Pith. sign in

REVIEW 3 major objections 5 minor 53 references

Across 20 motor-imagery EEG datasets, Bayesian complete-pooling improves reliability only slightly and at 13x the energy cost, leaving accuracy and discrimination unchanged.

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

2026-08-01 03:58 UTC pith:I7QQJW4C

load-bearing objection A well-executed negative-result benchmark whose practical conclusion survives the imperfect Bayesian/frequentist pairing, though the causal story about Bayesian averaging is weaker than the paper claims. the 3 major comments →

arxiv 2607.22980 v1 pith:I7QQJW4C submitted 2026-07-25 cs.LG

Bayesian Complete-Pooling in Cross-Subject Classification for Motor Imagery Electroencephalogram

Ethan Davis This is my paper
classification cs.LG
keywords Bayesian complete-poolingmotor imageryelectroencephalogramprobabilistic calibrationBrier scorerandom-effects meta-analysisbrain-computer interfaceMCMC posterior sampling
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper asks whether fully Bayesian, complete-pooling classifiers make practically better predictions for cross-subject left/right-hand motor-imagery EEG, a task where calibration matters because signals are nonstationary and point-estimate models risk overconfidence. The author paired six frequentist pipelines with Bayesian analogues that share identical feature engineering, fitted all by MCMC posterior sampling, and compared them across 20 datasets using Brier score decomposed into reliability and resolution, plus AUROC and Shannon entropy. The central claim is that Bayesian complete-pooling produces statistically detectable but practically negligible improvements in reliability and increases predictive uncertainty, while leaving Brier score, resolution, and discrimination unchanged. The author concludes that replacing point-estimate classifiers with Bayesian complete-pooling offers limited practical benefit for this task, and that hierarchical partial-pooling across subjects and sessions is the more promising direction. The finding matters because it disciplines the expectation that 'going Bayesian' alone fixes calibration in BCI, redirecting effort toward modeling group structure.

Core claim

The central discovery is that for cross-subject left/right hand motor imagery EEG, Bayesian complete-pooling — posterior averaging over all subjects combined, with no hierarchical structure — does not deliver practically meaningful gains. Across 20 datasets, the pooled effect on Brier score, resolution, and AUROC was statistically indistinguishable from zero, while reliability improved by 0.0015 (a statistically significant but tiny effect) and Shannon entropy increased by 0.0121 (predictions became more uncertain, i.e., less sharp). The author interprets the null discrimination result and positive reliability/sharpness results as showing that Bayesian model averaging smooths overconfident p

What carries the argument

The pair-comparison design: six frequentist pipelines (CSP+LDA, CSP+SVM, TS+LR, TS+SVM, SCNN, DCNN) each matched with a Bayesian counterpart (BLDA, Gaussian process classifier, BLR, Gaussian process classifier, Bayesian-last-layer SCNN/DCNN) sharing identical deterministic feature engineering, so the only difference is point-estimate optimization versus MCMC/NUTS posterior sampling and Bayesian model averaging. Performance is aggregated with random-effects meta-analysis (REML + Knapp-Hartung) on per-dataset effect sizes, using the Brier score decomposed into reliability and resolution plus AUROC and Shannon entropy. The mechanism speculated to produce the observed smoothing is the concentrat

Load-bearing premise

The Bayesian pipelines are faithful analogues of their frequentist controls, so the observed differences are attributable to Bayesian posterior averaging rather than to changing the model family; the paper itself concedes this mapping is imperfect for the top-2 Riemannian pipelines.

What would settle it

A replication of the same 20-dataset comparison — or a subset of them — in which the Bayesian and frequentist models are the exact same model class (e.g., an SVM and a Bayesian SVM with identical kernel and loss) and in which hyperparameters are optimized, could overturn the conclusion if it yields a practically meaningful Brier or reliability improvement (for example, a Brier difference larger than ~0.005).

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

If this is right

  • If correct, replacing frequentist classifiers with full Bayesian complete-pooling versions in cross-subject MI-EEG will not materially improve calibration or discrimination, so such swaps should not be the primary strategy for calibration-free BCI.
  • The statistically significant but practically negligible reliability gain, paired with a null Brier change, implies that point-estimate classifiers in this setting are only mildly overconfident — the reliability problem is smaller than often assumed.
  • The 13x energy overhead with no Brier/AUROC benefit means Bayesian complete-pooling is hard to justify on accuracy grounds alone, though the authors note the absolute cost is modest compared to household appliances.
  • The sensitivity of the reliability result to leave-one-out influence analysis (55% of iterations null) means conclusions about calibration improvement are dataset-dependent and should not be overgeneralized.
  • The null discrimination result suggests the learned decision boundary is largely determined by feature engineering and pooling structure, not by whether the classifier averages over parameters.

