Reproduce “EEG-Xplain: Decoding Neural Black-Boxes of EEG Foundation Models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Hansong Ma, Junxiao Wang

EEG-Xplain audits EEG foundation-model explanations by mapping attribution methods to channel-, time-, and frequency-level features, then masking ranked features to test score changes. In CBraMod’s Mumtaz2016 healthy-control task, 100 true-positive samples produced a channel-ranking correlation of ρ = 0.981, p < .001, with masking effects. Results vary by model, task, method, and masking protocol, and do not establish biological causes or clinical validity.

Key insights

  • Spatial explanations were generally more faithful than temporal explanations: models often identified plausible brain regions and rhythms, but were less dependable at isolating the precise timing of EEG events.
  • CBraMod and EEGMamba produced patterns consistent with known resting-state, depression-related, and abnormal-EEG characteristics, including posterior alpha activity and frontal or temporal abnormalities.
  • Explanation quality depended strongly on the attribution method and architecture; Grad-CAM was often ineffective for Transformer-style representations, while LaBraM showed weak temporal fidelity, underscoring the need for multi-method auditing.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.