acceptodds
Withdrawn from ICLR 2027

GRAFT-EEG: Gauge-Aware Physical Field Transport for Cross-Dataset Motor-Imagery EEG

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Cross-dataset EEG decoding often treats montage, reference, and active-sensor changes as representation shift, although they also alter the observation operator. We study target-statistics-blind transfer: fitting receives no target signal, label, or empirical statistic, and electrode coordinates only instantiate a fixed prediction operator. GRAFT-EEG removes the active-set common-reference gauge and transports each native mask through a template motor leadfield. A source-fitted spectral classifier anchors the decision; a bounded state-space-model (SSM) residual is secondary. We establish restricted identifiability and an operator-perturbation bound. Across five leave-one-dataset-out targets (253 participants), dynamic GRAFT-EEG achieved 60.865% equal-target balanced accuracy versus 52.684% for strict-clean 5/5 BIOT (hierarchical difference +8.182 points; 95% CI +6.382 to +10.196; positive in every target). On four EEGPT-clean targets, the descriptive means were 61.520% and 59.057%. In four post-hoc exact same-parent tests, FULL exceeded coordinate-basis, matched-spectrum, row-shuffled, and wrong-anatomy controls by 9.959, 9.698, 11.241, and 3.985 points, respectively (all 95% CIs excluded zero; Holm-adjusted P < 0.001). A separate 60-participant same-paradigm evaluation retained a positive physical-operator contrast (+4.029 points; +2.338 to +5.773), whereas dynamic-static was inconclusive; its preregistered status is partial confirmation rather than unrestricted external validation. The SSM added +1.059 points (+0.013 to +1.892). Claims remain bounded to five binary motor-imagery domains and the tested operators.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG decoding, motor imagery, cross-dataset transfer, domain generalization, heterogeneous montages, physical observation operators

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_snXqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes GRAFT-EEG for cross-dataset MI EEG decoding under heterogeneous electrode configurations. The method combines active-set reference projection, physics-informed transport into an eight-dimensional motor-cortex basis, a static spectral classifier, and a bounded SSM residual. Evaluations across five datasets report 60.87% balanced accuracy, outperforming BIOT under the target-statistics-blind protocol. Additional operator controls and external evaluations investigate the contribution of physical modeling.

Critical strengths

S1. Important and practically relevant research problem

The paper addresses cross-dataset EEG decoding under heterogeneous electrode montages, reference configurations, and missing channels without requiring target-domain calibration. This is relevant to improving the generalizability and reducing the calibration requirements of BCI systems.

S2. Extensive experiments with targeted mechanism validation

The evaluation covers five public datasets under a leave-one-dataset-out protocol. Four matched operator controls, FI/SSI comparisons, and an additional 60-participant evaluation provide useful empirical evidence for the contribution of task-matched spatial transport.

Critical weaknesses

W1. Limited methodological novelty relative to established physics-informed approaches

Physics-based montage interpolation is already supported by MNE-Python. Reference standardization, EEG inverse modeling, and source-space motor-imagery decoding also have established foundations. GRAFT-EEG appears largely to combine these techniques with task-specific feature engineering, with limited fundamental methodological novelty.

W2. Limited theoretical support for cross-dataset transfer

The two propositions rely on restrictive assumptions that do not directly correspond to the actual transfer setting. Proposition 1 considers joint observability of shared latent dynamics, whereas the datasets contain independent subjects and observations. Proposition 2 analyzes pseudoinverse reconstruction rather than the regularized inverse used in GRAFT-EEG. Neither establishes cross-montage representation consistency or generalization.

W3. Insufficient justification and validation of the eight-dimensional basis

The model fixes four SVD modes per hemisphere without analyzing alternative dimensions or truncation choices. Although the paper includes a source-only geometry sensitivity analysis and an additional external montage, its adaptability to more diverse, unseen electrode configurations remains insufficiently established.

W4. Insufficient isolation of physical modeling from task-specific anatomical priors

The reported improvements may arise primarily from the predefined bilateral motor-cortex structure rather than the physical leadfield itself. Existing operator controls do not adequately distinguish these contributions. Comparisons with simpler motor-region spatial representations would better clarify the necessity of the physical inverse model.

W5. Limited performance gains over the most relevant prior approach

Although GRAFT-EEG substantially outperforms several baselines, its improvement over FI + tangent-space classification is only 1.47%, with a 95% confidence interval of [-0.07, 2.96]. FI also performs better on two target datasets. Thus, the performance advantage over the closely related FI or Riemannian framework remains modest and statistically inconclusive.

W6. Unclear dataset characteristics and incomplete baseline reporting

Table 1 reports 4,710 retained Cho2017 trials containing inactive channels but only 28 excluded trials. Although the filtering criteria are described, the unusually high prevalence of inactive-channel patterns requires further explanation. In Table 2, CodeBrain results are entirely withheld despite completing all 25 evaluation cells because checkpoint provenance could not be established. The reporting could be more transparent.

