Pith. sign in

REVIEW 6 minor 3 cited by

EEG foundation models encode dataset identity, not portable disease information, on clinical benchmarks

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 01:27 UTC pith:IGP2XJR4

load-bearing objection Careful, honest negative-control benchmark showing frozen EEG foundation models underperform classical features on an external Korean cohort and encode dataset identity at ceiling; credible as stated, but the headline CAUEEG gap rests on recording-level folds and awaits code release.

arxiv 2607.24519 v2 pith:IGP2XJR4 submitted 2026-07-27 cs.LG cs.AIcs.NE

What EEG Foundation Models Encode: Dataset Identity and a Negative-Control Suite for Clinical Benchmarks

Marzieh Zare This is my paper
classification cs.LG cs.AIcs.NE
keywords EEG foundation modelsclinical EEGlinear probingdataset identitynegative controlscross-population transferCAUEEGictal detection
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper argues that current EEG foundation models, when evaluated with symmetric preprocessing and strict negative controls, do not transfer disease information across clinical populations. In a matched external-cohort test on Korean dementia-staging EEG, hand-crafted spectral features reached a higher macro-AUROC than every frozen foundation model tested. Frozen embeddings allow a linear probe to decode which dataset a recording came from at near-ceiling accuracy, while diagnosis decoding stays weak, suggesting these representations store dataset membership rather than portable clinical signal. One task-specific result survives the controls: cross-subject seizure detection, where a pretrained encoder outperforms its random-initialized counterpart and a label-permutation null collapses to chance. The paper distills these checks into a reporting protocol for clinical EEG foundation-model studies.

Core claim

Apparent transfer gains of EEG foundation models do not survive symmetric, controlled evaluation. On the Korean CAUEEG dementia-staging cohort, classical features reach 0.734 macro-AUROC versus 0.699 (BIOT), 0.669 (CBraMod), and 0.568 (REVE) under identical 19-channel five-fold splits; a patient-disjoint sensitivity preserves the ordering. Frozen REVE embeddings let a linear probe decode dataset identity at or near ceiling across Western–Korean and Western–Western pairs, even after PCA-50 and after removing line-frequency and amplitude-scale information, establishing dataset membership rather than a causal site or population effect. A matched random-initialized encoder exceeds pretrained REV

What carries the argument

The paper's central instrument is a negative-control suite: symmetric clean-montage construction so foundation-model and classical features see identical channels and folds; stronger classical comparators including enhanced spectral features and an MLP head; label permutation and scrambled-label fine-tuning to expose label-independent gains; matched random initialization to separate architecture from pretraining; and a dataset-identity probe, a linear classifier that decodes which dataset produced a recording from frozen embeddings. The dataset-identity probe does the load-bearing work: near-ceiling decodability across Western–Korean and Western–Western pairs, robust to PCA-50 and to removal

Load-bearing premise

The headline external-cohort result depends on recording-level five-fold validation because CAUEEG exposes no patient identifier, so the central comparison holds only if the public patient-overlap annotation correctly identifies all repeated recordings; if overlap differs between the classical and foundation-model pipelines, the reported gap could be partly a leakage artifact.

What would settle it

Run a fully patient-disjoint evaluation on CAUEEG with a verified patient map, or on a new external cohort with explicit patient identifiers and identical 19-channel preprocessing; if any frozen foundation model matches or exceeds the classical 0.734 macro-AUROC, and the classical-over-REVE ordering reverses, the paper's central claim is falsified.

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

If this is right

  • Published clinical gains of EEG foundation models should be re-audited with montage-matched folds, patient-disjoint splits, and stronger classical baselines before being attributed to portable disease knowledge.
  • Cross-population deployment of frozen EEG foundation models would be expected to underperform hand-crafted features unless specific adaptation and scrambled-label controls are passed.
  • The task-specific seizure-detection result (pretrained vs random-init separation, label-permutation collapse) indicates that frozen representations may still carry event-related signal, but its scope should be verified with an amplitude-aware classical comparator and paired subject-level statistics.
  • Dataset identity being decodable at ceiling implies that benchmark scores across datasets are not directly comparable; recording provenance must be treated as a confound in any cross-dataset leaderboard.
  • A minimal reporting protocol — clean symmetric channels, patient-disjoint evaluation or leakage sensitivity, stronger comparators, random init, label permutation, scrambled-label adaptation — becomes the standard bar for clinical EEG foundation-model claims.

