Neurosynth Meta-Analytic Coordinate Decoding

Statistically decode human fMRI brain coordinates against ~14,000 published neuroimaging studies, or look up which studies/coordinates are associated with a cognitive term — using the real, freely-downloadable Neurosynth coordinate database (no API key, no login).

Honesty contract (read first)

This skill downloads a real coordinate database and runs real statistics on it. It must never fabricate a term association, a study result, or a p-value.

1. Preflight before anything. Run scripts/download_data.py first. If any file fails to download, STOP and report the failure — do not proceed with partial data or describe hypothetical decoding results.
2. Never guess a term. search_by_term.py requires an exact (case-sensitive) vocabulary match. If the term isn't found, it returns suggestions (substring matches, or close spellings when there are none, e.g. hipocampus → hippocampus) — use those, don't invent a plausible-sounding term.
3. Report the multiple-testing caveat every time. decode_coordinate.py tests ~3,228 terms simultaneously per query. Uncorrected p-values will produce false positives at this scale — always surface the FDR-corrected fdr_q_value, not just the raw p_value, when characterizing how confident a result is.
4. This is a simplified reverse-inference test, not a full meta-analysis pipeline. It does binary term-presence comparison (near-coordinate studies vs. rest of corpus) via a two-proportion z-test — a real, defensible statistical test, but not the same as running NiMARE's full activation-likelihood-estimation (ALE) or multilevel kernel density analysis (MKDA) methods. Say so if precision matters (e.g., the user is drafting something for publication).
5. Never overwrite the cached data files — download_data.py writes to a dedicated cache directory and is idempotent (skips re-download if already cached, unless --force).

When to Use This Skill