mne-artifacts

MNE Artifact Correction (grill → analyze → critic)

Artifact correction of neurophysiology data via the MNE MCP server. This skill is skeptical by design: a bad decontamination runs without any error — ICA happily over-fits, removes neural signal, or is fit on the wrong data — so the discipline is to grill the artifact inventory before cleaning and critique the cleaned data before believing.

Companion skills: mne-mcp-guard for technical execution safety (ICA convergence, units); and mne-methodology-critic for Phase 3. Loaded objects persist in one MNE session.

---

PHASE 1 — GRILL (before cleaning anything)

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

Artifact inventory & claim

• Which artifacts are actually present — blink, saccade, ECG, muscle, line noise, electrode pop, drift? (Look first; don't assume.) Do EOG/ECG reference channels exist?
• What downstream analysis is this for, and could cleaning bias the comparison? (e.g. removing an ECG component differently across groups)

The two questions that decide validity

• Is ICA fit on a ~1 Hz high-passed COPY? ICA assumes stationarity; slow drifts make components unstable. Fit on a 1 Hz high-passed copy, then apply the unmixing to the 0.1 Hz data you actually analyze. (This is the single most common fatal error here.)
• Is n_components ≤ the data rank? Asking for more components than the rank yields unstable, uninterpretable components. ⚠️ An average reference and each interpolated channel REDUCE rank by ≥1 — count them.

Identification & selection

• How are artifact components identified — objectively (EOG/ECG correlation, ICLabel) or by eye? Subjective selection is not reproducible.
• How many components removed, and is the rule fixed across subjects (same threshold/labeller), or hand-picked per subject?