Where Pith is reading between the lines

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

  • A direct extension would be to test whether the dataset-identity signal is driven by country, device, or recording style by training a domain-adversarial classifier that removes dataset-specific directions and then re-measuring diagnosis decodability; the paper leaves this causal decomposition untested.
  • If dataset membership is near-ceiling decodable, multi-site clinical trials should publish per-site stratified results; a foundation model that wins on average could still be reading site-specific artifacts that happen to correlate with outcomes.
  • A testable prediction is that on a truly patient-disjoint, multi-site cohort with identical hardware and protocol, the classical-over-foundation-model gap on dementia staging would shrink or reverse; collecting such a cohort would settle whether the CAUEEG deficit is about population or about dataset shift.
  • The scrambled-label LoRA result (real equals null) suggests that any apparent cross-population adaptation gain should be reported with a label-shuffled control; future studies that omit this control should be read with caution.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

0 major / 6 minor

Summary. This paper is an audit-style benchmark of seven EEG foundation models (LaBraM, EEGMamba, CBraMod, REVE, LEAD, BENDR, BIOT) on five clinical tasks across four benchmark datasets, with a Korean external-cohort stress test (CAUEEG). Primary comparisons use frozen embeddings with linear probes under subject-disjoint LOSO or grouped five-fold CV where identifiers exist, and recording-level five-fold CV with a patient-disjoint sensitivity for CAUEEG. The headline results are: (i) on a symmetric clean 19-channel CAUEEG Normal/MCI/Dementia evaluation, classical features reach 0.734 macro-AUROC versus 0.699 for BIOT, 0.669 for CBraMod, and 0.568 for REVE, with the classical-over-REVE ordering preserved on an n=172 patient-disjoint subset; (ii) dataset identity is linearly decoded from frozen REVE embeddings at or near ceiling across Western–Korean and Western–Western dataset pairs, even after PCA-50 and removal of line-frequency/amplitude information; (iii) on CHB-MIT cross-subject ictal detection, pretrained REVE reaches 0.793 versus 0.701 for random initialisation and 0.505 for raw random features, while label permutation collapses the probe to chance. The paper also documents that scrambled-label LoRA reproduces apparent adaptation gains, that enhanced classical features match or exceed REVE on Alzheimer's detection, and that many apparent findings change after montage matching and stronger comparators. The authors are unusually transparent: they correct an earlier asymmetric CAUEE

Significance. If the results are taken at face value, this is a significant contribution to the clinical EEG foundation-model literature. The paper provides direct evidence that frozen EEG-FM representations can encode dataset identity rather than portable disease information, and it demonstrates that several reported gains disappear under symmetric preprocessing, stronger classical baselines, and negative controls. The negative-control suite — label permutation, scrambled-label adaptation, matched random initialisation, and symmetric montage construction — is a reusable methodological contribution that goes beyond the usual leaderboard-style benchmark. The manuscript is also exemplary in its self-correction practices: it explicitly retires unreliable estimates, discloses post-hoc rows, and attaches precise caveats to each claim (e.g., that the CHB-MIT result does not establish superiority over amplitude-aware handcrafted baselines). The main limitations are clearly stated, in particular the absence of patient identifiers for CAUEEG and the descriptive nature of several comparative statistics. The paper is likely to be influential as a cautionary benchmark and as a template for future EEG-FM eva

