MNE Preprocessing (grill → analyze → critic)

Preprocessing & data-quality cleanup of neurophysiology data via the MNE MCP server. This skill is skeptical by design: nearly every preprocessing mistake (a too-aggressive high-pass that eats your slow ERP, a reference that fabricates connectivity, filtering before epoching, differential bad-channel handling between groups) runs without any error and silently biases everything downstream — so the discipline is to grill the downstream analysis before cleaning, and critique the pipeline before believing.

Companion skills: mne-mcp-guard for technical execution safety; mne-methodology-critic for Phase 3. Loaded objects persist in one MNE session, and preprocessing edits are in place.

PHASE 1 — GRILL (before processing anything)

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

The question that decides every parameter

• What is the DOWNSTREAM analysis? Filter choices follow from it, and there is no neutral

default: 0.1 Hz high-pass for ERP (slow components survive), ~1 Hz high-pass for ICA (drift hurts decomposition), broadband / no/low high-pass for TFR / low-frequency power. Set the pipeline to the analysis, not the other way round.

Filtering

• High-pass edge — and ⚠️ will the filter edge distort the effect of interest? A 0.5–1 Hz HP can

attenuate/shift slow ERP components (CNV, P300, readiness potential). Justify the edge.

• Low-pass edge (anti-alias / smoothing) and transition bandwidth; FIR (linear-phase, default)

vs IIR (causal, ringing tradeoffs)?

• Line frequency: 50 or 60 Hz? (region-dependent) — notch it (+ harmonics) or rely on a low-pass

below it?

Resampling

• Needed at all? If so, before or after epoching? Downsample after epoching where possible —

resampling continuous data jitters event sample positions. New sfreq must stay > 2× the low-pass (anti-aliasing) — MNE low-passes on resample, but verify.

Reference & montage

• Montage system (standard_1020, biosemi64, GSN-HydroCel-128, …)? Needed for interpolation,

topomaps, and source work.

• Reference rationale: average / REST / linked-mastoid / bipolar — and **does it suit the later

analysis? Average/common references inflate apparent connectivity and matter for source** modeling; pin the reference to those needs now, not after.

Bad channels & quality

• Bad-channel criteria: objective (variance/flatness/correlation thresholds, e.g. PyPREP/RANSAC)

or eyeballed? Will the two groups be treated differently (differential handling fabricates group effects)?

• Will interpolation be used, and how many channels — interpolation reduces data rank, which

caps later ICA / source n_components.

• Segments to crop or annotate-and-reject (and the same objective vs eyeballed concern)?

PHASE 2 — ANALYZE

1. Capability + look first. mne_check_status; then look before cleaning: mne_plot_psd

(read the PNG — find the line-noise spike at 50/60 Hz + harmonics, drift, dead/noisy channels) and mne_plot_raw (drifts, flat lines, jumps). mne_plot_sensors to confirm the montage.

1. Montage (if positions aren't already set) — needed for interpolation/topomaps:

mne_set_montage(name="raw", montage="standard_1020").

1. Filter + notch, with the edge chosen for the downstream analysis:

   # ERP pipeline: gentle high-pass preserves slow components; low-pass + line notch
   raw.filter(l_freq=0.1, h_freq=40.0, fir_design="firwin")   # FIR, linear phase
   raw.notch_filter(freqs=[50, 100, 150])                       # 50 Hz region + harmonics

(Structured form: mne_filter(name="raw", l_freq=0.1, h_freq=40, notch=50). For ICA use l_freq=1.0; for TFR/low-freq power keep the high-pass minimal/off.)

1. Reference, chosen for the downstream need:

   raw.set_eeg_reference("average", projection=False)          # whole-head, common for source/ERP
   # linked mastoid:  raw.set_eeg_reference(["TP9", "TP10"])
   # REST (infinity): raw.set_eeg_reference("REST")            # needs a forward / sphere model

(Structured form: mne_set_reference(name="raw", ref_channels="average") / "REST" / "TP9,TP10".)

1. Bad channels + interpolate — prefer an objective criterion, then spline-interpolate:

   # objective flag: flat or extreme-variance channels (illustrative thresholds)
   data = raw.get_data(picks="eeg"); v = data.var(axis=1)
   names = [raw.ch_names[i] for i in range(len(v))]
   flat = [names[i] for i in np.where(v < 1e-14)[0]]
   noisy = [names[i] for i in np.where(v > np.median(v) + 5 * np.std(v))[0]]
   raw.info["bads"] = sorted(set(flat + noisy))
   raw.interpolate_bads(reset_bads=True)                        # spline; needs montage

(Structured form: mne_mark_bad_channels(name="raw", bads="Fp1,T7") then mne_interpolate_bads(name="raw"). Record how many channels were interpolated — it lowers data rank for any later ICA/source step.)

1. Resample — after epoching when you can; if on continuous data, keep events safe:

mne_resample(name="raw", sfreq=250) (verify new sfreq > 2× the low-pass).

1. Crop / annotate bad segments: mne_crop(name="raw", tmin=0, tmax=300), or mark

BAD_* annotations so epoching drops them.

1. Re-inspect. mne_plot_psd again — confirm the line-noise spike is gone, the high-pass

removed drift, and no channel still looks dead/noisy. Then archive the equivalent code + figures (the mne-analyst archiving convention).

Best-practice reminders: filter before epoching to avoid edge artifacts at every epoch boundary; pick the high-pass for the analysis; report per-group bad-channel/rejection counts; track how many channels were interpolated (rank).

PHASE 3 — CRITIC (before believing the result)

Hand the pipeline (filter band, reference, bad channels, resampling, order of operations) to `mne-methodology-critic` (invoke the skill, or dispatch it as a subagent with references/methodology-checklist.md). For preprocessing it will specifically check:

• high-pass distortion of slow ERP components (≳0.1–0.3 Hz can attenuate/shift CNV/P300);
• filter applied before vs after epoching (post-epoch filtering injects edge artifacts at

every boundary);

• reference choice biasing later connectivity / source (average/common references inflate

apparent connectivity and shift source estimates);

• differential bad-channel handling between groups (unequal interpolation fabricates group

effects);

• interpolation reducing data rank (caps later ICA / source n_components);
• resampling-induced event jitter (downsampling continuous data before epoching shifts event

timing).

Report its BLOCK / REVISE / PASS verdict to the user and act on it before stating that the data are "clean."

See references/preprocess-methods.md for deeper recipes (FIR vs IIR & transition bandwidth, notch strategies, reference math, objective bad-channel detection, rank bookkeeping, and pipeline order).