Representational and Functional Robustness to Electrode Montages in EEG Foundation Models
Authors: Jakob Steglich, Justus Meyer zu Bexten, Shakiba Moradi, Laure Ciernik, Simon M. Hofmann, Mina Jamshidi Idaji
Organizations: Max Planck Institute for Human Cognitive and Brain Sciences, Leipzig, Germany · ScaDS.AI Dresden/Leipzig · BIFOLD–Berlin Institute for the Foundations of Learning and Data, Berlin, Germany · Machine Learning Group, Technische Universität Berlin, Berlin, Germany · Hector Fellow Academy, Karlsruhe, Germany
EEG foundation models (EEG-FMs) are intended to generalize across different datasets by learning representations that, ideally, are invariant to dataset-specific EEG configurations such as electrode montages. However, EEG-FMs that accept different montages as input do not guarantee that representations and predictions remain stable across different electrode configurations, especially outside the training setting. In this work, we investigate the effects of different electrode montages through a joint functional and representational analysis of four EEG foundation models selected to span distinct montage-handling designs. We evaluate embeddings on cross-subject resting-state eyes-open/closed and within-subject motor-imagery classification under spatially informed channel reduction. Functional robustness is tested through the generalizability of linear probes across channel counts, while representational robustness is assessed through within-subject similarity and preservation of between-subject geometry. The four models show distinct robustness profiles, and the two axes dissociate: large changes in embedding similarity need not come with comparable probe degradation, and stable embeddings can still lose downstream performance. Comparing two readouts of the same encoder further shows that aggregation, not the encoder alone, determines functional robustness: pooling into anatomically aligned regions degrades less than a learned global readout, despite being montage-invariant by construction. Montage robustness is therefore a joint property of the encoder and its aggregation, and characterizing it requires both a representational and a functional axis. Input compatibility alone is evidence for neither.
Figures & tables
Figure 1: Functional within- and cross-montage robustness. (a) Within-montage baselines (N2N-probes with concat-all as pooling) for inter-subject RS and intra-subject MI classification; bars show mean AUROC and standard deviation across seeds (n=10) for RS and subjects for MI ( Table S3 ). (b, c) N2M-probes on RS and MI, respectively, with matched train and test samples, and (d, e) with held-out test data. Rows give the training and columns the test montage. The full montage has 128 channels for RS and 62 for MI ( Appendix C ).
Figure 2: Representational cross-montage robustness for resting-state EEG. Each point compares the embedding of the full montage ( Cfull ) to that of a subsampled montage with Csub channels, displayed on the x-axis. For within-subject similarities, each point is an average over subjects; variance is displayed over the first three EEG epochs per label (six trials per subject). (a) Within-subject cosine similarity. (b) Centered Kernel Alignment (CKA) between the subject-to-feature matrices of the two montages.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
SingLEM
CBraMod
LUNA
REVE
Category
Channel-independent
Spatial-temporal channel interaction
Latent channel unification
Continuous coordinate-aware
Pretraining corpus
71 datasets
1 (TUEG)
2 (TUEG + Siena)
92 datasets
Cross-channel interaction
Shared weights across channels; no interaction
Separate spatial and temporal self-attention
Learned query cross-attention
Cross-channel self-attention
Spatial encoding
None
Spatio-temporal convolution
3D electrode coordinates
4D Fourier encoding of electrode and temporal coordinates
Channel/layout handling
Channel dimension retained
Channel dimension retained
Fixed-size latent representation
Coordinate-aware channel-temporal representation
Pretraining channel configurations
17–256
19
20, 22 (TUEG) 29 (Siena)
3–129
Appendix
Table S1: Overview of the evaluated EEG foundation models.
Model
Channel-Fusion Mechanism
Positional / Spatial Encoding
Output Shape
CBraMod
Criss-cross transformer: separate spatial and temporal self-attention branches
Asymmetric Conditional Positional Encoding (ACPE)
C×P×F
LUNA
Cross-attention: Q≪C learned latent queries attend to channel-wise patch features, collapsing the channel axis
3D electrode coordinates via NeRF-style sinusoidal projection
(Q⋅P)×F
SingLEM
None during encoding. Channels processed fully independently (shared weights); cross-channel combination deferred to late-fusion concatenation downstream
(channel-independent)
C×P×F
REVE
Full self-attention over the flattened (C×P) token sequence + one summary token (cf. LUNA)
4D Fourier features + learned linear projection (3D coords + temporal patch index)
C×P×F
Appendix
Table S2: Overview of channel-fusion mechanisms across evaluated EEG foundation models. All four models are pretrained with a masked-reconstruction objective; C , P , F , Q follow the notation introduced in subsection B.2 .
Figure S1: Regions of Interest (ROIs) mapping for 128-channel GSN-HydroCel and 62-channel Standard-1005 montages. These ROIs were used in experiments 1 and 2. Channels were assigned to ROIs manually based on their anatomical location.
