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?

Bias & reporting (pin this down NOW)

• Will the artifact rate differ across groups/conditions (clinical vs control blink rates;

high- vs low-load muscle)? Differential cleaning manufactures effects.

• Over-cleaning: how do you guard against removing neural signal (e.g. an occipital alpha or a

genuine frontal component)?

• Plan to report n components removed, the labelling method, and per-group rejection rates.

PHASE 2 — ANALYZE

1. Capability + look first. mne_check_status; mne_plot_raw and read the PNG — confirm

which artifacts are present (blinks, heartbeat, muscle bursts, line noise) before deciding.

1. High-pass a copy for fitting (do not high-pass the data you analyze):

   raw_for_ica = raw.copy().filter(l_freq=1.0, h_freq=None)   # ICA-only copy

1. Fit ICA on the copy, with n_components ≤ rank. Use the structured tool, or mne_run_code

when you need the copy / rank control: mne_fit_ica(name="raw_for_ica", n_components=0.99, method="picard"). (fastica/infomax/picard; pass random_state for reproducibility.)

1. Plot and read the components. mne_plot_ica_components (scalp topographies) and

mne_plot_ica_sources (time courses) — read both PNGs: blink = frontal topo + slow square waves; ECG = ~1 Hz periodic spikes; muscle = high-freq edge/temporal; line = narrowband.

1. Identify objectively, not by eye (via mne_run_code): correlate with EOG/ECG channels, or

label with ICLabel:

   eog_idx, _ = ica.find_bads_eog(raw)        # needs an EOG channel (or by name)
   ecg_idx, _ = ica.find_bads_ecg(raw, method="correlation")
   ica.exclude = sorted(set(eog_idx + ecg_idx))
   # automatic labelling (extra deps): from mne_icalabel import label_components
   #   labels = label_components(raw_for_ica, ica, method="iclabel")

1. Exclude and apply to the 0.1 Hz data (not the high-passed copy):

mne_apply_ica(ica_name="ica", inst_name="raw", exclude="0,3").

1. Alternatives where appropriate — SSP (project out a blink/ECG subspace), or autoreject

(automated per-channel epoch thresholds / interpolation). Both via mne_run_code and need extra deps from the `[full]` extra (autoreject, mne-icalabel).

1. Archive the equivalent code + figures, and report n components removed (the mne-analyst

archiving convention).

Best-practice reminders: fit on the 1 Hz copy, apply to the 0.1 Hz data; count avg-ref/interp toward rank; use a fixed, objective selection rule across subjects; report per-group rejection.

PHASE 3 — CRITIC (before believing the cleaned data)

Hand the plan + cleaned data to `mne-methodology-critic` (invoke the skill, or dispatch it as a subagent with references/methodology-checklist.md). For artifact work it will specifically check:

• ICA fit on non-high-passed data (unstable components);
• n_components > data rank (avg-ref / interpolated channels reduce rank by ≥1);
• subjective / inconsistent component selection vs objective (EOG/ECG/ICLabel);
• over-cleaning that removes neural signal;
• differential rejection between conditions/groups manufacturing an effect;
• not reporting the number of components removed or the labelling method.

Report its BLOCK / REVISE / PASS verdict to the user and act on it before stating conclusions.

See references/artifact-methods.md for deeper recipes (ICA prerequisites & rank, fastica/infomax/ picard, component identification, ICLabel, SSP, autoreject, and regression EOG).