| name | eeg-preprocessing-reliability |
| description | EEG preprocessing reliability methodology for quantifying and mitigating preprocessing-induced prediction instability in EEG deep learning. Based on arXiv:2605.07212 (Hou et al., 2026). Use when: (1) evaluating EEG model robustness to preprocessing pipeline choices, (2) designing preprocessing-stable BCI/decoding systems, (3) analyzing counterfactory prediction flip rates (CFR), (4) implementing Preprocessing Uncertainty (PU) diagnostics, (5) applying NA-PGI regularization for graph-structured preprocessing consistency, (6) performing Walsh-Hadamard decomposition of preprocessing intervention space. Activation: eeg preprocessing reliability, preprocessing sensitivity, CFR, preprocessing uncertainty, PU, NA-PGI, EEG pipeline stability, counterfactual intervention, preprocessing flip rate, EEG deep learning reliability, 预处理可靠性, 脑电预处理敏感性.
|
EEG Preprocessing Reliability
Based on: "Same Brain, Different Prediction: How Preprocessing Choices Undermine EEG Decoding Reliability" (Hou et al., arXiv:2605.07212, May 2026).
Core Problem
EEG deep learning predictions are unstable under preprocessing choices. Across 6 datasets, up to 42% of trial-level predictions flip when only preprocessing changes (same raw data, same model, same labels). Standard uncertainty methods (softmax entropy, MC Dropout, ensembles) hold preprocessing fixed and do not capture this.
Seven Atomic Preprocessing Interventions
| # | Intervention | Option A (0) | Option B (1) | Impact |
|---|
| 1 | Reference | Common avg ref | Original | Medium |
| 2 | High-pass filter | 0.1 Hz | 0.5 Hz | High |
| 3 | Low-pass filter | 45 Hz | 30 Hz | High |
| 4 | Baseline correction | None | 200ms subtractive | High |
| 5 | Artifact attenuation | Off | ASR | High |
| 6 | Epoch rejection | Off | Autoreject | Med-High |
| 7 | Bad-channel repair | Off | RANSAC+interp | Medium |
These form a 7D Boolean lattice (128 pipelines, 448 edges).
Key Findings
Sensitivity Varies by Task
| Dataset | Task | Channels | Mean Acc (%) | CFR (%) |
|---|
| BCI-IV-2a | MI (4-cls) | 22 | 37.6 | 42.4 |
| SEED-IV | Emotion (4-cls) | 62 | 31.9 | 35.8 |
| PhysionetMI | MI (2-cls) | 64 | 57.7 | 21.7 |
| Sleep-EDF | Sleep (5-cls) | 2 | 85.6 | 9.6 |
| Lee2019-ERP | P300 (2-cls) | 62 | 84.1 | 4.1 |
| P300 | ERP (2-cls) | 16 | 83.4 | 2.6 |
CFR inversely correlates with accuracy: preprocessing matters most when model is least confident.
Sensitivity is Near-Additive
Walsh-Hadamard decomposition: interactions contribute 0.2% of total variance. Greedy step-by-step optimization achieves within 2.5% of oracle best-of-128 on all datasets. However, additivity may not extend to continuous preprocessing parameters.
Dominant Interventions are Task-Specific
- BCI-IV-2a: epoch rejection dominates (20.9%)
- Sleep-EDF: high-pass filtering dominates
- P300: high-pass filtering dominates
- Spearman rank correlations of intervention importance: mean 0.32 (non-transferable across tasks)
Diagnostic Metrics
Counterfactual Flip Rate (CFR)
Fraction of test trials whose prediction changes across all 128 pipelines:
CFR = (1/|P|-1) * sum_{p != p_ref} I[f_p(x) != f_{p_ref}(x)]
Preprocessing Uncertainty (PU)
Per-trial measure of pipeline disagreement, complementary to softmax entropy. Correlates only moderately with entropy (rho=0.42) and MC Dropout (rho=0.31).
Per-Intervention Effect
Average accuracy change when toggling one intervention across 64 pipeline pairs.
Mitigation: NA-PGI
Normalized Adaptive Preprocessing-guided Intervention (NA-PGI): graph-structured regularizer exploiting the compositional lattice of preprocessing interventions.
- Reduces CFR by up to 35%
- Single transferable hyperparameter (lambda)
- Edge-level logit consistency with logit-variance normalization
- Most effective on high-density EEG settings
Protocol for Evaluating Preprocessing Sensitivity
- Select anchor pipeline (p0)
- Train model on p0 only
- Evaluate on all 128 counterfactual pipelines
- Compute CFR and per-intervention effects
- Identify dominant interventions for the task
- (Optional) Apply NA-PGI for mitigation
Implementation Notes
- All preprocessing via MNE-Python
- 128 variants: ~min/subject (CPU)
- ERM training: ~GPU-hour/dataset on A100
- EEGNet-v4 with 3-fold subject-wise CV
- Generate 128 counterfactual views per raw recording
When to Apply
- Before reporting EEG decoding results: quantify preprocessing sensitivity
- When deploying BCI systems across sites: verify pipeline stability
- When comparing EEG studies: account for preprocessing-induced variability
- When training EEG foundation models: evaluate preprocessing robustness
Code Reference
Original implementation: https://github.com/dengzhe-hou/EEG-Preprocessing-Sensitivity
Related
- vlm-visual-cortex-alignment-robustness: complementary robustness analysis
- eeg-preprocessing-reliability: EEG decoding reliability assessment
- same-brain-different-prediction: methodological concern for all EEG studies
