From Theory to Application: Fine-Tuning Large EEG Model with Real-World Stress Data

Dung Truong; Shitou Zhang; Siwen Wang; Tzyy-Ping Jung; Wan-Lin Chen

REVIEW 4 major objections 6 minor 16 references

From Theory to Application: Fine-Tuning Large EEG Model with Real-World Stress Data

T0 review · 4 major / 6 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 The paper reports that fine-tuning a large EEG foundation model on a small, real-world classroom stress dataset yields 90.47% balanced accuracy with a 5-second window, above the 78.94% achieved by the traditional classifier on the same…

desk verdict Useful real-world stress EEG application, but the recording-level split likely leaks subject identity, so the 90.47% headline is not a reliable generalization estimate. read the letter →

arxiv 2505.23042 v1 pith:FWVRLYLX submitted 2025-05-29 cs.LG cs.AI

Siwen Wang , Shitou Zhang , Wan-Lin Chen , Dung Truong , Tzyy-Ping Jung This is my paper · ORCID

classification cs.LGcs.AI

keywords EEGfoundationmodelLaBraMstressclassificationfine-tuningreal-worldbalancedaccuracychannelreductionresting-state

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper asks whether a large EEG foundation model can be adapted to a messy, real-world stress-detection task when only a small dataset is available. The authors fine-tune LaBraM, a foundation model pretrained on over 2,500 hours of EEG, on resting-state recordings from 18 graduate students collected during a real class session, then train a binary classifier separating normal from elevated stress. The best data split reaches 90.47% balanced accuracy on 5-second windows, while the average across four random splits is about 81%, far above the 53.76% obtained without pretraining. This matters because real-world brain-computer interfaces need short windows and fast inference, not long controlled lab recordings. The result supports a data-centric development path where curating good EEG data and fine-tuning a shared foundation model replaces task-specific architecture design.

What carries the argument

The central object is LaBraM, a large EEG foundation model pretrained on a mixed corpus of over 2,500 hours of EEG. Its pretraining has two stages: a neural tokenizer converts EEG signals into discrete neural tokens by reconstructing their Fourier spectra, and the model then learns transferable features through masked token reconstruction. The downstream machinery is a full-parameter fine-tuning pipeline on the stress dataset, after band-pass filtering, artifact subspace reconstruction, ICA-based rejection of non-cortical components, downsampling to 200 Hz, and segmentation into 5-second patches with overlapping windows on the minority elevated-stress class. The pretrained weights carry the inductive bias: fine-tuning from scratch on the small dataset yields near-chance accuracy, while fine-tuning the pretrained model yields the reported gains.

What would settle it

Run the same 30-channel, 5-second fine-tuning protocol with leave-one-subject-out cross-validation over the 18 participants and compare the average balanced accuracy with the traditional baseline's 78.94%. If the leave-one-out average falls below that baseline or near the 53.76% no-pretraining level, the 90.47% headline is an artifact of the two-subject elevated-stress test set.

Watch

Extended reading notes

Core claim

The paper's central claim is that LaBraM, after full-parameter fine-tuning on the classroom stress dataset, learns stress-relevant features from real-world EEG and outperforms the traditional stress classifier evaluated on the same dataset. With 30 channels and a 5-second window, the best checkpoint reaches 90.47% balanced accuracy; across four random data splits the average is 81.04%, while a model trained from random initialization only reaches 53.76%, near chance. When restricted to the 11 channels used by the traditional baseline, the fine-tuned model reaches 72.23% (seed 42), about six points below the baseline's 78.94% that used a 15-second window, but with all 30 channels the average performance is about two points above that baseline. The paper interprets this as evidence that large EEG models can transfer from controlled pretraining data to real-world recordings and that fine-tuning them is a practical route for near-real-time stress detection.

Load-bearing premise

The central estimate assumes that a test set containing only two elevated-stress subjects is representative enough to produce a reliable balanced-accuracy number, and the paper itself notes this makes the result highly sensitive to out-of-distribution samples.

