PaperlayerView exact revision on arXiv

Source-linked AI summary

Probing Spatial and Spectral Grounding in EEG Foundation Models

Kleanthis Avramidis, Ivan Tashev, Dimitra Emmanouilidou

arXiv:2610.29367v1cs.NE

TL;DR

Downstream accuracy alone does not reveal whether EEG foundation models rely on physiologically interpretable signal structure. The paper probes spatial and spectral grounding across five models, finding that representational sensitivity and downstream reliance can diverge. These dissociations suggest that benchmarks alone do not establish what models encode or what their predictions depend on.

  • Problem

    Downstream accuracy indicates how often an adapted EEG model is correct, but not which signal properties support its predictions.

    Source: §1 · p. 1

  • Method

    The study applies spatial and spectral perturbation probes to five frozen EEG foundation models, examining channel identity and locality, phase-dependent structure, oscillations, and clinical morphology.

    Source: §1 · p. 1, §2 · p. 2, §4.1 · p. 3

  • Results

    Across the probes, representational sensitivity did not reliably predict downstream reliance: spatial codes could alter representations without corresponding accuracy loss, and REVE alone exceeded the spectral baseline on clinical morphology specificity.

    Source: §5.1 · p. 4, §5.2 · p. 4, §5.3 · p. 4

  • Takeaways & Limitations

    Evaluation should distinguish whether models access and encode physiological structure from whether downstream decisions rely on it.

    Source: §1 · p. 1, §6 · p. 4

    The analysis is limited by the small number of Siena Scalp participants and by linear readouts on frozen encoders, which capture only linearly accessible information.

    Source: §6 · p. 4

Abstract

from arXiv · show

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.

1. INTRODUCTION

EEG foundation-model accuracy does not establish whether predictions rely on physiologically meaningful signal structure. The paper introduces spatial and spectral perturbation probes to distinguish what models can access, encode, and use.

1. INTRODUCTION

  • Performance alone does not show whether EEG models use spatial context or waveform characteristics that experts rely on when interpreting signals.Clinical interpretation may depend on spatial distribution and waveform morphology even when spectral properties overlap.Source: §1 · p. 1
  • EEG noise, device artifacts, and participant variability can permit shortcut-based performance, while learned representations may preserve information that prediction heads do not use.The paper separates architectural access, representational encoding, and downstream reliance as distinct questions.Source: §1 · p. 1
  • The probes intervene on electrode identity, channel order, positional encoding, phase-dependent structure, and oscillatory content across five EEG foundation models.The models are BIOT, LaBraM, CBraMod, CSBrain, and REVE.Source: §1 · pp. 1–2
  • The study introduces spatial probes of identity and distance-dependent cross-channel influence, plus spectral probes of phase-dependent structure and frequency-band activity calibrated against handcrafted baselines.It argues that access to information, its preservation in representations, and its use in predictions cannot be inferred from benchmarks alone.Source: §1 · p. 1
  • Prior evaluation work reports inconsistent gains over lighter decoders and leaves unclear whether EEG representations preserve interpretable electrode-location or waveform information beyond spectral power.The cited studies report gains around 1% on some benchmarks and no advantage when labels are scarce.Source: §1 · p. 1

2. DATA AND PRE-PROCESSING

The study probes two datasets with different montages, environments, and tasks, using frozen encoders and linear readouts. The subsets and preprocessing are specified for each corpus.

2. DATA AND PRE-PROCESSING

  • PhysioNet-MI is a four-class motor-imagery corpus with 109 participants and 64 channels; probing uses the 20 participants and 1,805 samples in the prior class-balanced test set to avoid leakage.Signals are average-referenced, filtered, resampled to 200 Hz, and epoched from cue onset to four seconds without baseline correction.Source: §2 · p. 2
  • Siena Scalp provides seizure versus non-seizure windows from 14 participants and 29 electrodes; probing uses 374 balanced 10-second windows from four held-out subjects.Signals are band-passed at 0.1–75 Hz, notch-filtered at 50 Hz, and resampled to 200 Hz.Source: §2 · p. 2
  • All five encoders remain frozen, and the protocol probes only linear readouts of pooled penultimate-layer activations.For BIOT, probing uses 16 bipolar derivations, each assigned the midpoint of its electrode pair.Source: §2 · p. 2

3. SPATIAL GROUNDING

The spatial probes separate sensitivity to channel identity and order from cross-channel locality and effects on downstream predictions. They use controlled channel perturbations and specified evaluation procedures.

3.1. Spatial Probes

  • SPA1 measures embedding displacement after deranging positional labels with waveforms fixed, and separately tests channel-order and signal-to-location correspondence.CBraMod's order-derived spatial structure is tested by permuting input order; other models are invariant to that permutation.Source: §3.1 · p. 2
  • SPA2 perturbs each input channel with white noise and measures its influence on each output channel, then summarizes off-channel influence and its relation to electrode distance.The probe is applied after each transformer block, excluding the perturbed channel from the distance correlation.Source: §3.1 · p. 2
  • SPA3 tests whether spatial sensitivity affects motor-imagery predictions by permuting positional information at test time while keeping the input unchanged.A linear head is fit on clean PhysioNet-MI features; the study reports latent displacement and balanced-accuracy change, excluding CBraMod.Source: §3.1 · p. 2

3.2. Experimental setup

  • SPA1 uses ten label derangements on both datasets, SPA2 runs through the penultimate layer on PhysioNet-MI, and SPA3 uses subject-disjoint cross-validation with three positional permutations per fold.SPA1 and SPA3 use PCA fitted to clean embeddings and applied unchanged to perturbed representations.Source: §3.2 · p. 2

4. SPECTRAL GROUNDING

The section probes spectral information beyond power, frequency-specific oscillatory sensitivity, and whether these sensitivities carry into clinical predictions. It uses controlled interventions and linear readouts, including a phase-free spectral baseline.