Where Pith is reading between the lines

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

  • The paper's own logic implies that the same benchmark applied to partial-pooling (hierarchical) Bayesian models — where subject and session effects are explicitly modeled — would likely show larger calibration and discrimination gains; this is the author's stated future direction and a natural next experiment.
  • A stronger test of the 'Bayesian model averaging works in principle' hypothesis would be to use exactly the same model class on both sides (e.g., an SVM and a Bayesian SVM with the same kernel and loss) to rule out model-family mismatch as the cause of the null result; the paper's GPC-for-SVM pairing is an approximation.
  • One might predict from the typical-set argument that the reliability gain from Bayesian averaging grows as the feature dimensionality or model capacity increases, because the mode's neighborhood occupies a smaller fraction of the posterior mass — a testable extension that would refine when complete-pooling is worth its cost.
  • If the goal is calibration-aware BCIs, these results suggest reporting Brier score and reliability alongside AUROC in future benchmarks, since discrimination alone missed the (small) calibration gains.

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

3 major / 5 minor

Summary. The paper reports a large-scale benchmark comparing six Bayesian complete-pooling classifiers with paired frequentist baselines for cross-subject left/right-hand MI-EEG classification across 20 MOABB datasets. Using Brier score (decomposed into reliability and resolution), AUROC, and Shannon entropy, the authors fit random-effects meta-analyses (REML, Knapp–Hartung) to held-out probabilistic predictions. They find non-significant pooled differences in Brier score, resolution, and AUROC, but statistically significant improvements in reliability (p = 0.0396) and reduced sharpness (higher entropy, p = 0.0195), with low between-study heterogeneity. The reliability result is fragile under leave-one-out influence analysis, and Bayesian pipelines consume roughly 13.1× more energy. The paper concludes that Bayesian complete-pooling alone offers limited practical benefit and suggests partial pooling as a more promising direction.

Significance. If the result holds, this is a useful negative result for the BCI calibration literature: it challenges the assumption that Bayesian model averaging alone will materially improve cross-subject probabilistic predictions. Strengths include the use of an external benchmark (MOABB), 20 datasets, held-out probabilistic predictions, a proper scoring rule (Brier score with CORP decomposition), random-effects meta-analysis with REML and Knapp–Hartung adjustment, LOO influence diagnostics, and publicly archived code/data. The main weakness is that the paired Bayesian pipelines are not always faithful analogues of their frequentist controls, particularly for the deep-learning pairs, which undermines the causal attribution of the observed reliability/sharpness differences to Bayesian complete-pooling. The practical negative conclusion is less affected by this confound, so the paper remains defensible after revision.

major comments (3)
  1. [§3.3.2, Table 2] The Bayesian neural network pairs (BSCNN/BDCNN) are described as Bayesian last-layer (BLL) models: the SCNN/DCNN feature extractors are first trained end-to-end with a deterministic linear head, then the head is replaced by BLR. The frequentist controls are trained end-to-end as full classifiers. This changes the feature representation and training objective in addition to introducing Bayesian averaging, so the reliability/sharpness effects for these two pairs cannot be attributed solely to Bayesian complete-pooling. Since two of six pairs carry this confound, the causal claim in §5.1 that 'Bayesian classification is doing real work' is not established for the deep-learning pipelines. The authors should either re-run with a fully Bayesian or otherwise clean deep-learning analogue, or restrict the causal interpretation to the non-confounded pairs and present the BNN results as exploratory
  2. [§3.3.2, 'As a Bayesian analog to SVMs'] The GPC treatment differs from the SVM control in more than posterior averaging: GPC uses a probit likelihood, hand-set hyperpriors (amplitude, lengthscale), and sparse DTC approximation with 100 inducing points, whereas the SVM uses an RBF kernel with Platt scaling. The MAP-equivalence argument is asymptotic and does not guarantee that posterior predictive calibration matches the SVM's Platt-scaled output. Thus the CSP+GP and TS+GP effect sizes also include model-class and approximation differences. The paper should explicitly state the counterfactual as 'Bayesian versions of our chosen model families' and consider sensitivity to GP hyperparameters and inducing-point counts.
  3. [§4.2, Table 4, LOO analysis] The reliability effect is statistically significant only marginally (p = 0.0396) and is fragile: 11/20 leave-one-out iterations yield a null pooled effect, and the 95% prediction interval crosses zero. The abstract discloses this sensitivity, but §5.1 states that the statistical significance 'tells us that ... Bayesian classification is doing real work.' Given the LOO fragility and the confounding issues above, this interpretation is too strong. Please temper the language or provide additional robustness checks (e.g., sensitivity to the meta-analytic estimator, exclusion of influential studies).
minor comments (5)
  1. [§2.3] Typo: 'DerSimionian-Laird' should be 'DerSimonian-Laird'.
  2. [Table 2] The table references 'Zhou2016' in the Chevallier et al. list, but the dataset table (Table 1) uses 'Zhou2020'; clarify whether these refer to the same dataset or different sources.
  3. [Table 3] The column headers 'ESS bulk <400 Min. ESS bulk' and 'ESStail <400 Min. ESS tail' are unclear; suggest separate columns such as 'Min ESS bulk', 'Min ESS tail', 'Max R-hat'.
  4. [§3.3.1] For CSP, 'we decomposed 6 components along orthogonal discriminative axes' is vague; specify that CSP with 6 components was used.
  5. [§4.3] The energy measurements in Table 6 should state whether they are per training run, per fold, or per pipeline across all folds; the current text is ambiguous.

