Technical Summary: Interpreting EEG Foundational Transformers with LRP

Problem Statement

The adoption of Foundation Models (FMs) in electroencephalography (EEG) offers a potential solution to data scarcity in deep learning diagnostics and brain-computer interfaces (BCIs). However, the "opaque" nature of these models, particularly those based on Transformer architectures, hinders their wider adoption and clinical trust. While feature attribution methods like Layer-wise Relevance Propagation (LRP) have been successfully applied to convolutional neural networks (CNNs) and multi-layer perceptrons (MLPs) in EEG, their application to modern Transformer-based EEG-FMs remains unexplored. The core challenge is to determine if LRP can effectively probe the decision processes of these complex models to verify their validity or uncover spurious learning strategies (e.g., "Clever Hans" effects) without relying on ground-truth labels for every task.

Methodology

The authors propose the application of attention-aware LRP [8] to the LaBraM model [2], a prominent EEG Foundation Model. The methodology involves three primary components:

1. Model Configuration: The study evaluates LaBraM under three configurations to isolate pretraining effects:

   • Finetuned: Standard approach where the whole model is adapted.
   • Frozen: The pretrained backbone is frozen, with only a trainable classification head.
   • From Scratch: Random initialization without pretraining.
     Each configuration is trained five times to account for variability in dropout, initialization, and data sampling.

2. Attribution and Visualization:

   • LRP is applied to generate relevance scores for every input element (time step and channel) for each classification decision.
   • To manage the high dimensionality of EEG data ($n_{channels} \times t_{in}$) and the volume of generated scores, the authors employ an aggregation strategy. They aggregate attributions across time to analyze spatial patterns (topography) and across channels to analyze temporal patterns.
   • Spatial visualizations utilize discrete Voronoi-cell coloring to preserve electrode topology, avoiding the distortion of rectangular heatmaps.

3. Experimental Paradigms:

   • R-Peak Detection (Validation): A binary classification task predicting cardiac field artifacts (CFA) in raw EEG. This serves as a ground-truth validation because the physiological source (heart) and expected electrode location (Iz) are known.
   • PhysioMI L/R (Established Paradigm): A motor imagery/execution task (left vs. right fist) using the PhysioNet dataset. This tests the model's ability to generalize and reveals potential shortcut learning.
   • AffectiveVR (Exploratory): A naturalistic paradigm predicting subjective arousal and valence during a VR experience. This task lacks clear ground truth, serving as a testbed for exploratory hypothesis generation.

Key Contributions

The paper makes three specific contributions:

1. Validation of LRP on EEG-FMs: It demonstrates for the first time that attention-aware LRP produces plausible attributions on foundational EEG Transformers. This is validated against the R-Peak detection task, where relevance correctly concentrates on the Iz electrode, the site closest to the heart.
2. Detection of Spurious Behavior: The method successfully identifies "Clever Hans" behavior in motor imagery models. It reveals that models with poor generalization rely heavily on ocular (eye-movement) artifacts rather than motor cortex signals, likely exploiting visual cues from the experimental design.
3. Exploration of Novel Hypotheses: In the absence of established intuition for complex affective states, the authors characterize the unique promise and limits of LRP. They identify a recurring, biologically plausible pattern of central electrode relevance (C2, Cz, FC2, CP2) in arousal prediction, suggesting a novel sensorimotor signature of arousal distinct from valence.

Results

• Performance:
  • R-Peak Detection: The finetuned LaBraM achieved a balanced accuracy (BAC) of ~75%, confirming the model learns the correct physiological signal.
  • PhysioMI L/R: Performance was mixed. Two out of five finetuned runs achieved >70% BAC, while others performed near random guessing. Crucially, the attribution maps distinguished these groups: successful runs showed lateralized temporo-frontal patterns, while unsuccessful runs focused on ocular regions.
  • AffectiveVR: All models (finetuned, from scratch, frozen) and the CSP-LDA baseline performed similarly, with BACs around 55% (slightly above the 50% random baseline for median-split binary classification).
• Attribution Patterns:
  • Sign Inversion: Patterns were consistent in spatial structure but occasionally sign-inverted (reversed coloring), which does not affect the magnitude of relevance.
  • Motor Imagery: Successful models showed lateralized patterns, but even these focused on temporal electrodes closer to eye-movement regions than hand-motor regions, suggesting a reliance on task-correlated eye movements rather than pure neural motor signatures.
  • Arousal vs. Valence: Arousal prediction consistently highlighted a central electrode cluster (sensorimotor cortex), potentially linked to readiness potentials modulated by emotional context. This pattern was absent in valence prediction, supporting the hypothesis that it is an arousal-specific signature rather than a general motor decoding artifact.
  • Valence Prediction: Patterns were less consistent with established literature and showed ambiguity, highlighting the difficulty of interpreting heatmaps in low-intuition domains where spurious artifacts (e.g., EMG from facial muscles) are hard to distinguish from neural signals.

Significance and Claims

The paper positions LRP as a dual-purpose tool for EEG-FMs: verification and exploration.

• Verification: LRP can validate that models are learning the intended physiological signals (as seen in R-Peak detection) and expose overfitting or shortcut learning (as seen in the ocular reliance of failed motor imagery models). This is critical in EEG, where signals are not directly interpretable and shortcut learning has historically invalidated research lines.
• Exploration: By translating learned representations into neuroscientific hypotheses, LRP can surface novel, biologically plausible relationships (e.g., the central sensorimotor signature in arousal). The authors claim this moves EEG-FM research beyond pure predictive performance toward understanding brain-behavior relationships.

Limitations and Modesty:
The authors explicitly acknowledge three caveats:

1. Interpretability: Heatmap interpretation in EEG remains subjective and requires domain expertise; nonsensical strategies are harder to identify than in natural images.
2. Methodological Basis: The choice of LRP is based on comparisons conducted on non-Transformer models; further faithfulness measures (e.g., AOPC) on EEG-FMs are needed.
3. Model Capability: The lack of consistent performance gains over "from scratch" or specialized models (like CSP-LDA) suggests that current EEG-FMs have not yet fully leveraged their "underlying understanding" of EEG. Consequently, the discovery potential of LRP is currently constrained by the capabilities of the underlying models.

The paper concludes that while LRP is a powerful tool for the trajectory of EEG-FM development, its utility for clinical deployment or definitive discovery is currently bounded by the maturity of the foundation models themselves.