W7. Poor readability and unnecessarily complicated presentation

The manuscript contains extensive defensive qualifications, formulaic language, and unconventional compound terminology, making the presentation unnecessarily difficult to follow. Several abbreviations, mathematical symbols, and technical figures are also insufficiently explained, obscuring the main methodological ideas and experimental findings.

Questions for authors

Q1. What motivates the fixed eight-dimensional motor basis? How sensitive are performance and cross-montage representation consistency to the number of retained SVD modes, the geometry of unseen montages and more MI labels?

Q2. How much of the improvement can be attributed specifically to the physical leadfield rather than the predefined motor-cortex prior? Can the authors further clarify the limited improvement over FI + tangent-space classification?

Reviewer_nBc9OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies zero-shot cross-dataset motor imagery decoding in terms of various conditions. It applied common-average as "reference" over good channels, followed by a very sophisticated and complicated gated procedure, to boost zero-shot performance.

Decoding an unseen montage without any target data is a relevant question for EEG. However, this paper would be greatly benefited from human writing, and scientific improvement, I am afraid (sorry it is very difficult to read).

Critical strengths
  • This is a very important and practical problem in MI decoding, or more general BCI, indeed.
Critical weaknesses

Clarity and Presentation

  • The writing would be benefited from significant revision in terms of "accesibility", there are many “jargons” appearing without context, even for people with EEG background. There are many good stuff overflowed by some very likely AI-written traces(!)! I would suggest the authors to significantly revise their writing, and manually, please.
  • A good "related work" section could be more organized, from macro and micro, and highlighting the research gap. Unfortunately, I am very lost up until Related Works, what this work actually
  • There are many terms that are very AI, for isntance, in Section 3, the heading, nobody will use the term "Contract" , I am afraid, unless your favourite AI. Human will just say experimental settings or something similar.

Methodology

  • The common-average referenncing, actually, is very close and alike to classical source-space decoding. For instance, Mellot et al. (2024) cited in the paper. The novelty over this line of work is therefore a bit limited.
  • Several design choices are fixed without ablation.
  • Proposition 1 relies on stakced multi-montage observations with known dynamics, while the model decodes each trail with a static ridge inverse. I am a bit kind of concerned about this setting.

Experiment and Results

  • Most baselines appear "broken" by the protocol rather than weak (!), what happened. Why these comparsion is so bad performing?
  • L307, Foundation-model baselines are frozen with a logistic probe, and coordinate-aware models such as REVE and LUNA, which address exactly this problem, were not evaluated.

Minor Issues and Suggestions

  • I don't think "observation coordiates" is a very commonly used term in EEG area, and also, to make the paper more friendly to lay people, please describe what it means at the very beginning, or just change it to electrode locations or so.
  • "GRAFT" is never expanded.
Questions for authors
  1. What is the exact common-channel intersection for the classical and compact baselines, and how were inputs normalised?

  2. Under the primary protocol used in this paper, does GRAFT-EEG significantly outperform field interpolation with tangent-space features?

  3. How does the method compare with coordinate-aware EEG foundation models such as REVE or LUNA when fine-tuned rather than frozen?

Reviewer_qTKhOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper addresses cross-dataset decoding of MI-EEG with heterogeneous montages. The setting allows only electrode coordinates, with no target signals, labels, or statistics. Graft-EEG removes the common-reference component, maps the signal into an 8-dimensional motor latent via an fsaverage leadfield inverse, and classifies band log-power with a logistic regression plus a bounded SSM residual. In leave-one-dataset-out evaluation on five MOABB datasets, it achieves 60.9% BA versus 52.7% for BIOT, and outperforms four post-hoc operator controls by 4.0–11.2 pp.

Critical strengths
  • The access contract is rigorously enforced, and negative results are reported transparently. However, this is a strength of the evaluation procedure and does not by itself support the method’s contribution.
Critical weaknesses
  • The comparison does not support the +8.2 pp gain.
    • Most baselines collapse to chance, and even BIOT reaches only 53.6% source BA.
    • In contrast, FI+tangent under the same contract essentially matches Graft static.
    • Coordinate-aware foundation models (REVE, LUNA etc) are excluded, and all label-free adaptation (re-centering, BN adaptation, etc.) is ruled out, so the method's practical position is unclear.
  • The contribution of the physical leadfield is not isolated. The collapsed controls all lack hemispheric structure, whereas the hemisphere-structured wrong-anatomy control reaches 56.9%. The current evidence therefore shows the effect of hemisphere-localized bases, not of a physical model.
Questions for authors

Clarifications

  • How were amplitude units harmonized across datasets?
  • Does the advantage hold against the following baselines?
    • Baselines run with their native preprocessing and per-trial normalization
    • A simple C3/C4 lateralization baseline
    • REVE and LUNA under the same protocol
  • Under the same-parent protocol, does FULL still outperform a non-physical, hemisphere-localized basis?

Future work

  • An auxiliary tier that allows label-free target adaptation for all methods would clarify whether the advantage persists once distribution shift is partially corrected.
  • Device-type perturbation experiments and replication on a dataset not used during development would strengthen the robustness and generality claims.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.