Circularity Check

0 steps flagged

No circularity: the paper's claims are empirical comparisons of held-out predictions, with no fitted quantity renamed as a prediction and no load-bearing self-citation.

full rationale

The paper's derivation chain is empirical rather than definitional. For each of 20 datasets, pairwise effect sizes are computed from held-out LOSO/10-fold predictions and synthesized with REML/Knapp-Hartung random-effects meta-analysis. The Bayesian-versus-frequentist contrast is implemented through paired pipelines with shared feature extraction; the central results are measured differences in Brier score, reliability, resolution, AUROC, and Shannon entropy. No parameter is fitted to the target metric and then reported as a prediction. The Bayesian analogues (GPC for SVM, BLR for LR, BLL for CNN, BLDA for LDA) are model-specification choices whose faithfulness could be questioned, but that is an internal-validity concern, not circularity. The only self-referential element is Section 5.3's speculative claim that partial pooling 'would yield a greater effect,' which the paper explicitly frames as a belief for future work rather than a conclusion derived from the present data. The Limitations section acknowledges omitted HPO and the two-stage meta-analysis design, but these are acknowledged limitations, not circular reductions. The external MOABB/Chevallier benchmarks and the cited GPC/SVM theory are independent support; there are no self-citations, no imported uniqueness theorems, and no ansatz smuggled in via author self-citation. Accordingly, no circular step meets the required evidentiary standard.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The paper is empirical; the ledger consists of hand-chosen hyperparameters that shape the Bayesian posterior approximation, the structural analogy assumptions behind the paired design, and standard meta-analytic assumptions. No invented entities.

free parameters (5)
  • GPC priors (amplitude eta ~ HN(1); lengthscale l ~ LN(0,0.5) for the RBF-kernel GPC) = eta scale 1; l log-scale 0.5
    Hand-chosen hyperpriors in Section 3.3.2. The calibration and sharpness effects that constitute the significant findings depend on how informative these priors are; no HPO was run.
  • GPC sparse approximation: number of inducing points = 100 (k-means)
    Hand-chosen in Section 3.3.2; limits the fidelity of the posterior that is the treatment under test.
  • MCMC sampling budget = 1000 warmup + 1000 tuning draws; target acceptance 0.95; 4 chains
    Hand-chosen in Section 3.3.2; finite MCMC error is part of what distinguishes the realized Bayesian treatment from an idealized posterior average.
  • CSP component count = 6
    Hand-chosen in Section 3.3.1; shared across paired pipelines, so it does not bias the contrast.
  • Neural network hyperparameters (lr=1e-3, weight decay=1e-2, max 300 epochs, batch 64, early-stopping patience 150) = as listed in Section 3.3.1
    Chosen without HPO (disclosed in Section 5.2); shared between frequentist and Bayesian variants, so it affects absolute performance and possibly the size of any treatment effect.
axioms (5)
  • domain assumption Exchangeability of the 20 datasets in the random-effects meta-analysis (zeta_k independent of k)
    Stated in Section 2.3 (Equation 5). If the datasets are not exchangeable, the pooled effect and its confidence interval are not interpretable.
  • ad hoc to paper Paired Bayesian models are faithful analogues of the frequentist controls (SVM to GP, LR to BLR/GPC, LDA to BLDA, CNN to BLL)
    Invoked in Section 3.3.2. The causal reading of the results presupposes the analogy; the paper admits the top-2 Riemannian pipelines 'did not map cleanly to a Bayesian ML treatment.'
  • ad hoc to paper Cross-subject (fully pooled) evaluation is the regime where uncertainty-quantification benefits would be most visible
    Stated in Section 3.4. The scope of the conclusion ('Bayesian complete-pooling alone offers limited benefit') is only meaningful if this regime is favorable to the treatment.
  • standard math Brier score is a proper scoring rule and the CORP decomposition yields the stated reliability/resolution components
    Invoked in Sections 2.2 and 3.4 via references [5], [41], [42]; standard results taken from the literature.
  • domain assumption The 'typical set' argument explains the sharpness decrease
    Section 5.1 uses the volume-of-the-typical-set argument to explain why Bayesian averaging smooths predictions; this is an interpretive assumption, not a result established by this study.

pith-pipeline@v1.3.0-alltime-deepseek · 14077 in / 19310 out tokens · 182687 ms · 2026-08-01T03:58:07.901328+00:00 · methodology

