mne-timefreq

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.