Figure S2: Sampled montages for the 128-channel and 62-channel montages. . These sampled channels were used in Experiment 2.
Resting-State task
Motor Imagery task
Model
32-ch
64-ch
128-ch
16-ch
32-ch
62-ch
CBraMod
0.91 ± 0.02
0.90 ± 0.01
0.91 ± 0.01
0.63 ± 0.12
0.67 ± 0.13
0.71 ± 0.14
LUNA
0.90 ± 0.02
0.89 ± 0.02
0.91 ± 0.02
0.60 ± 0.12
0.61 ± 0.12
0.62 ± 0.11
REVE
0.94 ± 0.01
0.92 ± 0.02
0.93 ± 0.01
0.69 ± 0.13
0.69 ± 0.13
0.69 ± 0.13
SingLEM
0.89 ± 0.03
0.88 ± 0.02
0.91 ± 0.02
0.62 ± 0.12
0.65 ± 0.12
0.67 ± 0.13
Appendix
Table S3: AUROCs of within-montage baselines (N2N-probes with concat-all pooling) in Experiment 1 for RS and MI data . For the inter-subject RS task, the mean ± standard deviation of AUROC values across multiple seeds (n=10) is reported. For the intra-subject MI task, the values indicate the mean and standard deviation of AUROC values across subjects. Therefore, the high variance for MI task is associated with the high variance of subject-specific performances. This variation is in line with the CSP classification results reported in Lee et al. [2019] .
Method
concat-all
ROI-concat-all
ch-mean
attention-pool
MI task – 62-channel
CBraMod
0.71 ± 0.14
0.70 ± 0.13
0.54 ± 0.08
-
LUNA
0.62 ± 0.11
-
-
-
REVE
0.69 ± 0.13
0.68 ± 0.13
0.56 ± 0.09
0.58 ± 0.09
SingLEM
0.67 ± 0.13
0.67 ± 0.13
0.51 ± 0.05
-
RS task – 128-channel
Appendix
Table S4: AUROCs of within-montage baselines (N2N-probes) in Experiment 1 by pooling method for RS and MI datasets . Here, results for the full montages (128-channel for RS, and 62-channel for MI) are reported. For the inter-subject RS task, the mean ± standard deviation of AUROC values across multiple seeds (n=10) is noted. For the intra-subject MI task, the numbers indicate the mean and standard deviation of AUROC values across subjects. Therefore, the high variance for MI task is associated with the high variance of subject-specific performances. Additionally to the pooling methods introduced above, we report results for ch-mean pooling, where the embeddings are averaged over the channel dimension, and concatenate patches and features, yielding 1D embeddings of size (C⋅P⋅F) . This method serves as baseline, since it removes most information across channels, while providing montage-agnostic embeddings.
Figure S3: Representational cross-montage robustness for motor imagery EEG. Each point compares the embedding of the full montage ( Cfull ) to that of a subsampled montage with Csub channels, displayed on the x-axis. For within-subject similarities, each point is an average over subjects; variance is displayed over the first three EEG trials per label (six trials per subject). (a) Within-subject cosine similarity. (b) Centered Kernel Alignment (CKA) between the subject-to-feature matrices of the two montages.
Figure S4: Average cosine similarity across ROIs and Wasserstein Distances. Each point compares the embedding of the full montage ( Cfull ) to that of a subsampled montage with Csub . Each point is an average over subjects, variance is displayed over the first three EEG trials per label (six trials per subject). (a, c) ROI-level within-subject cosine similarity: the ROI-pooled embeddings were first averaged over the patch dimension to produce ∈RR×F embeddings. Then we computed ⟨s(fr(1),fr(2))⟩r , where fr(i)∈RF is the feature vector of ROI r for montage i and s(.) is the cosine similarity. (b, d) Within-subject Wasserstein distances W2 computed per subject between the two (ROI⋅P)×F embedding matrices, treating each of the N=ROI⋅P rows as a sample in RF , except for LUNA, which has N=Q rows in RF .
Figure S5: Representational cross-montage robustness for resting-state and motor-imagery EEG with an alternative channel sampling seed. Each point compares the embedding of the full montage ( Cfull ) to that of a subsampled montage with Csub channels, displayed on the x-axis. For within-subject similarities, each point is an average over subjects; variance is displayed over the first three EEG trials per label (six trials per subject). (a, c) Within-subject cosine similarity. (b, d) Centered Kernel Alignment (CKA) between the subject-to-feature matrices of the two montages.