0 comments
read the original abstract

Brain-computer interfaces (BCIs) have long sought calibration-free operation, but classifiers are typically benchmarked by discrimination alone, blind to whether predicted probabilities are well calibrated - a meaningful gap given nonstationary electroencephalogram (EEG) signals and the risk of overconfident point-estimate classifiers under distribution shift. We conducted a large-scale study contrasting Bayesian complete-pooling models against frequentist baselines for cross-subject, left-hand versus right-hand motor imagery EEG classification across 20 datasets. Six frequentist pipelines were each paired with an analogous Bayesian pipeline sharing identical feature engineering, fit via Markov chain Monte Carlo posterior sampling. Our primary metric was the Brier score, decomposed into reliability and resolution, alongside AUROC for discrimination and Shannon entropy for sharpness. Each metric was analyzed via random-effects meta-analysis (REML, Knapp-Hartung adjustment), verified by leave-one-out influence analysis. Bayesian complete-pooling produced statistically but not practically significant improvements in reliability and increases in predictive uncertainty (lower sharpness); Brier score, resolution, and discrimination showed no significant differences. Between-study heterogeneity was low across all metrics, though the reliability result was sensitive to leave-one-out removal. We additionally profiled computational cost, finding that Bayesian pipelines consumed roughly thirteen times more energy than their frequentist counterparts, a cost that remains modest relative to common household appliances. These results suggest that Bayesian complete-pooling alone offers limited practical benefit for cross-subject motor imagery classification, and that partial-pooling across subjects and sessions is a more promising direction for future work.

Figures

Figures reproduced from arXiv: 2607.22980 by Ethan Davis.

Figure 1
Figure 1. Figure 1: The MI-EEG datasets after preprocessing. [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The PGM of BLDA where P is the number of trials and Q is the number of electrodes. and fit a probabilistic BLR predictor as a replacement. SVMs belong to a category of kernel methods [7]. As a Bayesian analog to SVMs we used Gaussian processes classifiers (GPCs). The MAP of GPCs with probit loss yields highly similar solutions to SVMs with its characteristic hinge loss function [40]. For CSP features with … view at source ↗
Figure 3
Figure 3. Figure 3: The Brier score forest plot [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Forest and Baujat plots for reliability. [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Forest and Baujat plots for Shannon entropy. [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

53 extracted references · 8 canonical work pages

  1. [1]

    Transfer learning for BCIs

    Vinay Jayaram, Karl-Heinz Fiebig, Jan Peters, and Moritz Grosse-Wentrup. Transfer learning for BCIs. In Chang S. Nam, Anton Nijholt, and Fabien Lotte, editors,Brain–Computer Interfaces Handbook: Technological and Theoretical Advances, pages 425–441. CRC Press, Boca Raton, FL, 2018. ISBN 978-1-4987-7343-0. doi: 10.1201/9781351231954-22

    work page doi:10.1201/9781351231954-22 2018

  2. [2]

    Bayesian learning for EEG analysis

    Yu Zhang. Bayesian learning for EEG analysis. In Chang S. Nam, Anton Nijholt, and Fabien Lotte, editors,Brain–Computer Interfaces Handbook: Technological and Theoretical Advances, 16 pages 407–424. CRC Press, Boca Raton, FL, 2018. ISBN 978-1-4987-7343-0. doi: 10.1201/ 9781351231954-21

    2018

  3. [3]

    The largest eeg-based bci reproducibility study for open science: the moabb benchmark

    Sylvain Chevallier, Igor Carrara, Bruno Aristimunha, Pierre Guetschel, Sara Sedlar, Bruna Junqueira Lopes, S´ ebastien Velut, Salim Khazem, and Thomas Moreau. The largest eeg-based bci reproducibility study for open science: the moabb benchmark. working paper or preprint, 4 2024. URLhttps://universite-paris-saclay.hal.science/hal-04537061

    2024

  4. [4]

    Sculley, Sebastian Nowozin, Joshua V

    Yaniv Ovadia, Emily Fertig, Jie Ren, Zachary Nado, D. Sculley, Sebastian Nowozin, Joshua V. Dillon, Balaji Lakshminarayanan, and Jasper Snoek. Can you trust your model’s uncertainty? evaluating predictive uncertainty under dataset shift, 2019

    2019

  5. [5]

    Daniel S. Wilks. Chapter 9 - forecast verification. InStatistical Methods in the Atmospheric Sciences (Fourth Edition), pages 369–483. Elsevier, fourth edition edition, 2019. ISBN 978-0- 12-815823-4. doi: https://doi.org/10.1016/B978-0-12-815823-4.00009-2. URLhttps://www. sciencedirect.com/science/article/pii/B9780128158234000092

    doi:10.1016/b978-0-12-815823-4.00009-2 2019

  6. [6]

    Murphy.Probabilistic Machine Learning: An Introduction

    Kevin P. Murphy.Probabilistic Machine Learning: An Introduction. MIT Press, 2022. URL http://probml.github.io/book1

    2022

  7. [7]

    Bishop.Pattern Recognition and Machine Learning

    Christopher M. Bishop.Pattern Recognition and Machine Learning. Information Science and Statistics. Springer New York, New York, NY, 1 edition, 2006. ISBN 978-0-387-31073-2

    2006

  8. [8]

