 7372, pp. 228–231, 2011

2011

[4] [4]

Reach and grasp by people with tetraplegia using a neurally controlled robotic arm,

L. R. Hochberg, D. Bacher, B. Jarosiewicz, N. Y . Masse, J. D. Simeral, J. V ogel, S. Haddadin, J. Liu, S. S. Cash, P . V an Der Smagt et al. , “Reach and grasp by people with tetraplegia using a neurally controlled robotic arm,” Nature, vol. 485, no. 7398, pp. 372–375, 2012

2012

[5] [5]

A high-performance neuroprosthesis for speech decoding a nd avatar control,

S. L. Metzger, K. T. Littlejohn, A. B. Silva, D. A. Moses, M . P . Seaton, R. Wang, M. E. Dougherty, J. R. Liu, P . Wu, M. A. Berger, I. Zhuravleva, A. Tu-Chan, K. Ganguly, G. K. Anumanchipalli, and E. F. Chang , “A high-performance neuroprosthesis for speech decoding a nd avatar control,” Nature, vol. 7976, no. 620, pp. 1037–1046, 2023

2023

[6] [6]

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

work page 2012

[7] [7]

Optimizing spatial ﬁlters for robust eeg single-trial ana lysis,

B. Blankertz, R. Tomioka, S. Lemm, M. Kawanabe, and K.-r. Muller, “Optimizing spatial ﬁlters for robust eeg single-trial ana lysis,” IEEE Signal Processing Magazine , vol. 25, no. 1, pp. 41–56, 2008

work page 2008

[8] [8]

xDAW N algorithm to enhance evoked potentials: Application to brain–comput er interface,

B. Rivet*, A. Souloumiac, V . Attina, and G. Gibert, “xDAW N algorithm to enhance evoked potentials: Application to brain–comput er interface,” IEEE Trans. on Biomedical Engineering , vol. 56, no. 8, pp. 2035–2043, 2009

work page 2009