Editorial extensions

If this is right

A 5-second window is sufficient for stress classification at near-chance-free accuracy, making the approach more suitable for near-real-time BCI applications than the 15-second windows used by the traditional baseline.
Pretraining is the load-bearing component: the same fine-tuning protocol on a randomly initialized model reaches only 53.76%, so the foundation model's pretrained representations, not the architecture alone, drive the gain.
Channel count strongly affects performance: reducing from 30 to 20 channels lowers average balanced accuracy from 81.04% to 71.60%, while 11 carefully selected channels reach 72.23%, showing that channel selection can partly offset the loss.
Performance varies substantially across random data splits, from 66.84% to 90.47%, because the elevated-stress test group contains only two subjects, so the headline number is not a stable population estimate.
The success of fine-tuning on a small real-world dataset supports a shift from designing task-specific models to curating high-quality data and adapting a shared foundation model.

Reading between the lines

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

Editor's inference: the 90.47% figure depends on a single split whose elevated-stress test set contains two subjects; a leave-one-subject-out evaluation over the 18 participants would give a much more trustworthy estimate of cross-person generalization.
Editor's inference: the comparison with the traditional baseline confounds window length and channel count; a matched 5-second, 11-channel comparison would isolate whether the foundation model's advantage comes from pretraining or from using all 30 channels.
Editor's inference: a testable extension would be to fine-tune the same checkpoint on a published multi-subject stress EEG dataset and see whether the 5-second window advantage holds across a larger, more diverse elevated-stress group.
Editor's inference: since the model without pretraining is near chance, an informative ablation would vary the pretraining corpus composition to test whether domain match (resting-state versus task EEG) or raw scale of pretraining data matters more.

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

Colleague,

The thing to know: the headline 90.47% balanced accuracy for fine-tuning LaBraM on real-world classroom stress EEG is best-of-four-seeds, and the split is by recording, not by participant. That second point is the bigger problem. With 92 recordings from 18 participants, the same person almost certainly shows up in both training and test. EEG is highly subject-specific, so the model may be recognizing the person rather than the stress state. A larger test set wouldn't fix that; only a subject-disjoint split would.

What the paper does well: it applies a foundation EEG model to a messy, real-world dataset, which is genuinely underexplored. The pre-training ablation is a useful sanity check—random init lands near chance (53.76%) while pre-trained averages 81.04% across seeds. The channel-count sensitivity experiment is informative, and the authors are transparent about the four seeds and the small elevated-stress test group. That transparency earns credit.

The soft spots go beyond the split. The test set has two elevated-stress subjects, so the balanced-accuracy estimate is extremely noisy; no confidence intervals are given. Seed 1 gives 66.84%, which shows how much the result depends on the split. The abstract highlights the best seed. The 400-second recording cutoff is post hoc and removes ten recordings; the justification is plausible but not evidenced. The comparison to [13] is muddled: at the same 11 channels with seed 42 the fine-tuned LaBraM gets 72.23% versus 78.94%, so the "significantly outperforming" claim rests on the 30-channel result under a possibly leaked split.

The paper is honest about its limitations, but the subject-leakage issue is load-bearing. I'd want to see a participant-disjoint split, preferably with more than two elevated-stress subjects in the test set, before trusting the accuracy claim. The idea is worth pursuing and the dataset is real; this is fixable. I'd send it to review with a request for major revision, but I wouldn't cite the current numbers.

Referee Report

4 major / 6 minor

Summary. The paper fine-tunes LaBraM, a large EEG foundation model, on a real-world classroom stress dataset collected from 18 graduate students (92 recordings, binary normal vs. elevated stress). The authors report a best balanced accuracy of 90.47% using 5-second windows, alongside robustness experiments across four data-splitting seeds, a pre-training ablation, and channel-reduction analyses. They conclude that large EEG models can be effectively adapted to real-world stress data and outperform traditional classifiers, while also noting compute and interpretability limitations.