    Bishop and Hugh Bishop.Deep Learning: Foundations and Concepts

    Christopher M. Bishop and Hugh Bishop.Deep Learning: Foundations and Concepts. Springer Cham, 1 edition, 2023. ISBN 978-3-031-45467-7. doi: 10.1007/978-3-031-45468-4. URL https://link.springer.com/book/10.1007/978-3-031-45468-4

    doi:10.1007/978-3-031-45468-4 2023

  9. [9]

    Murphy.Probabilistic Machine Learning: Advanced Topics

    Kevin P. Murphy.Probabilistic Machine Learning: Advanced Topics. MIT Press, 2023. URL http://probml.github.io/book2

    2023

  10. [10]

    Chapman and Hall/CRC, 2 edition, 2020

    Richard McElreath.Statistical Rethinking: A Bayesian Course with Examples in R and STAN. Chapman and Hall/CRC, 2 edition, 2020. doi: 10.1201/9780429029608

    doi:10.1201/9780429029608 2020

  11. [11]

    Tilmann Gneiting, Fadoua Balabdaoui, and Adrian E. Raftery. Probabilistic forecasts, cali- bration and sharpness.Journal of the Royal Statistical Society Series B: Statistical Method- ology, 69(2):243–268, 04 2007. ISSN 1369-7412. doi: 10.1111/j.1467-9868.2007.00587.x. URL https://doi.org/10.1111/j.1467-9868.2007.00587.x

    arXiv 2007

  12. [12]

    Duda, Peter E

    Richard O. Duda, Peter E. Hart, and David G. Stork.Pattern Classification. Wiley- Interscience, New York, 2nd edition, 2001. ISBN 0-471-05669-3

    2001

  13. [13]

    Weinberger

    Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. On calibration of modern neural networks. In Doina Precup and Yee Whye Teh, editors,Proceedings of the 34th International Conference on Machine Learning, volume 70 ofProceedings of Machine Learning Research, pages 1321–1330. PMLR, 2017

    2017

  14. [14]

    Chapman & Hall/CRC Press, Boca Raton, FL and London, 1st edition, 2021

    Mathias Harrer, Pim Cuijpers, Furukawa Toshi A, and David D Ebert.Doing Meta-Analysis With R: A Hands-On Guide. Chapman & Hall/CRC Press, Boca Raton, FL and London, 1st edition, 2021. ISBN 9780367610074

    2021

  15. [15]

    Hedges, Julian P

    Michael Borenstein, Larry V. Hedges, Julian P. T. Higgins, and Hannah R. Rothstein.Intro- duction to Meta-Analysis. Wiley, 2009. doi: 10.1002/9780470743386. 17

    doi:10.1002/9780470743386 2009

  16. [16]

    Moabb: trustworthy algorithm benchmarking for bcis.Journal of Neural Engineering, 15(6):066011, 9 2018

    Vinay Jayaram and Alexandre Barachant. Moabb: trustworthy algorithm benchmarking for bcis.Journal of Neural Engineering, 15(6):066011, 9 2018. doi: 10.1088/1741-2552/aadea0. URLhttps://doi.org/10.1088/1741-2552/aadea0

    doi:10.1088/1741-2552/aadea0 2018

  17. [17]

    Miller, Gernot Mueller-Putz, Guido Nolte, Gert Pfurtscheller, Hubert Preissl, Gerwin Schalk, Alois Schl¨ ogl, Carmen Vidaurre, Stephan Waldert, and Benjamin Blankertz

    Michael Tangermann, Klaus-Robert M¨ uller, Ad Aertsen, Niels Birbaumer, Christoph Braun, Clemens Brunner, Robert Leeb, Carsten Mehring, Kai J. Miller, Gernot Mueller-Putz, Guido Nolte, Gert Pfurtscheller, Hubert Preissl, Gerwin Schalk, Alois Schl¨ ogl, Carmen Vidaurre, Stephan Waldert, and Benjamin Blankertz. Review of the bci competition iv.Frontiers in ...

    Pith/arXiv arXiv 2012

  18. [18]

    Motor imagery under distraction— an open ac- cess bci dataset.Frontiers in Neuroscience, Volume 14 - 2020, 2020

    Stephanie Brandl and Benjamin Blankertz. Motor imagery under distraction— an open ac- cess bci dataset.Frontiers in Neuroscience, Volume 14 - 2020, 2020. ISSN 1662-453X. doi: 10.3389/fnins.2020.566147. URLhttps://www.frontiersin.org/journals/neuroscience/ articles/10.3389/fnins.2020.566147

    arXiv 2020

  19. [19]

    A multi-paradigm EEG dataset for studying upper limb rehabilitation exercises.Scientific Data, 12(1):1877, 2025

    Wenwen Chang, Weixuan Kong, Guanghui Yan, Renjie Lv, Kaiyue Du, Muhammad Tariq Sadiq, Bin Guo, Rong Yin, and Xuan Liu. A multi-paradigm EEG dataset for studying upper limb rehabilitation exercises.Scientific Data, 12(1):1877, 2025. ISSN 2052-4463. doi: 10.1038/s41597-025-06147-6. URLhttps://doi.org/10.1038/s41597-025-06147-6

    work page doi:10.1038/s41597-025-06147-6 2025

  20. [20]

