mne-decoding

MNE Decoding / MVPA & BCI (grill → analyze → critic)

Multivariate decoding of neurophysiology data via the MNE MCP server. This skill is skeptical by design: the most damaging decoding mistakes — data leakage and a wrongly assumed chance level — produce a clean, plausible accuracy curve without any error, so the discipline is to grill the cross-validation before fitting and critique before believing.

Companion skills: mne-mcp-guard for technical execution safety; mne-methodology-critic for Phase 3. Loaded objects persist in one MNE session. Decoding needs scikit-learn.

---

PHASE 1 — GRILL (before fitting anything)

Do not fit a classifier until these are answered. If the user can't answer one, propose a sensible default and explicitly flag the open risk — never silently choose.

What is being decoded

• Which two (or more) conditions / labels, and what is the scientific claim tied to decodability?
• Class balance and sizes — n trials per class, per subject? (Imbalance silently inflates accuracy and breaks the nominal chance level.)
• Feature space: sensors × time? band power? source space? What is the classifier actually seeing?

The two questions that decide validity

• Cross-validation structure. Subject-level (leave-one-subject-out) or trial-level? If trial-level on pooled multi-subject data, do trials from one subject leak across train/test folds (⇒ identity decoding, inflated)? Is the split stratified by class? (This + leakage are the two fatal errors here.)
• Is every transform fit INSIDE the fold? Scaling, feature selection, ICA, PCA, even baseline z-scoring must be fit on training data only within each CV fold (use an sklearn Pipeline). Anything fit on the full dataset before CV = leakage ⇒ optimistic, invalid.

Inference plan (pin this down NOW, not after seeing results)

• Chance level — established by label permutation (shuffle labels, re-decode many times), not the nominal 1/n_classes. Imbalance and small n move true chance off 1/n.
• Multiple comparisons across time — a classifier per time point ⇒ many tests; plan a cluster-based permutation test of scores-vs-chance, not per-time-point thresholding.
• Temporal-generalization claims. Will off-diagonal generalization be read as maintenance / reactivation of a representation? That is a strong claim — state it in advance and guard it (it can also reflect a slow/sustained component, not reactivation).
• Metric: ROC-AUC / balanced accuracy (imbalance-robust) over raw accuracy?