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

AI summary

Overview

Research area: Machine learning for brain–computer interfaces, specifically cross-dataset transfer for motor-imagery EEG decoding using physically motivated observation-operator modeling. Technical level: Advanced — the paper combines Riemannian/linear-algebraic gauge arguments, a lead-field transport operator, a bounded state-space model, and a multi-dataset statistical protocol with hierarchical bootstrap and exact same-parent controls. Scope (one sentence): The paper asks whether an EEG decoder can be fitted using only source datasets — never touching target signals, labels, or statistics, only the target electrode coordinates — and still decode a completely unseen native montage, and it answers with a gauge-removal plus template motor lead-field transport method evaluated across five public datasets.

What This Paper Is About

Public motor-imagery EEG datasets differ in how many electrodes they use, where those electrodes sit, how they are referenced, and which channels fail — so a decoder trained on one dataset usually cannot be applied directly to another. Most transfer methods treat this as a statistical "representation shift," but the authors argue that changing the montage and reference also changes the observation operator: the mathematical map from brain sources to recorded sensors. GRAFT-EEG's goal is to transport each dataset's own electrode mask through a physically grounded template motor lead field, under a strict contract in which the target dataset's signals, labels, and empirical statistics stay sealed during all fitting.

Key Contributions

1. A precise access contract for cross-dataset EEG. The paper formalizes "target-statistics-blind" transfer: source signals and labels may be used for fitting, target electrode coordinates may be used to instantiate a fixed linear operator, but target signals, labels, covariance, normalization statistics, calibration, validation loss, and stopping signals are all forbidden. Everything is committed to a hash-bound manifest before a separate process opens the target data.
2. Gauge removal plus physical motor-field transport. An active-set common-reference quotient removes the unidentifiable common-offset component on each trial's retained channels, and a mask-specific ridge-regularized inverse transports the remaining field into a fixed eight-dimensional bilateral motor basis (four modes per hemisphere) built once from declared montage geometry and an fsaverage motor region.
3. Restricted theory. A restricted-identifiability proposition shows that heterogeneous montages whose joint observability matrix has rank d uniquely determine the latent state up to an invertible similarity transform, and an operator-perturbation bound separates measurement noise from montage/operator mismatch in the reconstruction error.
4. A layered evaluation with exact same-parent operator controls. Five leave-one-dataset-out targets, strict versus partial provenance denominators, a dataset–subject hierarchical bootstrap, four post-hoc controls that swap only the lead field (coordinate basis, matched-spectrum random, row-shuffled, wrong-anatomy), and a separately registered 60-participant same-paradigm check.

Main Findings

• Dynamic GRAFT-EEG beat the strongest clean pretrained baseline. Equal-target balanced accuracy was 60.865% for the dynamic model versus 52.684% for BIOT, a difference of +8.182 points (hierarchical 95% CI +6.382 to +10.196), positive in every one of the five targets (individual target effects +6.506 to +10.437 points; leave-one-target-out estimates 7.618–8.600 points).
• The static anchor alone was already strong. Static GRAFT-EEG reached 59.806% equal-target balanced accuracy, so most of the gain comes from cross-montage harmonization rather than the temporal model.
• The physical lead field mattered, not just the dimensionality. In four post-hoc exact same-parent tests that fixed split, seed, head, features, ridge rule, and budget, the full model exceeded the coordinate basis by 9.959 points, a matched-spectrum random operator by 9.698, a row-shuffled operator by 11.241, and a wrong-anatomy visual-cortex lead field by 3.985 — all with 95% CIs excluding zero and Holm-adjusted P < 0.001. The authors explicitly label this ladder post-hoc, not independently confirmatory.
• The state-space residual added little and was target-dependent. Dynamic minus static was +1.059 points (hierarchical 95% CI +0.013 to +1.892); four targets were positive while BNCI2014_001 was −0.586 points (−2.423 to +1.497).
• Against EEGPT on its four clean targets, the descriptive means were 61.520% versus 59.057% (+2.463 points), with per-target effects of −1.328, +8.857, −0.667, and +2.989. The authors keep this out of the strict 5/5 ranking.
• The separate 60-participant Dreyer2023A evaluation partially confirmed the mechanism. Physics versus no-physics was +4.029 points (95% CI +2.338 to +5.773), while dynamic versus static was +0.57 points (−0.08 to +1.23) and therefore inconclusive; the preregistered status is partial confirmation, not unrestricted external validation.
• Withholding target geometry entirely cost little. In the strict source-only-geometry sensitivity, where all 25 fits were sealed before the target montage was opened, dynamic accuracy was 60.460% (only 0.406 points below the canonical result) and still exceeded static by +0.747 points.
• Cho2017 was the hardest target for both methods (dynamic GRAFT 58.172%, BIOT 50.086%).
• Small label budgets did not close the gap dramatically. Revealing 1%, 5%, and 10% of target labels under the registered few-shot procedure gave 61.323%, 61.477%, and 61.722%, versus 60.865% in the zero-target protocol — increments the authors frame as context, not an oracle ceiling.
• Coordinate perturbations tracked performance loss. Across 50 target–condition units, the change in the reduced operator tracked the change in dynamic balanced accuracy (Spearman ρ = −0.449, 95% CI −0.760 to −0.282, permutation P < 0.001); at 20 mm, equal-target accuracy changed by −1.064 points for tangent errors and −2.220 points for rigid errors.
• Baseline health was audited, not assumed. BIOT was nondegenerate but had marginal source validation (mean balanced accuracy 53.6%), and several classical baselines (CSP/FBCSP/Riemann) produced systematic single-class target predictions; the authors therefore decline to use collapsed baselines as mechanistic evidence.

