EEG-Xplain: Decoding Neural Black-Boxes of EEG Foundation Models
cs.AI
Quick verdict
A unified XAI framework interpreting EEG foundation models; spatial fidelity is strong, but TP-only sampling, small n, and unvalidated LLM reports weaken it.
Summary
The paper introduces EEG-Xplain, a unified post-hoc interpretability framework for EEG foundation models such as CBraMod, BIOT, LaBraM, EEGPT, and EEGMamba. It adapts six attribution algorithms from three paradigms — gradient-based (Integrated Gradients, Gradient-SHAP), perturbation-based (Occlusion, KernelSHAP, LIME), and activation-based (Grad-CAM) — through a common model-task adapter, and projects attributions onto spatial (channel/topographic), temporal (patch-level), and frequency-band (delta–gamma ablation) dimensions. Reliability is assessed at population level via Area Over the Perturbation Curve against random-masking baselines and Spearman correlation between attribution rankings and single-feature intervention effects, plus cross-method consistency. On Mumtaz2016 (depression vs. control) and TUAB (normal vs. abnormal), with a temporal analysis on TUEV, the authors report high spatial fidelity for CBraMod (e.g., IG rho=0.981 on Mumtaz HC; rho=0.979 on TUAB abnormal), topographic patterns matching known markers (posterior alpha for controls, frontal channels for MDD, temporal chains for TUAB abnormal), and generally weaker temporal fidelity, with small samples (n=5–15); Grad-CAM performs worst. A contribution claimed but not quantitatively evaluated in the extracted text is LLM-generated natural-language clinical reports from structured attributions. Main limitations: analysis restricted to true-positive samples, no multiple-comparison correction, minimal detail on LLM evaluation, and unaddressed confounds in band-ablation attribution.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 4 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A single adapter framework can apply six attribution methods across five heterogeneous EEG foundation models and project results onto spatial, temporal, and frequency dimensions.
- Population-level spatial attributions of stronger models are faithful and neurophysiologically aligned (CBraMod IG ρ=0.981 on Mumtaz HC; ρ=0.979 on TUAB abnormal, with clinically expected channel patterns).
- No attribution method is uniformly best — method choice couples to model architecture and dimension, with Grad-CAM worst overall — justifying multi-method evaluation with a recommendation mechanism.
- Spatial fidelity of current EEG foundation models is generally higher than temporal (patch-level) fidelity.
- LLMs can transform structured attribution evidence into clinically readable reports, closing the gap between attribution maps and semantic reasoning.
Strengths
Genuine cross-architecture scope with a unified adapter
Fig. 1 and §3.2 wire five heterogeneous backbones (CBraMod, BIOT, LaBraM, EEGPT, EEGMamba) and six attribution algorithms from three paradigms through one model-task interface, enabling the systematic model×method×dimension comparisons in Tables 3–5 rather than single-model case studies.
Intervention-based fidelity, not just visual plausibility
§3.5 grounds faithfulness causally: progressive top-k masking (channel mean replacement spatially, variance-matched noise temporally) yields AOPC versus random-order masking (Fig. 4c/d: 2.068 vs 0.576 for HC; 2.147 vs 0.857 for MDD), plus Spearman ρ between attribution ranks and single-feature logit drops (Eq. 17–18).
Physiologically aligned results with pre-specifiable expectations
Candid reporting of negative results
Weaknesses
Fidelity evaluated only on true positives
Statistical hygiene: no multiple-comparison correction, tiny temporal cells
Headline LLM contribution is unevaluated
Band-ablation and masking confounds unaddressed
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
EEG-Xplain (arXiv 2609.15687, cs.AI) proposes a unified post-hoc interpretability framework for EEG foundation models, evaluated on five backbones (CBraMod, BIOT, LaBraM, EEGPT, EEGMamba), six attribution algorithms (Integrated Gradients, Gradient-SHAP, KernelSHAP, Occlusion, LIME, Grad-CAM), and three public datasets (Mumtaz2016 depression triage, TUAB abnormality screening, TUEV event detection; Table 2). The authors claim three contributions (§1): a cross-architecture attribution platform, a population-level fidelity paradigm (progressive occlusion + AOPC + intervention correlation), and an LLM-based pathway from structured attributions to clinical reports. The first two are substantially delivered; the third is asserted but never quantitatively evaluated in the visible text.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 4 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is EEG-Xplain?
EEG-Xplain is a framework for explaining predictions of EEG foundation models (CBraMod, BIOT, LaBraM, EEGPT, EEGMamba). It runs six standard attribution methods (Integrated Gradients, Gradient-SHAP, Occlusion, KernelSHAP, LIME, Grad-CAM) through one adapter and reports which EEG channels, time segments, and frequency bands (delta to gamma) drove each prediction, then checks faithfulness by masking the highlighted features and measuring how much model confidence drops.
Which datasets does the paper use?
Three public clinical EEG benchmarks: Mumtaz2016 (depression vs healthy controls), TUAB (normal vs abnormal EEG), and TUEV (six-class EEG event detection). Mumtaz2016 is freely downloadable via figshare; TUAB and TUEV come from the TUH corpus, which requires registration.
What are the main findings?
Channel-level explanations of the strongest model (CBraMod) match clinical knowledge — occipital channels for healthy controls, frontal channels for depression, bilateral temporal chains for abnormal EEG — with high fidelity scores (e.g., rho=0.981 on the control class). Explanation quality varies strongly by model and method, Grad-CAM works poorly on transformer/SSM architectures, and time-segment (temporal) explanations are much less reliable than spatial ones, partly because temporal evaluations used only 5–15 samples.
Is code or data available?
The paper states that code is available on GitHub, though the full repository URL is not visible in the extracted text. All five evaluated models have public pretrained checkpoints, and all three datasets are public benchmarks (the TUH corpora require registration).
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Text extraction has omitted sections (parts of Related Work, §3.2 System, portions of §4.1); the review is based on visible fragments only.
- Figures 3–6 exist only as caption/garbled text; topographic maps and the radar chart could not be visually verified.
- Table 4 (TUAB) and parts of the temporal-results table are truncated; some ρ values are cited from surrounding prose rather than full tables.
- Apparent source-text anomalies (e.g., 'Krishnadisagreement et al. [22]', inconsistent 'CBraMod/CBRAMOD' naming) may be extraction artifacts rather than paper errors.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.