minor comments (6)
  1. [Abstract and §4.6, Table 9] The headline CAUEEG value 0.734 is the best over the three probe families (SVM-RBF), while the stated primary protocol is frozen linear probing. The all-LogReg comparison (Classical 0.720 vs REVE 0.528) preserves the direction, but the abstract should either state 'best-of-probe-family' or headline the linear-probe value to avoid implying the 0.734 number comes from the primary linear probe.
  2. [§3.2 and §4.6] Please provide more detail on the provenance and availability of 'dementia-no-overlap.json' and the 'authors' fixed split': where they are obtained, whether they cover all three diagnostic groups, and how the 1,187-recording CAUEEG analysis cohort is derived from the 1,379 released recordings. This is important for reproducibility, since the patient-disjoint sensitivity depends entirely on this file.
  3. [§3.4, footnote] There is an internal inconsistency between the stated feature count '9×N_ch + 2 features per epoch (spectral/temporal features plus hemispheric alpha asymmetry and frontal theta power)' and the footnote saying Tables 3–5 were computed without alpha asymmetry and frontal theta. Please state explicitly which tables use which feature set and report the exact feature dimensionality for each classical row (e.g., 9×19 vs 9×19+2).
  4. [Abstract and Conclusion] The CAUEEG gap is presented in the abstract as 'classical features achieve 0.734 ... versus ... 0.568 for REVE' without the immediately relevant caveat that the primary CAUEEG folds are recording-level. The body (§5.4 #12) is appropriately candid, but the abstract should carry a short qualifier such as 'recording-level primary folds' so readers do not over-interpret the point estimate.
  5. [§2.2 and References] References [37], [41], and [42] appear in the reference list but are not cited in the visible text. Either cite them where relevant (e.g., analysis variability, scaling laws, the companion manuscript) or remove them from the list.
  6. [§4.7] The conclusion that scrambled labels reproduce the LoRA movement relies on near-equality of means (0.7039 vs 0.7050 and 0.8392 vs 0.8397). Reporting seed-level or fold-level paired values would strengthen this claim; as written, the reader cannot assess whether the tiny 0.1 pp differences are within seed variability.

Circularity Check

0 steps flagged

No significant circularity: the paper is an empirical audit/benchmark whose claims are tested against independent controls, not derived from its own inputs.

full rationale

This is a benchmarking and negative-control paper, not a derivation. The main claims are empirical comparisons: frozen embeddings from pretrained models are scored by linear probes against classical features and randomized controls. No reported quantity is defined in terms of the quantity it is claimed to predict, and no fitted parameter is renamed as a prediction. The CAUEEG external-cohort comparison is explicitly recording-level and is accompanied by a patient-disjoint sensitivity; reliance on the authors' dementia-no-overlap.json is an acknowledged data-validity limitation (leakage risk), not a circular step, because the paper does not claim that file proves the gap. The dataset-identity probe trains a linear classifier on dataset labels, but the reported result is held-out/out-of-pair generalization at or near ceiling; this is an empirical measurement of representation content, not an equivalence-by-construction. The scrambled-label fine-tuning and label-permutation controls are designed to test whether label information explains observed movements, and the null results are presented as null results. The only self-citation is reference [42], a companion manuscript under review, which is not load-bearing in the provided text: no central premise is justified by it, and no uniqueness theorem or ansatz is imported from it. All benchmark model scores are produced independently from public weights and external datasets, and the paper explicitly flags its own limitations, including in-domain pretraining exposure, missing patient identifiers, and unpaired uncertainty estimates. Consequently there is no circular step to report under the criteria of this review. Correctness risks, such as possible unannotated patient overlap in CAUEEG or probe-family selection favoring the classical comparator, are separate from circularity and do not raise the score.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

This is an empirical benchmark rather than a derivation, so the ledger records the analysis choices the conclusions ride on rather than fitted constants. The free parameters are hand-set probe and preprocessing decisions (C values, PCA-200, 1:3 ratio, z-scoring, TUAB subsampling), each disclosed in the text. The axioms are domain assumptions about dataset metadata — patient-identifier absence, pretraining-inventory completeness, the reliability of the dementia-no-overlap.json trace — plus the fairness of one shared preprocessing pipeline across seven differently-pretrained models. No invented entities: 'dataset identity' is explicitly treated as a measured, non-causal property, not a postulated object.