    Eeg datasets for motor imagery brain–computer interface.GigaScience, 6(7):gix034, 5 2017

    Hohyun Cho, Minkyu Ahn, Sangtae Ahn, Moonyoung Kwon, and Sung Chan Jun. Eeg datasets for motor imagery brain–computer interface.GigaScience, 6(7):gix034, 5 2017. ISSN 2047-217X. doi: 10.1093/gigascience/gix034. URLhttps://doi.org/10.1093/gigascience/ gix034

    doi:10.1093/gigascience/gix034 2017

  21. [21]

    Experimenters’ influence on mental-imagery based brain-computer interface user training.International Journal of Human-Computer Studies, 149:102603, 2021

    L´ ea Pillette, Aline Roc, Bernard N’Kaoua, and Fabien Lotte. Experimenters’ influence on mental-imagery based brain-computer interface user training.International Journal of Human-Computer Studies, 149:102603, 2021. ISSN 1071-5819. doi: https://doi.org/10. 1016/j.ijhcs.2021.102603. URLhttps://www.sciencedirect.com/science/article/pii/ S1071581921000215

    arXiv 2021

  22. [22]

    When should MI-BCI feature optimization include prior knowledge, and which one?Brain-Computer Interfaces, 9(2):115–128, 4 2022

    Camille Benaroch, Maria Sayu Yamamoto, Aline Roc, Pauline Dreyer, Camille Jeunet, and Fabien Lotte. When should MI-BCI feature optimization include prior knowledge, and which one?Brain-Computer Interfaces, 9(2):115–128, 4 2022. doi: 10.1080/2326263X.2022.2033073. URLhttps://inria.hal.science/hal-03920680

    arXiv 2022

  23. [23]

    Integrat- ing simultaneous motor imagery and spatial attention for eeg-bci control.IEEE Transactions on Biomedical Engineering, 71(1):282–294, 2024

    Dylan Forenzo, Yixuan Liu, Jeehyun Kim, Yidan Ding, Taehyung Yoon, and Bin He. Integrat- ing simultaneous motor imagery and spatial attention for eeg-bci control.IEEE Transactions on Biomedical Engineering, 71(1):282–294, 2024. doi: 10.1109/TBME.2023.3298957

    arXiv 2024

  24. [24]

    Beamforming in noninvasive brain–computer interfaces.IEEE Transactions on Biomedical Engineering, 56 (4):1209–1219, 2009

    Moritz Grosse-Wentrup, Christian Liefhold, Klaus Gramann, and Martin Buss. Beamforming in noninvasive brain–computer interfaces.IEEE Transactions on Biomedical Engineering, 56 (4):1209–1219, 2009. doi: 10.1109/TBME.2008.2009768

    arXiv 2009

  25. [25]

    Dataset combining EEG, eye- tracking, and high-speed video for ocular activity analysis across BCI paradigms.Sci- entific Data, 12(1):587, 2025

    Eva Guttmann-Flury, Xinjun Sheng, and Xiangyang Zhu. Dataset combining EEG, eye- tracking, and high-speed video for ocular activity analysis across BCI paradigms.Sci- entific Data, 12(1):587, 2025. ISSN 2052-4463. doi: 10.1038/s41597-025-04861-9. URL https://doi.org/10.1038/s41597-025-04861-9. 18

    work page doi:10.1038/s41597-025-04861-9 2025

  26. [26]

    HEFMI-ICH: a hybrid EEG-fNIRS motor imagery dataset for brain-computer interface in intracerebral hemorrhage.Scientific Data, 12(1):1816, 2025

    Jian Shi, Danyang Chen, Xingwei Zhao, Zhixian Zhao, Shengjie Li, Yeguang Xu, Tao Ding, Zheng Zhu, Peng Zhang, Qing Ye, Yingxin Tang, Ping Zhang, Bo Tao, and Zhouping Tang. HEFMI-ICH: a hybrid EEG-fNIRS motor imagery dataset for brain-computer interface in intracerebral hemorrhage.Scientific Data, 12(1):1816, 2025. ISSN 2052-4463. doi: 10.1038/ s41597-025-...

    work page doi:10.1038/s41597-025-06100-7 2025

  27. [27]

    Transfer learning promotes acquisition of individual bci skills.PNAS Nexus, 3(2): pgae076, 02 2024

    Satyam Kumar, Hussein Alawieh, Frigyes Samuel Racz, Rawan Fakhreddine, and Jos´ e del R Mill´ an. Transfer learning promotes acquisition of individual bci skills.PNAS Nexus, 3(2): pgae076, 02 2024. ISSN 2752-6542. doi: 10.1093/pnasnexus/pgae076. URLhttps://doi. org/10.1093/pnasnexus/pgae076

    work page doi:10.1093/pnasnexus/pgae076 2024

  28. [28]

    Eeg dataset and openbmi toolbox for three bci paradigms: an investigation into bci illiteracy.GigaScience, 8(5):giz002, 1 2019