Significance. If the reported accuracy held under a clean evaluation protocol, this would be a useful proof-of-concept: it would show that a large EEG foundation model can be fine-tuned on a small, noisy real-world dataset and achieve high classification accuracy with short windows, which is relevant for near-real-time BCI applications. The paper honestly reports all four seeds, includes a pre-training ablation showing that random initialization gives only 53.76% accuracy, and provides a channel-count sensitivity analysis. These are genuine strengths. However, the central accuracy claim is currently not supported because the test set is constructed in a way that violates independence and may allow subject-identity shortcuts. The headline result is the best of four highly variable seeds, and no confidence intervals or significance tests are provided.

major comments (4)

[Section II-D, Table I] The 75% overlapping-window augmentation appears to have been applied to the validation and test sets in addition to the training set. The test elevated-stress group contains 391 patches originating from only two recordings; under a non-overlapping 5-second segmentation, two recordings of up to 400 seconds would yield only about 160 patches. The six normal-stress test recordings yield 455 patches, consistent with no augmentation. This means the test set consists of highly overlapping, non-independent windows from just two elevated-stress subjects, so the reported balanced accuracy can be inflated by the model memorizing those particular recordings. Augmentation should be restricted to the training set, or the authors should report accuracy at the recording or subject level.
[Section II-D] The train/validation/test split is performed at the recording level, not the participant level. With 92 recordings from 18 participants, recordings from the same participant can appear in multiple splits. Since EEG is highly subject-specific, a model could exploit subject identity or session-specific noise as a shortcut, and the reported accuracy would then not reflect stress generalization. The authors should either perform a subject-disjoint split or provide an analysis showing the degree of participant overlap across splits and demonstrating that the observed accuracy is not driven by subject identity.
[Section III, Fig. 2, Abstract] The headline 90.47% balanced accuracy is the best result among four random seeds; the other seeds produce 66.84%, 87.01%, and 87.05%. Given that the test elevated-stress group contains only two subjects (as the authors acknowledge) and that no confidence intervals or significance tests are reported, the abstract's claim that the model 'achieves' 90.47% and 'significantly outperforms' traditional classifiers is not statistically supported. The mean and variance across seeds should be the primary result, and the comparison to the baseline in [13] should be accompanied by a proper significance test or confidence interval.
[Section II-D] The 400-second recording-length cutoff is introduced post hoc with a qualitative justification ('typically result from poor signal quality'). The authors do not report how many recordings were removed from each stress group or whether this exclusion changes the class balance. Since the elevated-stress group initially has only 21 recordings, removing even two could meaningfully alter the dataset. A sensitivity analysis around this cutoff, or an objective pre-registered exclusion criterion, is needed to establish that the filtering does not bias the results.

minor comments (6)

[Affiliation] There is a typo in the affiliation of Wan-Lin Chen: 'San Deigo' should be 'San Diego'.
[Section III, Figs. 2-4] The figures report average accuracies across seeds but do not show per-seed markers or error bars; adding them would clarify the large variability visible in the text (e.g., 66.84% to 90.47%).
[Section III] The sentence 'We expect variation in precision' should refer to 'balanced accuracy' or 'performance', since the reported metric is balanced accuracy, not precision.
[Section IV] The claim that the 5-second window makes the system 'suitable for near-real-time stress detection' is not supported by any measurement of inference time or computational latency; a runtime comparison would be needed to substantiate this.
[Section III] The 20-channel configuration is selected randomly, while the 11-channel configuration is based on [13]; the comparison between these two reduced-channel settings is therefore confounded by channel selection, and this should be stated more prominently when interpreting Fig. 4.
[References] Reference [13] is a prior study by an overlapping set of authors; the direct comparison in the Discussion would benefit from explicitly noting this relationship for transparency.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: accuracy is a measured test-set result; self-citations to baseline [13] and thesis [14] are contextual, not load-bearing.