Figure S6: Average cosine similarity across ROIs and Wasserstein Distances with an alternative channel sampling seed. Each point compares the embedding of the full montage ( Cfull ) to that of a subsampled montage with Csub . Each point is an average over subjects, variance is displayed over the first three EEG trials per label (six trials per subject). (a, c) ROI-level within-subject cosine similarity: the ROI-pooled embeddings were first averaged over the patch dimension to produce ∈RR×F embeddings. Then we computed ⟨s(fr(1),fr(2))⟩r , where fr(i)∈RF is the feature vector of ROI r for montage i and s(.) is the cosine similarity. (b, d) Within-subject Wasserstein distances W2 computed per subject between the two (R⋅P)×F embedding matrices, treating each of the N=R⋅P rows as a sample in RF , except for LUNA, which has N=Q rows in RF .
Figure S7: Functional within- and cross-montage robustness for 20 ROIs. (a) Within-montage baselines ( N=M ) for inter-subject RS classification; bars show mean AUROC and standard deviation across seeds for RS ( Table S3 ). (b) N2M-probes on RS, with matched train and test samples, and (c) with held-out test data. Rows give the training and columns the test montage. The full montage has 128 channels for RS ( Appendix C ).
Electroencephalography (EEG) foundation models learn reusable representations of brain signals for transfer across people, tasks, and recording setups.
EEG foundation models (EEG-FMs) have been evaluated predominantly on clean, in-distribution accuracy, demonstrating modest gains over supervised baselines and weak frozen representations. This study examines whether these conclusions hold beyond clean accuracy by evaluating six EEG-FMs and a supervised baseline across ten datasets along three layers of analysis: (i) Robustness: we apply test-time perturbations including additive noise, random and region-based channel dropout and region-specific noise injection. Our analyses show that no single model dominates all failure modes. The most noise-robust model is among the most fragile under channel dropout and much of the dropout fragility disappears when channels are removed rather than zero-padded. (ii) Interpretability: using attribution methods in EEG-FMs, we show that models broadly concentrate relevance on task-appropriate brain regions consistent with known neurophysiology. (iii) Expressiveness: we demonstrate that the poor head-only performance previously attributed to low-quality pre-trained representations is largely explained by the pooling strategy and that EEG-FMs possess sufficient representational capacity when their token-level embeddings are preserved. Furthermore, with block-wise probing and attention analysis we show that late blocks are repurposed during fine-tuning, while early blocks already hold task-related information. Our results show that conclusions about EEG-FMs depend on evaluation choices and we recommend that future evaluation of EEG-FMs should report robustness per perturbation type, produce attribution maps and examine multiple pooling strategies.
EEG foundation models are advancing at an unprecedented pace, yet their evaluation is driven by downstream accuracy, which reveals little about whether their predictions rely on physiologically interpretable signal structure. We introduce a set of perturbation probes in two grounding axes: spatial grounding in channel identity and topographic locality, and spectral grounding in oscillatory content and waveform morphology. We apply these probes to five published EEG foundation models across motor-imagery and seizure recordings. Our experiments reveal differences in both representational sensitivity and downstream reliance. Electrode identity may be encoded without influencing downstream classification: permuting LaBraM's positional codes heavily displaced its embeddings yet reduced motor-imagery accuracy by only 1.2 points, while spectral details may be discriminable without producing specific responses to simulated clinical markers, as only REVE exceeded a spectral baseline on this task. These dissociations are hidden in common benchmarks, demonstrating how grounding probes can complement performance-based evaluation by revealing what EEG foundation models encode and what their predictions depend on.
Kleanthis Avramidis, Ivan Tashev, Dimitra Emmanouilidou
Signal Analysis and Interpretation Lab, University of Southern California, USA · Microsoft Research, Redmond, WA, USA
Scaling EEG foundation models requires pooling data across heterogeneous electrode montages, a prerequisite both for larger pretraining corpora and for downstream deployment. We present the first systematic comparison of four channel adaptation methods (Conv1d projection, spherical spline interpolation (SSI), source-space decomposition, and Riemannian re-centering) across five pretrained EEG foundation models (5M--157M parameters), five downstream tasks, and two training regimes with 10--15 random seeds each. We find that rigid-montage models (BENDR, Neuro-GPT) require external adaptation, while flexible models (EEGPT, CBraMod) match or exceed it natively when fine-tuned but benefit from external methods under frozen-encoder deployment. A probe-SFT asymmetry exists: external adaptation can cause severe negative transfer during fine-tuning of flexible models. The optimal method is architecture-dependent (Conv1d for BENDR, SSI/Riemannian for Neuro-GPT, source-space decomposition for depression detection), and 5M-parameter CBraMod outperforms models up to 31× larger on 4/5 datasets, consistent with independent findings that compact EEG-specific architectures can match larger models.
Kuntal Kokate, Bruno Aristimunha, Dung Truong +1
Swartz Center for Computational Neuroscience, Institute for Neural Computation, University of California San Diego, La Jolla, CA 92093, USA · Yneuro, Paris, France · Centre de Recherche Cerveau et Cognition, CNRS, Université Paul Sabatier, Toulouse, France