free parameters (5)
  • Probe regularization C = C=1.0 (CHB-MIT, TUAB); C=0.1 (ds004504, CAUEEG)
    Hand-set per pipeline, not selected by nested CV; the paper flags this as a limitation for cross-task absolute comparisons (§3.5). CHB-MIT result is insensitive to C over {0.1, 1.0, 10}.
  • In-fold PCA-200 on FM embeddings = 200 components (CHB-MIT)
    Applied to foundation-model embeddings but not to classical features (Table 3 note). §4.10 shows the PCA sweep is flat (0.789–0.794), so low impact on the REVE estimate.
  • Ictal:interictal ratio with same-session negatives = 1:3
    Task-construction choice (§3.2); AUC-PR 0.525 and Brier 0.213 are quoted at this constructed prevalence, not natural continuous-recording prevalence (§4.11).
  • TUAB epoch subsample = 4,807 of 8,130 epochs
    Stratified per subject, seed 42; the full-set classical AUROC is 2.4 pp higher (77.1 vs 74.7), and full-set FM estimates were not run (§5.4 #8).
  • Per-epoch z-score normalization and clip = z-score per channel, clip ±8 SD
    Chosen as a pretraining-compatibility constraint (§3.3); removes absolute amplitude and thereby disables an amplitude-aware handcrafted comparator on CHB-MIT (§3.2 note), directly shaping the scope of the surviving positive claim.
axioms (6)
  • domain assumption One shared preprocessing pipeline is a fair interface to all seven models
    §3.3: 'a single uniform pipeline cannot simultaneously match seven distinct pretraining pipelines'; BENDR (256 Hz pretraining) runs at 200 Hz with truncation/padding, so its near-chance result is confounded (§5.4 #4).
  • domain assumption CHB-MIT is absent from REVE's pretraining inventory
    §3.2: 'treated as out-of-domain subject to the completeness of that inventory'; if REVE's published list omitted CHB-MIT, the CHB-MIT pretrained-weight claim weakens.
  • domain assumption dementia-no-overlap.json is a faithful patient-overlap trace
    §4.6: 'the authors' dementia-no-overlap.json file is the only public trace of patient overlap'; repeat recordings wholly within training remain unquantified (§5.4 #12).
  • domain assumption Approximate p-values from overlapping LOSO training sets are meaningful
    §4.11: 'The p-value is approximate because LOSO test subjects are disjoint but training sets overlap heavily across folds.'
  • domain assumption The pure-PyTorch Mamba2 port is faithful
    §3.1: CUDA-only selective scan reimplemented in pure PyTorch; a port bug would only affect the post-hoc EEGMamba rows, which are non-decisive.
  • standard math Standard statistics and DSP machinery behave as assumed
    L2 logistic regression, Welch band power, Hjorth parameters, PCA, Wilcoxon signed-rank; none load-bearing beyond the usual statistical caveats.

pith-pipeline@v1.3.0-alltime-deepseek · 22493 in / 20993 out tokens · 209159 ms · 2026-08-04T01:27:43.979614+00:00 · methodology

0 comments
read the original abstract

Pretrained EEG foundation models are proposed for clinical decoding, but whether reported gains transfer across populations or survive negative controls is unclear. We benchmark LaBraM, EEGMamba, CBraMod, REVE, LEAD, BENDR, and BIOT on five clinical tasks across four datasets. Primary analyses use frozen linear probes with subject-disjoint LOSO or grouped five-fold validation. Because CAUEEG releases no patient identifiers, it is evaluated at recording level with a patient-disjoint sensitivity. We challenge apparent gains using stronger classical comparators, label permutation, scrambled-label fine-tuning, and random-initialisation controls. In a matched 19-channel CAUEEG evaluation (Normal/MCI/Dementia; N = 1,187 recordings), classical features achieve 0.734 macro-AUROC versus 0.699 for BIOT, 0.669 for CBraMod, and 0.568 for REVE. A patient-disjoint sensitivity retains the classical-over-REVE ordering (0.717 versus 0.565). Dataset identity is decoded from frozen REVE embeddings at or near ceiling across Western-Korean and Western-Western pairs, including after PCA-50 and removal of line-frequency and amplitude-scale information. This establishes dataset membership, not a causal site or population effect. A matched random-initialised encoder exceeds pretrained REVE on CAUEEG (0.659 versus 0.570). On CHB-MIT cross-subject ictal detection (n = 23), REVE reaches 0.793, versus 0.739 for the best enhanced nonlinear comparator, 0.701 for random initialisation, and 0.505 for raw-signal random features. Because preprocessing removes absolute amplitude, this does not establish superiority over every plausible handcrafted baseline. Conclusions change materially after montage matching, patient-overlap checks, stronger comparators, and representation controls. We distill these checks into a reporting protocol for clinical EEG foundation-model studies.

Figures

Figures reproduced from arXiv: 2607.24519 by Marzieh Zare.

