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

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

Problem Statement
Electroencephalography (EEG) Foundation Models (EEG-FMs) have emerged as a promising approach for Brain-Computer Interfaces (BCIs), aiming to learn general-purpose representations from large unlabelled corpora via self-supervision. However, current evaluations of these models are predominantly limited to "clean," in-distribution accuracy metrics. This focus overlooks critical factors for real-world deployment: the inherent noise of EEG signals (due to electrode impedance, movement, and environmental artifacts), the need for interpretability in high-risk clinical domains, and the actual quality of the learned representations. Previous benchmarks have shown that EEG-FMs often offer only marginal gains over supervised baselines like EEGNet, yet the root causes—whether poor robustness, misaligned feature learning, or low-quality pre-trained representations—remain unexamined.

Methodology
The authors conducted a systematic benchmarking study involving six state-of-the-art EEG-FMs (BIOT, CBraMod, LaBraM, NeuroRVQ, BrainOmni, and REVE) and a supervised baseline (EEGNet) across eight diverse datasets covering motor imagery, motor execution, event-related potentials (ERP), working memory, sleep staging, and eye-state classification. The study employed three distinct layers of analysis:

1. Robustness Analysis: Models trained on clean data were evaluated at test time under five perturbation conditions: additive white and pink noise (at varying Signal-to-Noise Ratios), random channel dropout (zero-padding), region-based channel dropout (zero-padding and true removal), and region-specific noise injection.
2. Interpretability Analysis: The study applied Attention-Aware Layer-Wise Relevance Propagation (AttnLRP) to attribute model predictions to specific brain regions. For models where AttnLRP was numerically unstable (NeuroRVQ and BrainOmni), Gradient × Input (G×I) was used. The analysis examined whether models focused on neurophysiologically plausible regions and how these attention maps shifted under perturbation.
3. Expressiveness Analysis: Using block-wise linear probing, the authors investigated where task-relevant information emerges within the transformer architecture (early vs. late blocks) and whether pre-trained representations are sufficient for downstream tasks. A specific focus was placed on the impact of pooling strategies (mean pooling vs. token flattening) on the performance of "head-only" (frozen backbone) adaptation.

Key Results

• Robustness: No single model dominates all failure modes.

  • Noise: BrainOmni demonstrated the highest robustness to additive noise, losing only 5.4% accuracy at -5 dB compared to 14–21% for others.
  • Dropout: Paradoxically, BrainOmni was among the most fragile under random channel dropout when channels were zero-padded (-13.3% degradation). However, when dropped channels were removed entirely (true dropout) rather than zero-padded, variable-channel models (including BrainOmni) recovered significantly, suggesting their fragility stems from handling zero-valued inputs rather than a lack of task information.
  • Regimes: At extreme noise levels, all models converged to a task-specific performance floor, indicating a signal-limited regime where architecture matters less. At moderate noise, architectural differences remained decisive.

• Interpretability:

  • Spatial Alignment: All models broadly concentrated relevance on task-appropriate brain regions (e.g., central electrodes for motor tasks, occipital sites for ERP), aligning with known neurophysiology.
  • Perturbation Stability: Under perturbation, attribution maps remained spatially stable even as prediction accuracy degraded. This suggests models continue to attend to the correct brain regions but fail to decode the corrupted signal content.
  • Exceptions: BrainOmni produced diffuse attention maps, likely due to its operation over abstract latent tokens rather than physical electrodes.

• Expressiveness:

  • Block-wise Probing: Task-relevant linearly decodable signals were present in early blocks of pre-trained models. Fine-tuning primarily repurposed late blocks to specialize for the downstream task, a pattern consistent with findings in NLP transformers.
  • Pooling Artifacts: The previously observed poor performance of head-only adaptation for models like LaBraM and NeuroRVQ was largely attributed to their use of mean pooling, which collapses token-level information. Replacing mean pooling with token flattening significantly improved head-only performance (gains of 14–21%), bringing these models closer to the EEGNet baseline. This indicates that EEG-FMs possess sufficient representational capacity when token-level embeddings are preserved.

Significance and Claims
The paper claims to provide the first systematic assessment of robustness, interpretability, and expressiveness for EEG foundation models. The authors argue that:

1. Robustness is task- and perturbation-dependent: Rankings of model performance change based on the type of corruption (noise vs. dropout) and the specific task signal.
2. Spatial grounding exists but is insufficient: While models learn spatially grounded representations consistent with neurophysiology, this does not guarantee faithful prediction under corruption.
3. Pooling strategies confound evaluation: The perceived "low quality" of pre-trained representations in head-only settings is often an artifact of pooling design (mean pooling vs. flattening) rather than the representations themselves.
4. Future Directions: The study highlights the need for future evaluations to use concatenation-based pooling to avoid conflating pooling design with representation quality and to consider true channel removal rather than zero-padding when testing variable-channel models.

The authors conclude that while EEG-FMs show promise, their development requires rigorous evaluation beyond clean accuracy to ensure reliability in clinical and real-world BCI settings.