What Do EEG Foundation Models Capture from Human Brain Signals?

Technical Summary: What Do EEG Foundation Models Capture from Human Brain Signals?

Problem Statement
Clinical electroencephalogram (EEG) analysis has historically relied on a hand-crafted feature catalog refined over decades, encompassing band power, connectivity, complexity, and other signal properties. Recently, EEG foundation models have emerged, utilizing self-supervised pretraining on raw signals to achieve state-of-the-art performance on clinical benchmarks, often bypassing the traditional feature engineering pipeline. A critical open question remains: do the high-dimensional representations learned by these foundation models align with the established clinical feature lexicon, or have they diverged to discover novel signal structures? Understanding this alignment is essential for validating model decisions against human knowledge and identifying gaps where new clinical descriptors may be needed.

Methodology
The authors propose a three-part audit framework to decompose the alignment question into: (1) What does the model learn? (Encoding), (2) What does the model use? (Causal reliance), and (3) How much can be explained? (Performance recovery). This audit is applied across three foundation models (CSBrain, CBraMod, LaBraM), five clinical tasks (MDD, Stress, ISRUC-Sleep, TUSL, Siena), and a lexicon of 63 hand-crafted features grouped into six families (time-domain morphology, spectral power/shape, time-frequency envelope, signal complexity, cross-frequency coupling, and cross-channel relations).

The methodology employs three distinct interpretability techniques for every (dataset, model, feature) triplet:

1. Layer-wise Ridge Probing: To determine encoding, linear ridge regression probes predict each hand-crafted feature from the model's activations at every layer. A feature is considered "encoded" if a probe achieves significant predictive power ($R^2$) on a held-out validation set, identifying the specific layer where the information resides.
2. LEACE-Style Cross-Covariance Subspace Erasure: To determine causal usage, the authors intervene on the model's activations. At the encoding peak layer, they remove the linearly decodable component of a target feature from the activation vector using a cross-covariance subspace erasure (a scalable approximation of the LEACE operator). The resulting drop in task performance on a held-out test set indicates causal reliance. This is rigorously controlled against three null targets (random subspace, shuffled pairing, Gaussian noise) and a linear residual probe to rule out spurious effects.
3. Transparent Classifier Closure: To measure explainability, a logistic regression classifier is trained solely on the set of confirmed "representation-causal" features. Its performance is compared against the foundation model and a same-dimension random-feature baseline. The "closure ratio" quantifies the fraction of the foundation model's advantage over the random baseline that is recovered by the hand-crafted features.

Key Contributions

• Causal Audit Framework: The paper introduces a rigorous pipeline that distinguishes between features a model encodes (has in its representation) and features it causally uses for decision-making, addressing the limitation of correlation-based interpretability methods.
• Comprehensive Lexicon Audit: The study audits 945 units (3 models $\times$ 5 tasks $\times$ 63 features), providing the first large-scale causal map of how foundation models relate to the clinical EEG feature canon.
• Feature Taxonomy: The authors categorize features into universal candidates (strong support across architectures/tasks), task-specific, model-specific, and encoded-only, offering a structured view of model reliance.

Results

• Encoding: The foundation models broadly recover the clinical lexicon. 89.6% (847/945) of the units pass the encoding criterion, with all 63 features encoded in at least one (model, task) cell. Frequency-domain features are encoded with high strength across all cells.
• Causal Usage: 68.6% (648/945) of the units are "representation-causal," meaning the model relies on them for task decisions. 21.1% are encoded but not causally used. Across tasks, 50 features qualify as universal candidates with strong support in two or more tasks. While frequency-domain features dominate, time-frequency, cross-channel, complexity, time-domain, and cross-frequency families each contribute substantial causal mass on at least four of the five tasks.
• Performance Recovery (Closure): Confirmed features recover, on average, 79.3% of the foundation model's advantage over a random-feature baseline. A clear task gradient exists: tasks near the performance ceiling (e.g., MDD, closure $\approx$ 0.99) are almost fully explained by the lexicon, while harder tasks (e.g., Stress, closure $\approx$ 0.56) leave a non-trivial residual.
• Residuals: The unexplained performance gap on difficult tasks suggests the existence of signal structures not captured by the current 63-feature lexicon or non-linear combinations not captured by the linear surrogate.

Significance
The paper claims that EEG foundation models, despite being trained without inductive bias toward hand-crafted features, recover the clinical feature catalog at scale. However, the distinction between encoding and usage is critical; models encode many features but rely causally on a specific subset. The findings suggest that for tasks near the performance ceiling, existing clinical descriptors are largely sufficient to explain model behavior. Conversely, the residual performance on harder tasks pinpoints concrete targets for future concept discovery, indicating that the current catalog of EEG descriptors is incomplete. This work establishes a feedback loop between model representations and human knowledge, providing a method to validate existing clinical descriptors and identify where new ones are needed for AI-driven neuroscience.