Figure 1
Figure 1. Figure 1: Evaluation pipeline. Six pretrained models are evaluated where their input constraints [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 1
Figure 1. Figure 1: Evaluation pipeline. Seven pretrained models are evaluated under a unified frozen [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: CHB-MIT cross-subject ictal detection on the full cohort. BIOT has CHB-MIT [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 2
Figure 2. Figure 2: CHB-MIT cross-subject ictal detection on the corrected 23-subject cohort. BIOT [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: REVE band ablation on ds004504 (AD vs. HC, 5-fold stratified group CV). Sub-alpha [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figure 3
Figure 3. Figure 3: REVE band ablation on ds004504 (AD vs. HC, 5-fold stratified group CV). Sub-alpha [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: ROI pooling ablation on AD vs. FTD (REVE frozen, 5-fold CV). Frontal, posterior, [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 4
Figure 4. Figure 4: ROI pooling ablation on AD vs. FTD (REVE frozen, 5-fold CV). Extracting frontal, [PITH_FULL_IMAGE:figures/full_fig_p015_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Frozen REVE 19-channel embeddings pooled across the Western (ds004504, [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗

Review history (2 revisions) →

discussion (0)

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

Forward citations

Cited by 3 Pith papers

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

  1. Cross-Cohort Spectral-Temporal Dissociation in Frozen EEG Foundation-Model Representations

    q-bio.NC 2026-07 conditional novelty 6.0

    EEG foundation models fail to encode the alpha-envelope DFA exponent, a disease-relevant temporal-scaling feature, while spectral-input models still encode the static 1/f slope.

  2. Cross-Cohort Spectral-Temporal Dissociation in Frozen EEG Foundation-Model Representations

    q-bio.NC 2026-07 conditional novelty 6.0

    Frozen EEG foundation models can decode the aperiodic spectral slope, but not the alpha-envelope DFA exponent reproducibly across two cohorts.

  3. Technological Advances in Detecting and Managing Cognitive Impairment in Older Adults: Trends, Challenges, and Future Directions

    cs.LG 2026-07 conditional novelty 3.0

    Narrative review of cognitive-impairment detection technologies concludes that reported accuracies are often inflated by weak validation and that progress depends on multimodal, longitudinally validated, externally te...

Reference graph

Works this paper leans on

43 extracted references · 7 linked inside Pith · cited by 2 Pith papers

  1. [1]

    A., Grabowski, H

    DiMasi, J. A., Grabowski, H. G., & Hansen, R. W. (2016). Innovation in the pharmaceutical industry: New estimates of R&D costs.Journal of Health Economics, 47, 20–33

    2016

  2. [2]

    W., Craighead, J

    Hay, M., Thomas, D. W., Craighead, J. L., et al. (2014). Clinical development success rates for investigational drugs.Nature Biotechnology, 32(1), 40–51

    2014

  3. [3]

    & Barachant, A

    Jayaram, V. & Barachant, A. (2018). MOABB: Trustworthy algorithm benchmarking for BCIs.Journal of Neural Engineering, 15(6), 066011

    2018

  4. [4]

    Bommasani, R., et al. (2021). On the opportunities and risks of foundation models. arXiv:2108.07258

    Pith/arXiv arXiv 2021

  5. [5]

    He, K., Chen, X., Xie, S., et al. (2022). Masked autoencoders are scalable vision learners. Proc. CVPR, pp. 16000–16009

    2022

  6. [6]

    B., Zhao, L

    Jiang, W. B., Zhao, L. M., & Lu, B. L. (2024). Large Brain Model for Learning Generic Representations with Tremendous EEG Data in BCI.Proc. ICLR(Spotlight)

    2024

  7. [7]

    Wang, J., et al. (2025). EEGMamba: An EEG Foundation Model with Mamba.Neural Networks, 192, 107816

    2025

  8. [8]

    Wang, J., et al. (2025). CBraMod: A Criss-Cross Brain Foundation Model for EEG Decoding. Proc. ICLR. 24

    2025

  9. [9]

    El Ouahidi, Y., Lys, J., Th¨ olke, P., Farrugia, N., Pasdeloup, B., Gripon, V., Jerbi, K., & Lioi, G. (2025). REVE: A Foundation Model for EEG, Adapting to Any Setup with Large-Scale Pretraining on 25,000 Subjects.Advances in Neural Information Processing Systems (NeurIPS). arXiv:2510.21585

    arXiv 2025

  10. [10]