    Min-Ho Lee, O-Yeon Kwon, Yong-Jeong Kim, Hong-Kyung Kim, Young-Eun Lee, John Williamson, Siamac Fazli, and Seong-Whan Lee. Eeg dataset and openbmi toolbox for three bci paradigms: an investigation into bci illiteracy.GigaScience, 8(5):giz002, 1 2019. ISSN 2047-217X. doi: 10.1093/gigascience/giz002. URLhttps://doi.org/10.1093/gigascience/ giz002

    doi:10.1093/gigascience/giz002 2019

  29. [29]

    EEG datasets of stroke patients, 12 2022

    Haijie Liu and Xiaodong Lv. EEG datasets of stroke patients, 12 2022. URLhttps:// figshare.com/articles/dataset/EEG_datasets_of_stroke_patients/21679035

    arXiv 2022

  30. [30]

    An EEG Motor Imagery Dataset for Brain Computer Interface in Acute Stroke Patients.Scientific Data, 11(1):131,

    Haijie Liu, Penghu Wei, Haochong Wang, Xiaodong Lv, Wei Duan, Meijie Li, Yan Zhao, Qingmei Wang, Xinyuan Chen, Gaige Shi, Bo Han, and Junwei Hao. An EEG Motor Imagery Dataset for Brain Computer Interface in Acute Stroke Patients.Scientific Data, 11(1):131,

  31. [31]

    A. L. Goldberger, L. A. Amaral, L. Glass, J. M. Hausdorff, P. C. Ivanov, R. G. Mark, J. E. Mietus, G. B. Moody, C. K. Peng, and H. E. Stanley. PhysioBank, PhysioToolkit, and Phys- ioNet: Components of a New Research Resource for Complex Physiologic Signals.Circulation, 101(23):e215–e220, 2000. doi: 10.1161/01.cir.101.23.e215

    doi:10.1161/01.cir.101.23.e215 2000

  32. [32]

    Deep learning with convolutional neural networks for eeg decoding and visu- alization.Human Brain Mapping, 38(11):5391–5420, 2017

    Robin Tibor Schirrmeister, Jost Tobias Springenberg, Lukas Dominique Josef Fiederer, Martin Glasstetter, Katharina Eggensperger, Michael Tangermann, Frank Hutter, Wolfram Burgard, and Tonio Ball. Deep learning with convolutional neural networks for eeg decoding and visu- alization.Human Brain Mapping, 38(11):5391–5420, 2017. doi: https://doi.org/10.1002/h...

    doi:10.1002/hbm 2017

  33. [33]

    Open access dataset for eeg+nirs single-trial classification.IEEE Transactions on Neural Systems and Rehabilitation Engineering, 25(10): 1735–1745, 2017

    Jaeyoung Shin, Alexander von L¨ uhmann, Benjamin Blankertz, Do-Won Kim, Jichai Jeong, Han-Jeong Hwang, and Klaus-Robert M¨ uller. Open access dataset for eeg+nirs single-trial classification.IEEE Transactions on Neural Systems and Rehabilitation Engineering, 25(10): 1735–1745, 2017. doi: 10.1109/TNSRE.2016.2628057

    arXiv 2017

  34. [34]

    Stieger, Stephen A

    James R. Stieger, Stephen A. Engel, and Bin He. Continuous Sensorimotor Rhythm Based Brain Computer Interface Learning in a Large Population.Scientific Data, 8(1):98, 2021. doi: 10.1038/s41597-021-00883-1

    doi:10.1038/s41597-021-00883-1 2021

  35. [35]

    Evaluation of eeg oscillatory patterns and cognitive process during simple and compound limb motor imagery.PLOS ONE, 9(12):1–19, 12 2014

    Weibo Yi, Shuang Qiu, Kun Wang, Hongzhi Qi, Lixin Zhang, Peng Zhou, Feng He, and Dong Ming. Evaluation of eeg oscillatory patterns and cognitive process during simple and compound limb motor imagery.PLOS ONE, 9(12):1–19, 12 2014. doi: 10.1371/journal.pone.0114853. URLhttps://doi.org/10.1371/journal.pone.0114853. 19

    doi:10.1371/journal.pone.0114853 2014

  36. [36]

    A multi-day and high-quality EEG dataset for motor imagery brain-computer interface.Scientific Data, 12(1):488, 2025

    Banghua Yang, Fenqi Rong, Yunlong Xie, Du Li, Jiayang Zhang, Fu Li, Guangming Shi, and Xiaorong Gao. A multi-day and high-quality EEG dataset for motor imagery brain-computer interface.Scientific Data, 12(1):488, 2025. ISSN 2052-4463. doi: 10.1038/s41597-025-04826-y. URLhttps://doi.org/10.1038/s41597-025-04826-y

    doi:10.1038/s41597-025-04826-y 2025

  37. [37]

    Relative power correlates with the decoding performance of motor imagery both across time and subjects