full rationale

The paper's central claim is an empirical result: a LaBraM foundation model fine-tuned on a real-world EEG stress dataset reaches 90.47% balanced accuracy on a held-out test split. This is not a derivation or a first-principles prediction; it is a supervised learning evaluation. The target variable (elevated vs. normal stress per DASS labels) is not defined in terms of the model's output, and no parameter is fitted to the test set. LaBraM is an external pretrained model (cited as [6]), not the authors' own model. The comparison baseline [13] is a prior study with overlapping authorship, but it is used as an external benchmark on the same dataset, and the current model is evaluated independently on its own held-out recordings. Self-citations to [13] and [14] support contextual statements about channel informativeness and are not load-bearing for the main accuracy claim. Concerns about the small elevated-stress test set and the recording-level (rather than subject-level) data split are validity and generalization risks, not circularity. Therefore the paper is not circular.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the transferability of LaBraM's pretrained features, the validity of DASS-based labels, and the preprocessing choices. Several hyperparameters and a data-exclusion cutoff are tuned or selected during the study.

free parameters (6)

Learning rate = 1e-5
The paper states that 5e-5 (the TUAB value) causes overfitting within 3 epochs, so 1e-5 was selected on the validation set.
Batch size = 32
Reduced from the TUAB configuration; chosen based on validation performance and GPU memory.
Augmentation overlap = 75% (step 1.25 s)
Elevated-stress patches are generated with 75% overlapping windows to expand the minority class.
Recording length cutoff = 400 s
Recordings longer than 400 s were discarded as likely poor-quality; this post hoc exclusion affects the class balance (21 to 19 elevated, 71 to 63 normal).
Layer decay = 0.65
Hyperparameter from fine-tuning setup, chosen without systematic search.
Number of epochs = 50
Training ran 50 epochs with early stopping via best validation accuracy.

assumptions (4)

domain assumption LaBraM's pretrained representations transfer to real-world stress EEG.
The entire experiment assumes the 2,500-hour pretraining captures features useful for stress classification, without prior stress-specific validation.
domain assumption DASS survey scores provide a valid ground-truth stress label for each recording.
Stress labels are assigned from the DASS questionnaire completed before each session; the paper does not specify the threshold or verify the label reliability.
domain assumption The artifact-removal pipeline (band-pass, ASR, ICA with 80% threshold) preserves stress-relevant EEG signal.
Preprocessing choices are standard but could remove or distort stress-related neural activity; no validation is provided.
ad hoc to paper Discarding recordings longer than 400 seconds does not bias the stress groups.
The cutoff is introduced solely to clean the dataset; it removes 10 recordings and changes the class counts, with no evidence that the excluded recordings are non-informative.

how reviews work

0 comments

Cite this review

Pith. "Pith review of From Theory to Application: Fine-Tuning Large EEG Model with Real-World Stress Data." pith.science (2026). https://pith.science/paper/FWVRLYLX

@misc{pith2026250523042,
  author       = {Pith},
  title        = {Pith review of: From Theory to Application: Fine-Tuning Large EEG Model with Real-World Stress Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FWVRLYLX}},
  note         = {Machine review of arXiv:2505.23042}
}

read the original abstract

Recent advancements in Large Language Models have inspired the development of foundation models across various domains. In this study, we evaluate the efficacy of Large EEG Models (LEMs) by fine-tuning LaBraM, a state-of-the-art foundation EEG model, on a real-world stress classification dataset collected in a graduate classroom. Unlike previous studies that primarily evaluate LEMs using data from controlled clinical settings, our work assesses their applicability to real-world environments. We train a binary classifier that distinguishes between normal and elevated stress states using resting-state EEG data recorded from 18 graduate students during a class session. The best-performing fine-tuned model achieves a balanced accuracy of 90.47% with a 5-second window, significantly outperforming traditional stress classifiers in both accuracy and inference efficiency. We further evaluate the robustness of the fine-tuned LEM under random data shuffling and reduced channel counts. These results demonstrate the capability of LEMs to effectively process real-world EEG data and highlight their potential to revolutionize brain-computer interface applications by shifting the focus from model-centric to data-centric design.