4. SPECTRAL GROUNDING

  • The spectral probes address information beyond phase-free power, oscillatory activity across canonical bands, and clinical morphology across models with different spectral input biases.BIOT consumes STFT magnitudes; LaBraM uses Fourier amplitude and phase; CBraMod and CSBrain add magnitude-spectrum projections; REVE uses temporal patches.Source: §4 · p. 2

4.1. Spectral Probes

  • SPE1 uses IAAFT surrogates that preserve amplitude distributions and approximately preserve power spectra to test whether embeddings retain waveform structure beyond those properties.Encoder performance is benchmarked against phase-free spectral power on the same task.Source: §4.1 · p. 3
  • SPE2 injects oscillations in delta, theta, alpha, or beta bands and uses unit-normalized latent displacement to predict the injected band.Injection power is measured against the recording's power in the same band, so the aperiodic spectrum does not disproportionately weight frequencies.Source: §4.1 · p. 3
  • SPE3 tests whether spectral sensitivity carries into clinical predictions by measuring seizure-probability responses to epileptiform patterns versus pink-noise control.The patterns are injected at 0 dB into held-out non-seizure segments and a contiguous 20–40% channel subset.Source: §4.1 · p. 3

4.2. Experimental setup

  • SPE1 uses subject-disjoint five-fold cross-validation on PhysioNet-MI and leave-one-subject-out on Siena, while SPE2 and SPE3 use held-out-participant evaluation on their respective datasets.SPE2 averages across three carrier realizations per band and injection level, ranging from −20 to +5 dB relative to original band power.Source: §4.2 · p. 3

5. RESULTS AND DISCUSSION

Spatial and spectral probes reveal that information encoded in EEG foundation models does not necessarily shape their downstream decisions. Model-specific sensitivity varies across channel identity, topographic locality, oscillatory structure, and clinical waveform selectivity.

5.1. Spatial grounding

  • BIOT’s learned slot tokens produced the largest identity displacement, 0.687 on PhysioNet-MI and 0.517 on Siena Scalp, while REVE’s 3D coordinates produced the smallest, 0.272 and 0.140, respectively.Breaking signal-to-location correspondence gave a similar model-specific sensitivity ordering across both datasets.Source: §5.1 · p. 3
  • REVE and LaBraM showed stronger perturbation effects on nearby than distant electrodes, while CSBrain and CBraMod showed little distance structure and BIOT was approximately flat.For REVE and LaBraM, distance correlations were ρ = -0.829 and -0.811; the strongest topographic organization appeared in LaBraM’s intermediate layers.Source: §5.1 · p. 3
  • On PhysioNet-MI, permuting positional codes substantially altered REVE, LaBraM, and CSBrain representations, but LaBraM had no corresponding motor-imagery accuracy loss; BIOT’s decision-level effect was unassessable because its readout performed near chance.Representations became broadly mixed, but this did not systematically preserve or discard their spatial biases.Source: §5.1 · p. 4

5.2. Spectral grounding

  • REVE and LaBraM distinguished EEG from phase-randomized surrogates and decoded injected bands at lower perturbation levels, whereas CSBrain and CBraMod often matched or fell below the spectral baseline.CSBrain and CBraMod were most impaired by weaker theta and alpha injections, with performance recovering in beta.Source: §5.2 · p. 4, Figure 3 · p. 4, Figure 4 · p. 4
  • Only REVE’s epileptiform specificity exceeded the log-PSD control; several encoders also responded strongly to matched-power pink noise, so spectral sensitivity did not ensure clinical waveform selectivity.Specificity is the mean epileptiform response minus the noise-control response.Source: §5.2 · p. 4, Table 2 · p. 4

5.3. Representations do not predict decision reliance

  • BIOT showed the largest channel-identity displacement without measurable decision-level reliance, while REVE had the clearest positional and morphological decision effects despite the smallest identity displacement.LaBraM’s robust surrogate and oscillation decoding likewise did not translate into specificity to clinical morphology.Source: §5.3 · p. 4

6. CONCLUSION

Across five EEG foundation models, perturbation probes distinguish what spatial and spectral structure is encoded from what downstream predictions use. The results show that sensitivity or encoding alone need not govern decisions, while the analysis is bounded by limited Siena participation and linear readouts on frozen encoders.

6. CONCLUSION

  • Across five models, spatial and spectral profiles were not determined by architecture or captured by downstream accuracy.Channel identity could be encoded without topographic locality or classification influence, while oscillatory sensitivity did not imply pathological-morphology selectivity.Source: §6 · p. 4
  • The analysis is limited by the small number of Siena Scalp participants and linear readouts on frozen encoders, which capture only linearly accessible information and may not reflect fully fine-tuned models.The authors call for broader ecological and clinical validation.Source: §6 · p. 4
  • The paper concludes that what a representation encodes need not govern its predictions.Source: §6 · p. 4

7. COMPLIANCE WITH ETHICAL STANDARDS

The study used publicly available, de-identified human EEG data and collected no new participant data.

7. COMPLIANCE WITH ETHICAL STANDARDS

  • This study was a secondary analysis of publicly available, de-identified human EEG data.Source: §7 · p. 5
  • No new human-participant data were collected.Source: §7 · p. 5
  • Ethical approval and informed consent for the original data collections are documented in the corresponding dataset publications.Source: §7 · p. 5
Loading 2610.29367v1…