    Qing Zhou, Jiafan Lin, Lin Yao, Yueming Wang, Yan Han, and Kedi Xu. Relative power correlates with the decoding performance of motor imagery both across time and subjects. Frontiers in Human Neuroscience, 15, 2021. ISSN 1662-5161. doi: 10.3389/fnhum.2021.701091. URLhttps://www.frontiersin.org/journals/human-neuroscience/articles/10.3389/ fnhum.2021.701091

    arXiv 2021

  38. [38]

    Cambridge University Press, 2020

    Marc Peter Deisenroth, Aldo Faisal, and Cheng Soon Ong.Mathematics for Ma- chine Learning. Cambridge University Press, 2020. ISBN 9781108455145. URL https://www.cambridge.org/au/universitypress/subjects/computer-science/ pattern-recognition-and-machine-learning/mathematics-machine-learning?format= PB

    2020

  39. [39]

    Springer Series in Statistics

    Trevor Hastie, Robert Tibshirani, and Jerome Friedman.The Elements of Statistical Learning. Springer Series in Statistics. Springer New York, New York, NY, 2 edition, 02 2009. ISBN 978-0-387-84857-0. doi: 10.1007/978-0-387-84858-7

    doi:10.1007/978-0-387-84858-7 2009

  40. [40]

    Carl Edward Rasmussen and Christopher K. I. Williams.Gaussian Processes for Machine Learning. The MIT Press, 11 2005. ISBN 9780262256834. doi: 10.7551/mitpress/3206.001

    doi:10.7551/mitpress/3206.001 2005

  41. [41]

    Allan H. Murphy. A new vector partition of the probability score.Journal of Applied Meteo- rology, 12(4):595–600, 1973. doi: 10.1175/1520-0450(1973)012⟨0595:ANVPOT⟩2.0.CO;2

    doi:10.1175/1520-0450(1973)012 1973

  42. [42]

    URLhttps://doi.org/10.7551/mitpress/3206.001.0001

    doi:10.7551/mitpress/3206.001.0001

  43. [43]

    David J. C. MacKay.Information Theory, Inference, and Learning Algorithms. Cambridge University Press, Cambridge, UK, 2003. ISBN 0-521-64298-1

    2003

  44. [44]

    Timo Dimitriadis, Tilmann Gneiting, and Alexander I. Jordan. Stable reliability diagrams for probabilistic classifiers.Proceedings of the National Academy of Sciences, 118(8):e2016191118,

  45. [45]

    Julian P. T. Higgins and Simon G. Thompson. Quantifying heterogeneity in a meta-analysis. Statistics in Medicine, 21(11):1539–1558, 6 2002. ISSN 0277-6715. doi: 10.1002/sim.1186

    doi:10.1002/sim.1186 2002

  46. [46]

    Panel usages ´ electrodomestiques ann´ ee 6, 3 2026

    La librairie ADEME. Panel usages ´ electrodomestiques ann´ ee 6, 3 2026. URLhttps: //librairie.ademe.fr/batiment/9155-panel-usages-electrodomestiques-annee-6. html

    2026

  47. [47]

    Martin, Ravin Kumar, and Junpeng Lao.Bayesian Modeling and Computation in Python

    Osvaldo A. Martin, Ravin Kumar, and Junpeng Lao.Bayesian Modeling and Computation in Python. Chapman and Hall/CRC, 1 edition, 2021. doi: 10.1201/9781003019169. URL https://doi.org/10.1201/9781003019169

    work page doi:10.1201/9781003019169 2021

  48. [48]

    OptunaHub: A platform for black- box optimization.arXiv preprint arXiv:2510.02798, 2025

    Yoshihiko Ozaki, Shuhei Watanabe, and Toshihiko Yanase. OptunaHub: A platform for black- box optimization.arXiv preprint arXiv:2510.02798, 2025. 20

    Pith/arXiv arXiv 2025

  49. [49]

    Riley, Jayne F

    Richard D. Riley, Jayne F. Tierney, and Lesley A. Stewart, editors.Individual Par- ticipant Data Meta-Analysis: A Handbook for Healthcare Research. Wiley, 2021. doi: 10.1002/9781119333784

    work page doi:10.1002/9781119333784 2021

  50. [50]

    Chapman and Hall/CRC, 1st edition, 2019

    Max Kuhn and Kjell Johnson.Feature Engineering and Selection: A Practical Approach for Predictive Models. Chapman and Hall/CRC, 1st edition, 2019. doi: 10.1201/9781315108230

    work page doi:10.1201/9781315108230 2019

  51. [53]

    Cambridge University Press, 2006

    Andrew Gelman and Jennifer Hill.Data Analysis Using Regression and Multilevel/Hierarchical Models. Cambridge University Press, 2006. 21

    2006

  52. [2021]

    URLhttps://www.pnas.org/doi/abs/10.1073/pnas

    doi: 10.1073/pnas.2016191118. URLhttps://www.pnas.org/doi/abs/10.1073/pnas. 2016191118

    doi:10.1073/pnas.2016191118

  53. [2024]

    doi: 10.1038/s41597-023-02787-8

    doi:10.1038/s41597-023-02787-8