MNE Time-Frequency Analysis (grill → analyze → critic)

Time-frequency analysis of neurophysiology data via the MNE MCP server. This skill is skeptical by design: most time-frequency mistakes (edge artifacts read as real low-frequency effects, an unstated or contaminated baseline, evoked-vs-induced confusion) run without any error — so the discipline is to grill before computing and critique before believing.

Companion skills: mne-mcp-guard for technical execution safety (wavelet-length errors); mne-methodology-critic for Phase 3. Loaded objects persist in one MNE session.

PHASE 1 — GRILL (before computing anything)

Do not compute a TFR 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, post-stimulus × baseline)

• Within- or between-subject? Paired or independent? n per cell?
• Confirmatory (hypothesis + time-freq window/ROI pre-specified) or exploratory (whole map,

corrected)?

The two questions that decide validity

• Power or inter-trial coherence (ITC)? Power = spectral magnitude over time; ITC = phase

consistency across trials (0–1). They answer different questions — a stimulus can drive ITC with little power change (and vice versa). State which, or both.

• Evoked or induced (total) power? ⚠️ Power of the average (evoked) captures only

phase-locked activity; per-trial power then averaged (total/induced) also captures non-phase-locked oscillations. These are different claims; mixing them (e.g. computing total power but interpreting it as the evoked response) is the most common fatal error here.

Data & parameters

• Frequencies of interest, and is the LOWEST freq resolvable given epoch length? A Morlet

wavelet at f Hz with n_cycles cycles has half-length ≈ n_cycles / (2f) s — the epoch must extend that far beyond every time point you interpret, or you read edge artifacts.

• n_cycles choice (default freqs/2) and its time ↔ frequency resolution tradeoff: fewer

cycles = better time, worse frequency resolution (and shorter wavelet → less edge contamination).

• Baseline window + normalization TYPE (logratio / zscore / percent / mean) — and is the

baseline clean (no spillover from the previous trial, no anticipatory activity)?

• Method: Morlet, multitaper (DPSS smoothing), or Stockwell (S-transform)?
• Reference, channel selection, decimation, units.

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

• Time-freq window / ROI: pre-registered or data-driven? (data-driven = double-dipping)
• Multiple comparisons across time × frequency × channels — which correction? Per-pixel tests

are massively multiple and correlated → plan cluster-based permutation in time-frequency.

• Small n (≲ 15–20)? Then plan permutation / non-parametric, not a normality assertion.

PHASE 2 — ANALYZE

1. Capability + look first. mne_check_status; mne_describe("epochs") to read the epoch

window. Check the epoch is wide enough: for the lowest frequency, you need ≈ n_cycles/(2·fmin) s of pad on each side of the interpretable window. If not, re-epoch wider (e.g. -0.5 to 1.0+ s) from raw before computing.

1. Compute Morlet power (+ ITC). Quick path: mne_tfr_morlet(fmin=, fmax=, n_freqs=). For ITC

and full control use mne_run_code:

   import numpy as np
   freqs = np.linspace(4, 40, 20)
   n_cycles = freqs / 2.0                       # default tradeoff; raise for fine freq resolution
   power, itc = mne.time_frequency.tfr_morlet(
       epochs, freqs=freqs, n_cycles=n_cycles,
       use_fft=True, return_itc=True, average=True)   # average=True ⇒ total/induced power + ITC

(Multitaper: tfr_multitaper(..., time_bandwidth=4.0). Stockwell: tfr_stockwell(epochs, fmin=4, fmax=40). Evoked power: compute on epochs.average() instead of per-trial.)

1. Baseline-normalize — state the type. Apply on plot or in code; never leave it implicit:

   power.plot([0], baseline=(-0.5, 0), mode="logratio")   # or: zscore / percent / mean

logratio/percent = relative to baseline; zscore = SD units. Baseline must sit in the clean pre-stimulus window, clear of edge and anticipatory activity.

1. Plot and read the PNG. Inspect the time-frequency map; identify where the edge zone lies

(the cone of influence widens at low frequencies) and explicitly avoid interpreting it.

1. Induced vs evoked, explicitly. If the claim is about non-phase-locked oscillations, confirm

you used per-trial power (average=True over single-trial TFRs), not power of the evoked average.

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

Best-practice reminders: epoch wide then crop the display; report n_cycles, baseline window + mode, and method; mark the edge/cone region; prefer a time-freq window that was pre-specified.

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 time-frequency work it will specifically check:

• edge / wavelet-length effects read as real low-frequency effects (too-short epoch);
• baseline normalization unstated, or baseline window contaminated;
• evoked vs induced confusion (power of average vs per-trial power averaged);
• n_cycles time/freq-resolution tradeoff unjustified;
• double-dipping on a data-driven time-freq window / ROI;
• multiple comparisons across time × freq × channels (recommend cluster-based permutation);
• small-n normality asserted vs tested.

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

See references/timefreq-methods.md for deeper recipes (Morlet vs multitaper vs Stockwell, ITC, baseline modes, the n_cycles/edge tradeoff, and time-frequency inference choices).