Figures

Figures reproduced from arXiv: 2505.23042 by the authors.

**Figure 1.** Further Data Processing Pipeline Note that the optimal learning rate and batch size can vary significantly for different datasets. For example, a learning rate of 5e-5 works well for the TUAB dataset, but for our stress dataset, it leads to overfitting within just 3 epochs. We run each training session for 50 epochs and save the checkpoint with the best classification accuracy. On average, 50 epochs take ~2 hours on… view at source ↗

**Figure 3.** Fine-tuning results with pre-train v.s. without pre-train. Average [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

**Figure 4.** Fine-tuning results with different channel counts. Average results [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗

Figures from the paper (1 more)

**Figure 2.** Figure 2: Finetuning results using 4 different data splitting seeds. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

16 extracted references · 8 canonical work pages

[13]

Online Mental Stress Detection Using Frontal-channel EEG Recordings in a Classroom Scenario

Chi-Yuan Chang, Chieh Hsu, Ying Choon Wu, Siwen Wang, Darin Tsui, and Tzyy-Ping Jung. Online mental stress detection using frontal-channel eeg recordings in a classroom scenario. arXiv preprint arXiv:2405.11394, 2024

work page Pith review arXiv 2024
[1]

Foun- dation models for generalist medical artificial intelligence

Michael Moor, Oishi Banerjee, Zahra Shakeri Hossein Abad, Harlan M Krumholz, Jure Leskovec, Eric J Topol, and Pranav Rajpurkar. Foun- dation models for generalist medical artificial intelligence. Nature, 616(7956):259–265, 2023

work page 2023
[2]

Florence: A new foundation model for computer vision

Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021

arXiv 2021
[3]

Coca: Contrastive captioners are image-text foundation models

Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyed- hosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917 , 2022

arXiv 2022
[4]

On the opportunities and risks of foundation models

Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258 , 2021

arXiv 2021
[5]

Brant-2: Foundation model for brain signals

Zhizhang Yuan, Daoze Zhang, Junru Chen, Geifei Gu, and Yang Yang. Brant-2: Foundation model for brain signals. arXiv e-prints , pages arXiv–2402, 2024

work page 2024
[6]

Large brain model for learning generic representations with tremendous eeg data in bci

Wei-Bang Jiang, Li-Ming Zhao, and Bao-Liang Lu. Large brain model for learning generic representations with tremendous eeg data in bci. arXiv preprint arXiv:2405.18765 , 2024

arXiv 2024
[7]

Neuro-gpt: Developing a foundation model for eeg

Wenhui Cui, Woojae Jeong, Philipp Thölke, Takfarinas Medani, Karim Jerbi, Anand A Joshi, and Richard M Leahy. Neuro-gpt: Developing a foundation model for eeg. arXiv preprint arXiv:2311.03764 , 107, 2023

arXiv 2023

Show all 16 references

[8]

Eegformer: Towards transferable and interpretable large-scale eeg foundation model

Yuqi Chen, Kan Ren, Kaitao Song, Yansen Wang, Yifan Wang, Dongsheng Li, and Lili Qiu. Eegformer: Towards transferable and interpretable large-scale eeg foundation model. arXiv preprint arXiv:2401.10278, 2024

2024 arXiv
[9]

Eeg-gpt: exploring capabilities of large language models for eeg classification and interpretation

Jonathan W Kim, Ahmed Alaa, and Danilo Bernardo. Eeg-gpt: exploring capabilities of large language models for eeg classification and interpretation. arXiv preprint arXiv:2401.18006 , 2024

2024 arXiv
[10]

The temple university hospital eeg data corpus

Iyad Obeid and Joseph Picone. The temple university hospital eeg data corpus. Frontiers in neuroscience , 10:196, 2016

2016
[11]

Evaluation of artifact subspace reconstruction for automatic eeg artifact removal

Chi-Yuan Chang, Sheng-Hsiou Hsu, Luca Pion-Tonachini, and Tzyy- Ping Jung. Evaluation of artifact subspace reconstruction for automatic eeg artifact removal. In 2018 40th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC) , pages 124...

