MNE ERP / ERF Analysis (grill → analyze → critic)

Evoked-response analysis of neurophysiology data via the MNE MCP server. This skill is skeptical by design: most ERP mistakes (peak-picking a window after seeing the grand average, a high-pass that shifts latency, calling an occipital 100 ms deflection "N1") run without any error — so the discipline is to grill the component identity and measurement plan before averaging, 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.

PHASE 1 — GRILL (before averaging anything)

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

Design & claim

• What is the hypothesis, and what is the comparison? (condition × condition, group × group,

pre × post)

• Within- or between-subject? Paired or independent? n per cell?
• Confirmatory (component + window + electrodes pre-specified) or exploratory (whole-head, corrected)?

The questions that decide validity

• Which components, and what are their EXPECTED latency, topography, AND stimulus modality?

Latency/polarity/topography must match the named component for that modality. ⚠️ auditory N1 is fronto-central (~100 ms); visual N1 is occipito-temporal ~150–200 ms — a ~100 ms occipital deflection in vision is most likely P1, not N1. A topography/modality mismatch means the component is misnamed. (This is the single most common identity error here.)

• Is the measurement window + electrode set PRE-SPECIFIED? A mean/peak amplitude measured in a

window/channel chosen after seeing this dataset's grand average is circular (double-dipping). Pin the window and ROI down NOW, from prior literature, not from the data.

Data & parameters

• Baseline window (length + placement)? Baseline noise propagates into every component.
• Filter. ⚠️ high-pass **> 0.3 Hz distorts slow components and shifts/inverts apparent latency

and polarity** (CDA, LPP, P300, CNV especially); low-pass smooths and delays peaks. State the band.

• Epoching: tmin/tmax, peak-to-peak rejection threshold + its justification.
• Reference, montage, channel selection, units.

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

• Trial counts balanced across conditions? Unequal trials → unequal ERP SNR → spurious

amplitude/latency differences. Will rejection differ between conditions/groups (differential rejection fabricates effects)?

• Mean or peak amplitude? Mean-in-window is more robust; peak is biased by noise and by window

length. Single-subject peak latency is noisy → plan a jackknife (group latency) with the adjusted t.

• Multiple comparisons across components × electrodes × conditions — which correction, and is

its independence assumption valid? (neighbouring channels/times are correlated → consider cluster-based permutation rather than per-channel tests)

PHASE 2 — ANALYZE

1. Capability + look first. mne_check_status; confirm the epoching inputs (events, montage,

units) before averaging.

1. Get events. mne_find_events (trigger channel) or mne_events_from_annotations

(EDF/BrainVision/EEGLAB). Verify the event codes map to the intended conditions.

1. Epoch with a justified baseline and peak-to-peak rejection, naming the conditions:

   mne_make_epochs(event_id="target:1,standard:2", tmin=-0.2, tmax=0.8,
                   baseline="default", reject_eeg=100e-6)   # (None,0); 100 µV p2p

Inspect with mne_plot_epochs_image and read the PNG (single-trial consistency, latency jitter, residual artifact).

1. Average per condition — use DISTINCT evoked names. ⚠️ mne_average_evoked defaults

evoked_name="evoked"; calling it twice overwrites the first condition. Name each one:

   mne_average_evoked(condition="target",   evoked_name="evk_target")
   mne_average_evoked(condition="standard", evoked_name="evk_standard")

1. Difference wave (the contrast carries the hypothesis), via mne_run_code:

   evk_diff = mne.combine_evoked([evk_target, evk_standard], weights=[1, -1])
   evk_diff.comment = "target - standard"

1. Plot at the PRE-SPECIFIED times, then read the PNGs: `mne_plot_evoked(name="evk_diff",

style="joint") and mne_plot_topomap(name="evk_diff", times="0.3") (your registered latency, **not** "peaks"` on this data).

1. Measure in the pre-registered window (mean is robust; peak/latency via mne_run_code):

   # MEAN amplitude in a pre-specified window over a pre-specified ROI
   roi = ["Pz", "CPz", "POz"]; tmin, tmax = 0.30, 0.50           # registered, not picked here
   ev = evk_diff.copy().pick(roi)
   m = ev.get_data(tmin=tmin, tmax=tmax).mean()                  # volts
   # PEAK amplitude + latency (note: single-subject peak latency is noisy → jackknife for groups)
   ch, lat, amp = evk_diff.get_peak(tmin=tmin, tmax=tmax, mode="abs", return_amplitude=True)
   gfp = evk_diff.data.std(axis=0)                               # global field power over time
   print(f"mean={m*1e6:.2f} µV | peak={amp*1e6:.2f} µV @ {lat*1e3:.0f} ms ({ch})")

1. Archive the equivalent code + figures (the mne-analyst archiving convention).

Best-practice reminders: high-pass ≤ 0.1–0.3 Hz + line-notch before epoching; equalize trial counts (epochs.equalize_event_counts) before comparing; report per-condition trial counts and rejection rate; measure where you pre-registered, not where the peak landed.

PHASE 3 — CRITIC (before believing the result)

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

• measurement window / peak channel chosen post-hoc (circular / double-dipping);
• component topography vs stimulus modality mismatch (mislabelled N1/P1, etc.);
• high-pass filter distorting slow-component latency/polarity;
• baseline window length/placement and propagated noise;
• unequal trial counts → SNR and differential rejection between conditions;
• multiple comparisons across components × electrodes × conditions;
• single-subject peak-latency noise (use jackknife for group latency).

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

See references/erp-methods.md for deeper recipes (event handling, difference waves, mean vs peak vs jackknife latency, GFP, component windows by modality, and ERP inference choices).