    Wu, J., et al. (2025). AdaBrain-Bench: Benchmarking Brain Foundation Models for Brain- Computer Interface Applications.arXiv:2507.09882

    Pith/arXiv arXiv 2025

  11. [11]

    Wang, Y., Huang, N., Mammone, N., Cecchi, M., & Zhang, X. (2025). LEAD: An EEG Foundation Model for Alzheimer’s Disease Detection.arXiv preprint arXiv:2502.01678 [cs.LG, eess.SP]

    arXiv 2025

  12. [12]

    Kostas, D., Aroca-Ouellette, S., & Rudzicz, F. (2021). BENDR: Using transformers and a contrastive self-supervised learning task to learn from massive amounts of EEG data. Frontiers in Human Neuroscience, 15, 653659

    2021

  13. [13]

    B., & Sun, J

    Yang, C., Westover, M. B., & Sun, J. (2023). BIOT: Biosignal Transformer for Cross-data Learning in the Wild.Advances in Neural Information Processing Systems (NeurIPS), 36, 78240–78260

    2023

  14. [14]

    Chen, T., Kornblith, S., Norouzi, M., & Hinton, G. (2020). A simple framework for contrastive learning of visual representations.Proc. ICML, pp. 1597–1607

    2020

  15. [15]

    & Dao, T

    Gu, A. & Dao, T. (2024). Mamba: Linear-time sequence modeling with selective state spaces.Proc. COLM

    2024

  16. [16]

    Dao, T. & Gu, A. (2024). Transformers are SSMs: Generalized models and efficient algorithms through structured state space duality.Proc. ICML

    2024

  17. [17]

    Kuruppu, G., Wagh, N., Kremen, V., Pati, S., Worrell, G., & Varatharajah, Y. (2025). EEG Foundation Models: A Critical Review of Current Progress and Future Directions. arXiv:2507.11783

    arXiv 2025

  18. [18]

    Liu, D., Chen, Y., Chen, Z., Cui, Z., Wen, Y., An, J., Luo, J., & Wu, D. (2026). EEG Foundation Models: Progresses, Benchmarking, and Open Problems.arXiv:2601.17883

    arXiv 2026

  19. [19]

    A., Laskaris, N., & Zafeiriou, S

    Lee, N., Bakas, S., Barmpas, K., Panagakis, Y., Adamos, D. A., Laskaris, N., & Zafeiriou, S. (2025). Assessing the Capabilities of Large Brainwave Foundation Models.2025 IEEE 35th International Workshop on Machine Learning for Signal Processing (MLSP)

    2025

  20. [20]

    Wang, X., Yang, Y., & Coyle, D. (2026). EEG-FM-Audit: A Systematic Evaluation and Analysis Pipeline for EEG Foundation Models.arXiv:2605.26910

    Pith/arXiv arXiv 2026

  21. [21]

    Kastrati, A., B¨ urki, J., Lauer, J., Xuan, C., Iaquinto, R., & Wattenhofer, R. (2025). EEG- Bench: A Benchmark for EEG Foundation Models in Clinical Applications.Foundation Models for the Brain and Body Workshop (BrainBodyFM), NeurIPS 2025. arXiv:2512.08959

    arXiv 2025

  22. [22]

    G., Karaiskou, A.-I., Gagliardi, G., Strypsteen, T., Badiei, M

    Kontras, K., Osselaer, T., Mouslech, S. G., Karaiskou, A.-I., Gagliardi, G., Strypsteen, T., Badiei, M. H., Rani, A., Vanmarcke, M., Bhagubai, M., Ekbote, C., Hwang, J., Chatzichristos, C., Liang, P. P., & De Vos, M. (2026). NeuroAtlas: Benchmarking Foundation Models for Clinical EEG and Brain-Computer Interfaces.arXiv:2605.14698

    Pith/arXiv arXiv 2026

  23. [23]

    Xiong, W., Li, J., Li, J., Zhu, K., & Jiang, C. (2026). EEG-FM-Bench: A Comprehensive Benchmark for the Systematic Evaluation and Diagnostic Analyses of EEG Foundation Models.International Conference on Machine Learning (ICML 2026). arXiv:2508.17742. 25

    arXiv 2026

  24. [24]