2018
[12]

Independent component analysis of electroencephalographic data

Scott Makeig, Anthony Bell, Tzyy-Ping Jung, and Terrence J Se- jnowski. Independent component analysis of electroencephalographic data. Advances in neural information processing systems , 8, 1995

1995
[14]

Advancing the neurophysiological understanding of stress, a study based on recorded Electroencephalography (EEG) data in real-world classroom

Siwen Wang. Advancing the neurophysiological understanding of stress, a study based on recorded Electroencephalography (EEG) data in real-world classroom . PhD thesis, UC San Diego, 2022

2022
[15]

Classification of perceived mental stress using a commercially available eeg headband

Aamir Arsalan, Muhammad Majid, Amna Rauf Butt, and Syed Muhammad Anwar. Classification of perceived mental stress using a commercially available eeg headband. IEEE journal of biomedical and health informatics , 23(6):2257–2264, 2019

2019
[16]

Eeg based stress monitoring

Xiyuan Hou, Yisi Liu, Olga Sourina, Yun Rui Eileen Tan, Lipo Wang, and Wolfgang Mueller-Wittig. Eeg based stress monitoring. In 2015 IEEE international conference on systems, man, and cybernetics , pages 3110–3115. IEEE, 2015

2015

Pith tools

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

[1] [13]

Online Mental Stress Detection Using Frontal-channel EEG Recordings in a Classroom Scenario

Chi-Yuan Chang, Chieh Hsu, Ying Choon Wu, Siwen Wang, Darin Tsui, and Tzyy-Ping Jung. Online mental stress detection using frontal-channel eeg recordings in a classroom scenario. arXiv preprint arXiv:2405.11394, 2024

work page Pith review arXiv 2024

[2] [1]

Foun- dation models for generalist medical artificial intelligence

Michael Moor, Oishi Banerjee, Zahra Shakeri Hossein Abad, Harlan M Krumholz, Jure Leskovec, Eric J Topol, and Pranav Rajpurkar. Foun- dation models for generalist medical artificial intelligence. Nature, 616(7956):259–265, 2023

work page 2023

[3] [2]

Florence: A new foundation model for computer vision

Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021

arXiv 2021

[4] [3]

Coca: Contrastive captioners are image-text foundation models

Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyed- hosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917 , 2022

arXiv 2022

[5] [4]

On the opportunities and risks of foundation models

Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258 , 2021

arXiv 2021

[6] [5]

Brant-2: Foundation model for brain signals

Zhizhang Yuan, Daoze Zhang, Junru Chen, Geifei Gu, and Yang Yang. Brant-2: Foundation model for brain signals. arXiv e-prints , pages arXiv–2402, 2024

work page 2024

[7] [6]

Large brain model for learning generic representations with tremendous eeg data in bci

Wei-Bang Jiang, Li-Ming Zhao, and Bao-Liang Lu. Large brain model for learning generic representations with tremendous eeg data in bci. arXiv preprint arXiv:2405.18765 , 2024

arXiv 2024

[8] [7]

Neuro-gpt: Developing a foundation model for eeg

Wenhui Cui, Woojae Jeong, Philipp Thölke, Takfarinas Medani, Karim Jerbi, Anand A Joshi, and Richard M Leahy. Neuro-gpt: Developing a foundation model for eeg. arXiv preprint arXiv:2311.03764 , 107, 2023

arXiv 2023