Probing Spatial and Spectral Grounding in EEG Foundation Models

Kleanthis Avramidis; Ivan Tashev; Dimitra Emmanouilidou

cs.NEOct 21, 2026

Probing Spatial and Spectral Grounding in EEG Foundation Models

Authors: Kleanthis Avramidis, Ivan Tashev, Dimitra Emmanouilidou

Organizations: Signal Analysis and Interpretation Lab, University of Southern California, USA · Microsoft Research, Redmond, WA, USA

Abstract

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.

Figures & tables

Figure 1: SPA1 — Sensitivity to spatial encoding. Normalized latent displacement after permuting electrode labels while holding the waveforms fixed ( $S_{\mathrm{id}}$ ; left) and after breaking the correspondence between signals and their location ( $S_{\mathrm{sig}}$ ; right).

Figure 2: SPA2 — Topographic structure across depth. Spearman correlation between off-channel influence and inter-electrode distance. Negative values indicate stronger influence on nearby than distant channels. Results are shown on PhysioNet-MI.

Model	Clean	Permuted	$\Delta$	$\lVert\delta f\rVert/\lVert f\rVert$
REVE	0.455 ±0.006	0.340 ±0.012	+0.115 ±0.013	0.714
LaBraM	0.354 ±0.004	0.343 ±0.011	+0.012 ±0.010	0.805
BIOT	0.276 ±0.012	0.275 ±0.004	+0.001 ±0.009	0.589
CSBrain	0.440 ±0.008	0.337 ±0.004	+0.104 ±0.008	0.885

Table 1: SPA3 — Balanced accuracy before and after permuting the learned positional code at test time on PhysioNet-MI, with the input waveform unchanged and the readout frozen. The last column reports the induced latent change. Random chance is 25%.

Figure 3: SPE1 — Information beyond spectral power. Balanced accuracy of linear probes distinguishing clean recordings from IAAFT [ 17 ] surrogates on PhysioNet-MI and Siena. Error bars indicate SD from participant-clustered bootstrapping at 200 resamples.

Figure 4: SPE2 — Sensitivity to narrowband oscillations. Four-way accuracy for identifying the injected delta, theta, alpha, or beta band from the unit-normalized latent displacement, shown against injection power relative to the original band power.

Model	Ictal rhythmic	Spike wave	Pink noise	Specificity
REVE	+0.113 ±0.027	+0.097 ±0.029	-0.053 ±0.011	+0.158 ±0.018
LaBraM	+0.131 ±0.027	+0.055 ±0.015	+0.115 ±0.025	-0.022 ±0.039
BIOT	+0.069 ±0.059	+0.053 ±0.045	-0.014 ±0.032	+0.076 ±0.071
CSBrain	+0.114 ±0.035	+0.053 ±0.021	+0.044 ±0.005	+0.040 ±0.013
CBraMod	+0.036 ±0.008	+0.080 ±0.014	+0.040 ±0.008	+0.019 ±0.006
log-PSD	+0.117 ±0.023	+0.103 ±0.020	+0.083 ±0.023	+0.027 ±0.025

Table 2: SPE3 — Change in

P(\mathrm{seizure})

after a biomarker is injected into held-out non-seizure segments of Siena Scalp at

0

dB, with the readout fitted on clean folds held frozen. Specificity is the mean epileptiform response minus the noise control response. Values are participant-level means

\pm

standard error of the mean (SEM).

EEG Foundation Models

Electroencephalography (EEG) foundation models learn reusable representations of brain signals for transfer across people, tasks, and recording setups.

See all 101 papers

Beyond Accuracy: Robustness, Interpretability and Expressiveness of EEG Foundation Models
May 17, 2026Urban Širca, Maryam Alimardani, Stefanos Zafeiriou +1EEG Decoding Neural Network Interpretability
Vrije Universiteit Amsterdam · Imperial College London
What Do EEG Foundation Models Capture from Human Brain Signals?
May 12, 2026Ling Tang, Qian Chen, Jilin Mei +6Frequency-Domain Feature Learning Representation Probing
Shanghai Artificial Intelligence Laboratory · Shanghai Jiao Tong University · Tongji University +2
Representational and Functional Robustness to Electrode Montages in EEG Foundation Models
Sep 28, 2026Jakob Steglich, Justus Meyer zu Bexten, Shakiba Moradi +3EEG Foundation Models Electroencephalography
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 +2

May 17, 2026J/K move · Enter open · S save

Beyond Accuracy: Robustness, Interpretability and Expressiveness of EEG Foundation Models

Urban Širca, Maryam Alimardani, Stefanos Zafeiriou, Konstantinos Barmpas

Vrije Universiteit Amsterdam · Imperial College London

EEG Decoding Neural Network Interpretability Neural Network Expressivity EEG Foundation Models

Open paper AskPDF arXiv

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.

Abstract

EEG Foundation Models

Explore similar work

Beyond Accuracy: Robustness, Interpretability and Expressiveness of EEG Foundation Models

What Do EEG Foundation Models Capture from Human Brain Signals?

Representational and Functional Robustness to Electrode Montages in EEG Foundation Models