Collect Clinical Care Guidelines Skill

Overview

This skill finds the clinical care guidelines that describe the management of a dismech disorder, so their phenotype and treatment content can be mined against the existing knowledge-base entry (the workflow first used for the Fanconi anemia gap analysis, issue #4878).

A "clinical care description" here is operationalized as a PubMed citation whose Publication Type is Practice Guideline, published within the last 10 years (configurable). PubMed's Practice Guideline publication type is a curated NLM tag applied to society/consensus management guidelines, which makes it a high-precision filter for care descriptions — far better than free-text searching for the word "guideline".

The output is a tab-delimited citation table (one row per disorder × PMID) plus a JSONL discovery record that doubles as a prioritization ranking.

When to use

• Building an initial care-guideline citation set for a batch of disorders.
• Refreshing an existing set with newly published guidelines.
• Prioritizing which disorders have the richest recent guideline evidence to mine for phenotype/treatment gaps.

The workflow (two steps)

The helper script scripts/collect_guidelines.py uses NCBI E-utilities (esearch + esummary) with only the standard library plus PyYAML. Run it from the repo root.

Step 1 — search

For each disorder, it builds a clean search term (the top-level MONDO mapping label when present, else the name field with underscores as spaces) and runs:

esearch  db=pubmed
         term=("<disease name>"[MeSH Terms] OR "<disease name>"[Title/Abstract])
               AND "Practice Guideline"[Publication Type]
         datetype=pdat  reldate=3650   # last 10 years

The disease phrase must be field-tagged ([MeSH Terms] / [Title/Abstract]). A bare quoted phrase lets PubMed's Automatic Term Mapping shatter an unmatched name into individual all-fields words — e.g. "Alsahan-Harris syndrome" collapses to Harris + syndrome and falsely returns dozens of unrelated guidelines. Field-tagging makes an unmatched disorder correctly return zero.

# All disorders (≈1,560; ~9 min without an API key at 3 req/s)
python .claude/skills/collect-care-guidelines/scripts/collect_guidelines.py \
    search --out hits.jsonl

# Or a named subset
python .claude/skills/collect-care-guidelines/scripts/collect_guidelines.py \
    search --slugs Asthma,Marfan_Syndrome,Fanconi_Anemia --out hits.jsonl

Each JSONL line is {"slug", "search_name", "count", "pmids"}. Sorting by count descending is the prioritization ranking — disorders with the most recent practice guidelines float to the top.

Step 2 — fetch citation metadata → TSV

# Keep disorders with ≥1 guideline; take the top 40 by guideline count
python .claude/skills/collect-care-guidelines/scripts/collect_guidelines.py \
    fetch --hits hits.jsonl --out citations.tsv --min-count 1 --top 40

The TSV columns are:

column	meaning
disorder_slug	kb/disorders/<slug>.yaml stem
search_name	the disease term sent to PubMed
pmid	guideline citation PMID
title	article title
journal	journal abbreviation
pub_year	publication year
pubdate	full PubMed pubdate string
doi	DOI when present
publication_types	all NLM publication types (should include Practice Guideline)
guideline_count_for_disorder	how many guidelines that disorder returned

Tuning knobs

• --reldate <days> — lookback window (default 3650 = 10 years).
• --retmax <n> — max PMIDs captured per disorder (default 30).
• --top <n> / --min-count <n> — how many disorders to keep in the TSV.
• NCBI_API_KEY env var — raises the rate limit from 3 to 10 req/s.

Re-slicing by rarity (or any disorder attribute)

search scans all disorders and records the count per disorder, so any slice is a post-filter on the JSONL — no re-search needed. To build a rare-disease batch (the Fanconi-anemia-style target), filter to entries whose prevalence.prevalence_class is one of RARE, ULTRA_RARE, BAND_1_9_PER_100000, BAND_1_9_PER_1000000, BELOW_1_IN_1000000 (and not COMMON / ABOVE_1_IN_1000), intersect with count > 0, then feed the filtered JSONL to fetch:

# after producing hits.jsonl, keep only rare slugs (whatever your filter yields)
grep -F -f rare_slugs.txt hits.jsonl > hits_rare.jsonl   # or filter in Python
python .claude/skills/collect-care-guidelines/scripts/collect_guidelines.py \
    fetch --hits hits_rare.jsonl --out rare_citations.tsv --min-count 1 --top 10

Precision notes and caveats

• The bare "<name>" term relies on PubMed Automatic Term Mapping, which expands to the MeSH disease + all-fields text. This favors recall; a returned guideline is about the disease in the vast majority of cases, but the set is a candidate list for human curation, not final evidence.
• The Practice Guideline publication type is narrower than Guideline. Switch the PUBTYPE constant in the script if you want to widen to Guideline.
• Disorders whose name/MONDO label is a broad parent (e.g. "Glioblastoma, IDH-Wildtype") may under- or over-match; spot-check the top hits.
• This is a discovery step. Any citation ultimately used as dismech evidence must still pass the normal snippet-verification workflow (just fetch-reference, just validate-kb-references) — the .hpoa/guideline provenance alone does not satisfy the dismech PMID + verified-snippet policy.

Relationship to the Fanconi anemia precedent

The FA gap analysis (projects/FANCONI_ANEMIA_GAP_ANALYSIS.md) mined a single disorder's care guideline into a custom HPO profile and diffed it against the dismech entry. This skill generalizes the discovery half of that work across all disorders so the same mining can be prioritized and repeated at scale. See projects/CLINICAL_CARE_GUIDELINES/ for the tracked citation set.