    Lu, Z., Li, Z., Shen, X., Lou, K., Xin, Y., Chen, X., Wang, S., Chen, X., Fan, J., Huang, C., Xu, X., Hou, Z., Wei, C., & Liu, Q. (2026). OmniEEG-Bench: A Standardized Evaluation Benchmark for EEG Foundation Models.arXiv:2606.00815

    Pith/arXiv arXiv 2026

  25. [25]

    ˇSirca, U., Alimardani, M., Zafeiriou, S., & Barmpas, K. (2026). Beyond Accuracy: Robust- ness, Interpretability and Expressiveness of EEG Foundation Models.arXiv:2605.17562

    Pith/arXiv arXiv 2026

  26. [26]

    T., et al

    Schirrmeister, R. T., et al. (2017). Deep learning with convolutional neural networks for EEG decoding and visualization.Human Brain Mapping, 38(11), 5391–5420

    2017

  27. [27]

    Dosovitskiy, A., et al. (2021). An image is worth 16x16 words: Transformers for image recognition at scale.Proc. ICLR

    2021

  28. [28]

    Shoeb, A. H. (2009). Application of machine learning to epileptic seizure onset detection and treatment.PhD thesis, MIT

    2009

  29. [29]

    & Picone, J

    Obeid, I. & Picone, J. (2016). The Temple University Hospital EEG data corpus.Frontiers in Neuroscience, 10, 196

    2016

  30. [30]

    Miltiadous, A., et al. (2023). A dataset of scalp EEG recordings of Alzheimer’s disease, frontotemporal dementia and healthy subjects from routine EEG.Data, 8(6), 95

    2023

  31. [31]

    Kemp, B., et al. (2000). Analysis of a sleep-dependent neuronal feedback loop.IEEE Trans. Biomedical Engineering, 47(9), 1185–1194

    2000

  32. [32]

    C., & Paik, J

    Kim, M.-J., Youn, Y. C., & Paik, J. (2023). Deep learning-based EEG analysis to classify normal, mild cognitive impairment, and dementia: Algorithms and dataset.NeuroImage, 272, 120054

    2023

  33. [33]

    Gramfort, A., et al. (2013). MEG and EEG data analysis with MNE-Python.Frontiers in Neuroscience, 7, 267

    2013

  34. [34]

    Welch, P. (1967). The use of fast Fourier transform for the estimation of power spectra. IEEE Trans. Audio and Electroacoustics, 15(2), 70–73

    1967

  35. [35]

    Hjorth, B. (1970). EEG analysis based on time domain properties.Electroencephalography and Clinical Neurophysiology, 29(3), 306–310

    1970

  36. [36]

    J., et al

    Donoghue, T., Haller, M., Peterson, E. J., et al. (2020). Parameterizing neural power spectra into periodic and aperiodic components.Nature Neuroscience, 23(12), 1655–1665

    2020

  37. [37]

    Botvinik-Nezer, R., et al. (2020). Variability in the analysis of a single neuroimaging dataset by many teams.Nature, 582, 84–88. DOI: 10.1038/s41586-020-2314-9

    doi:10.1038/s41586-020-2314-9 2020

  38. [38]

    van Dijk, H., van Wingen, G., Denys, D., Olbrich, S., van Ruth, R., & Arns, M. (2022). The two decades brainclinics research archive for insights in neurophysiology (TDBRAIN) database.Scientific Data, 9, 333

    2022

  39. [39]

    Cavanagh, J. F. (2021). EEG: 3-Stim Auditory Oddball and Rest in Parkinson’s Disease. OpenNeuro, dataset ds003490

    2021

  40. [40]

    J., Shen, Y., Wallis, P., et al

    Hu, E. J., Shen, Y., Wallis, P., et al. (2022). LoRA: Low-Rank Adaptation of Large Language Models.Proc. ICLR

    2022

  41. [41]

    Kaplan, J., et al. (2020). Scaling laws for neural language models.arXiv:2001.08361

    Pith/arXiv arXiv 2020

  42. [42]

    Zare, M. (2026). Foundation Models for EEG Are Blind to Long-Range Temporal Cor- relations: A Spectral–Temporal Dissociation Behind Their Cross-Population Fragility. Companion manuscript, under review, 2026. 26

    2026

  43. [43]

    L., et al

    Goldberger, A. L., et al. (2000). PhysioBank, PhysioToolkit, and PhysioNet.Circulation, 101(23), e215–e220. 27

    2000