Methodology in Plain English

The core idea is that a shared reference electrode adds the same constant voltage to every channel, and that constant is physically meaningless — it is a "gauge" choice. So for each trial, the method first subtracts the average across that trial's surviving channels, and applies exactly the same subtraction to the corresponding rows of a precomputed lead-field matrix. This matters because bad-channel patterns differ per trial: Cho2017 alone produces 223 distinct retained active-channel masks, and 4,710 of its trials have at least one inactive native channel, so a single fixed pseudo-inverse per dataset would be wrong.

Next, the method builds a fixed eight-dimensional "motor basis" from the declared electrode coordinates and a template motor region on the fsaverage head. It combines this basis with the denoised electrode mask and inverts it with a ridge penalty scaled by the operator's largest singular value, so the regularization strength is comparable across different masks. Each trial is carried into this shared eight-dimensional space, where different montages become comparable. Nothing in this construction touches EEG samples, labels, or target statistics.

On top of that latent representation, a static classifier takes log power in 8–13 Hz, 13–30 Hz, and 8–30 Hz bands (24 features) and fits a class-balanced weighted logistic regression. A small secondary branch chops each trial into 0.5-s windows stepped by 0.25 s (11 windows of 24 features) and runs a 16-dimensional diagonal state-space recurrence whose decay gates are constrained between 0.5 and 0.999, producing a residual capped below 2 by a tanh. That residual is initialized to zero — so the model starts exactly at the static classifier and can only improve on it — and is added to the frozen static logit before the sigmoid.

Evaluation holds out one complete dataset at a time across five public collections (253 participants, 41,342 retained left/right-hand trials, native montages of 22–64 channels, signals resampled to 200 Hz and epoched to 600 samples over 0.5–3.5 s). Each fold uses five seeds, subject-disjoint source validation, and a dataset–subject hierarchical bootstrap with 20,000 replicates for uncertainty.

Why This Matters

Research impact. The paper reframes cross-dataset EEG transfer as a problem of the observation operator rather than only of feature distributions, and it attaches a testable access contract to that claim. The four same-parent controls are the distinctive methodological move: by holding everything fixed and swapping only the physical lead field, they rule out explanations based on latent dimensionality, singular spectrum, or electrode-wise response magnitude. The result pushes back on the assumption that ever-larger pretrained EEG models automatically dominate compact, physically structured decoders — several strong baselines in this study sat near chance.

Real-world applications:

• Calibration-free BCI spellers and prosthetics, where a new user's cap layout and reference differ from the training cohort and no labeled calibration session is available.
• Multi-site clinical or research studies that pool recordings from different amplifiers and electrode caps without forcing every site onto a shared channel subset.
• Wearable and low-channel EEG devices with ad hoc or reduced sensor layouts, where retaining native sensors instead of intersecting channel sets preserves information.
• Retrospective analysis of archived EEG corpora where only the montage specification, not the raw data, can be shared across institutions for model development.

Industry relevance. The zero-target-label protocol is directly relevant to organizations that must ship a decoder before they can collect data from a new customer population, and the hash-bound manifest with process-separated fitting, prediction, and scoring is a template for auditable regulated machine-learning pipelines. The paper is also candid that its evidence is bounded: five binary motor-imagery domains, an independent unit of only n = 5 held-out datasets, and dataset confounded with device and protocol.

Future Directions

1. Generalize beyond the binary hand motor-imagery task and beyond five datasets. The authors state that non-motor paradigms were not tested and that the independent generalization unit is the held-out dataset, so broader paradigm coverage and more independent sites are the obvious next step.
2. Replace the fixed template basis with participant-specific or anatomy-adaptive operators. The fsaverage basis is explicitly "a common coordinate system, not participant-specific localization," and participant-specific source recovery is declared beyond scope; closing that gap is a natural extension.
3. Characterize placement tolerance more fully. Coordinate sensitivity was characterized only for two synthetic error families over the five declared montages, so arbitrary cap deformation, real anatomical variation, and a clinical placement threshold remain open.
4. Make the target-statistics-blind contract geometry-blind too. The held-out montage must still be opened after sealing to instantiate the prediction operator, and the source-only-geometry sensitivity is presented as a sensitivity analysis rather than a replacement for the primary result — a genuinely geometry-free variant remains an open problem.

Target Audience

Researchers and graduate students in EEG decoding, brain–computer interfaces, and domain adaptation who want a rigorous example of how to separate what information a model is allowed to see from how well it performs. It will be most useful to readers comfortable with linear algebra, inverse problems, and state-space models who are designing cross-dataset benchmarks or auditing transfer claims. Practitioners building calibration-free BCI systems will find the access contract and the same-parent control ladder directly reusable, and reviewers or methodologists interested in provenance-aware evaluation will find the strict-versus-partial denominator bookkeeping instructive.

Authors’ 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. method 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 method 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&lt;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.