CortexBridge: Cortical Alignment of EEG Montages for Foundation Models
EEG foundation models are typically pretrained on fixed channel layouts, so they struggle when electrode montages differ at test time. This paper proposes CortexBridge, a lightweight adapter that routes arbitrary scalp montages into a fixed cortical atlas space using electrode coordinates and sparse attention. By injecting this anatomically grounded representation into frozen backbones, the authors show consistent accuracy gains across three models, five datasets, and three BCI paradigms.
CortexBridge offers a clean and anatomically motivated solution to montage-agnostic transfer: instead of forcing heterogeneous EEG channels into a fixed scalp vocabulary, it learns a shared cortical latent space via a lightweight atlas-based adapter. The SSVEP result is genuinely strong, but the average improvements for EEGPT and LaBraM are modest (under 1%), the paper lacks ablations isolating the cortical mechanism from simpler geometric baselines, and the absence of code or statistical testing limits confidence in the reported margins.
The strongest aspect is the anatomical grounding of the interface. Using Schaefer atlas ROIs and Yeo networks provides a principled alternative to scalp-space channel vocabularies, the residual gate initialization ensures a fair start from the raw baseline, and the 13.02% balanced accuracy gain on the 12-class SSVEP dataset is a clear demonstration of utility for very sparse montages. The qualitative finding that SSVEP activations concentrate in the Yeo Visual network adds useful neuroscientific face validity.
The primary concern is that the paper never demonstrates the cortical step is necessary. Without ablations removing the atlas, the geometric bias, or the sparse routing, the gains could simply come from adding approximately 0.10M extra parameters and an inductive bias on scalp distances. The improvements for EEGPT and LaBraM average under 1%, which is small relative to the likely variance of subject-grouped three-fold cross-validation on datasets with as few as eight subjects. Two conditions actually degrade, including a 0.96% drop for EEGPT on BNCI2014-008, yet no statistical test is reported to show any gain is significant.
The evidence supports feasibility but not the specific cortical claim. Table 1 compares only the raw foundation-model input against the same input plus CortexBridge, omitting comparisons to coordinate-driven scalp encodings, channel interpolation, or a coordinate-only MLP. The anatomical specificity is supported only by Figure 3, a qualitative visualization on two datasets using one backbone, which is suggestive but not conclusive evidence that the atlas geometry drives performance rather than acting as a generic spatial regularizer.
Reproducibility is partially addressed but incomplete. Architecture hyperparameters and the optimization recipe are clearly listed in Appendix B. However, no code repository, subject-fold splits, or preprocessing scripts are released, and the paper omits how electrode coordinates are obtained for each MOABB dataset — a critical detail because few public BCI releases include true 3D sensor locations. The column-balance penalty weight is given, but there is no description of the foundation-model patch-specific resampling pipeline beyond a passing mention, which would block independent reproduction.
Electroencephalography (EEG) foundation models are often pretrained with a fixed channel vocabulary or a limited set of montages, making transfer difficult when electrode layouts change. We propose CortexBridge, a lightweight adapter that combines EEG features with electrode and atlas coordinates to map arbitrary montages into a shared cortical latent space. Evaluated with three frozen foundation models on five brain-computer interface (BCI) datasets from the Mother of All BCI Benchmarks (MOABB), CortexBridge improves performance in 13 of 15 evaluations. The gains in balanced accuracy average 0.80% for EEGPT, 0.70% for LaBraM, and 3.26% for CBraMod, with a maximum gain of 13.02% on 12-class steady-state visual evoked potential (SSVEP) classification. Visualizations of the learned atlas representations reveal task-dependent spatial patterns, with SSVEP showing a more concentrated representation in the Yeo Visual network than auditory P300. These results establish cortical alignment as a learnable and anatomically grounded routing mechanism from heterogeneous EEG montages to pretrained foundation models.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.