Note (2026-08-28): the Metastatic_* entries named in this project were folded into their histologic parent entries per design decisions §3a (Metastatic_Prostate_Cancer → Prostate Adenocarcinoma, Metastatic_Colorectal_Cancer → Colon Adenocarcinoma, Metastatic_HCC → Hepatocellular Carcinoma, Metastatic_Renal_Cell_Carcinoma → Renal Cell Carcinoma). Historical tables below retain the old names.

Clinical Care Guideline Collection

Task

Issue #4878 — "collect clinical care guidelines" (Melissa Haendel): "we need to prioritize other care guidelines like we did for fanconi anemia. Lets come up with a search and prioritization strategy, as well as assessment against existing gaps in HPOA file / dismech content."

The Fanconi anemia work (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 project generalizes the discovery and prioritization half of that work across the whole knowledge base: find, for each dismech disorder, the recent clinical care descriptions that could be mined for phenotype and treatment gaps.

What counts as a "clinical care description"

A PubMed citation whose Publication Type is Practice Guideline, published within the last 10 years. The Practice Guideline type is a curated NLM tag applied to society/consensus management guidelines, so it is a high-precision proxy for care descriptions — far better than free-text searching for the word "guideline". Practice Guideline is narrower than the sibling Guideline type; widening the net is a documented follow-on option (see the skill).

Search and prioritization strategy

Fully reproducible via the collect-care-guidelines Agent Skill (.claude/skills/collect-care-guidelines/). Two steps:

1. Search — for every kb/disorders/*.yaml, build a clean disease term (the top-level MONDO mapping label when present, else the name field) and run PubMed E-utilities esearch:

("<disease>"[MeSH Terms] OR "<disease>"[Title/Abstract]) AND "Practice Guideline"[Publication Type] datetype=pdat reldate=3650

Ranking disorders by hit count is the prioritization: disorders with the most recent practice guidelines float to the top.

2. Fetch — pull citation metadata (esummary) for the ranked disorders into a tab-delimited citation table.

Reliability note (why the disease phrase is field-tagged): an unadorned quoted term lets PubMed's Automatic Term Mapping shatter an unmatched name into individual all-fields words. Early testing had "Alsahan-Harris syndrome" collapse to Harris + syndrome and falsely return 30 unrelated guidelines (pelvic floor, kidney cancer, …). Tagging the phrase with [MeSH Terms]/[Title/Abstract] makes an unmatched disorder correctly return zero. This single fix is the difference between a trustworthy worklist and noise.

Results — Batch 1 (top 40 by guideline count)

Searched 1,564 disorder entries. 464 returned at least one recent practice guideline; 340 returned ≥2, 280 returned ≥3.

The full ranking is preserved in CLINICAL_CARE_GUIDELINES/guideline_search_all.jsonl (one JSON record per disorder: slug, search_name, count, pmids).

This first batch takes the top 40 disorders by recent-guideline count and exports their citation metadata to CLINICAL_CARE_GUIDELINES/guideline_citations.tsv — 1,200 citation rows (up to 30 most-recent PMIDs per disorder; the guideline_count_for_disorder column records the true total, which exceeds the sampled rows for high-volume conditions).

Results — Batch 2 (rare diseases, Fanconi-anemia-style)

Batch 1's count-ranking surfaces common conditions. Batch 2 deliberately re-slices the same 1,564-disorder search to rare diseases — the case the FA work actually targeted, and the richest ground for phenotype-annotation gaps.

Rare is defined structurally from each entry's prevalence.prevalence_class: any of RARE, ULTRA_RARE, BAND_1_9_PER_100000, BAND_1_9_PER_1000000, or BELOW_1_IN_1000000, with no COMMON / ABOVE_1_IN_1000 / BAND_1_5_PER_10000 record. Of 288 rare disorders, 48 have ≥1 recent practice guideline.

This batch picks 10 rare disorders spanning distinct disease families (blistering skin disease, vasculitis, autoimmune myopathy, hereditary cancer/polyposis, lysosomal storage, bleeding disorder, neurodegenerative ataxia, inborn error of metabolism), each in the FA-comparable "mineable" range of 3–11 guidelines. Citations → CLINICAL_CARE_GUIDELINES/guideline_citations_rare_batch.tsv (60 rows).

For continuity, Fanconi_Anemia itself (4 guidelines, same rare tier) is not re-listed here — it already has its own gap-analysis project and serves as the benchmark this batch is modeled on.

Worked example — Epidermolysis Bullosa (guidelines in practice)

Epidermolysis_Bullosa is the first entry curated from this citation set, demonstrating the end-to-end flow. Its 11 guideline hits were filtered (2 were excluded — see below), the relevant abstracts fetched with just fetch-reference, and the multidisciplinary-care gap closed with snippet-verified evidence:

• 2 new phenotypes — Esophageal Stricture (HP:0002043) and Pseudosyndactyly / mitten deformity (HP:0010554), previously only mentioned in prose.
• 6 new treatments spanning the DEBRA International care guidelines — esophageal dilatation, hand surgery + therapy, palliative/end-of-life care, podiatric foot care, neonatal EB care, and orthodontic/dental care.

NEC catch in practice: two of EB's 11 hits — the pediatric autoimmune blistering guideline (PMID:41678328) and the pemphigoid/EB-acquisita guideline (PMID:31646663) — describe autoimmune EB acquisita, a distinct entity from this Mendelian entry, and matched only on the "epidermolysis bullosa acquisita" string. They were excluded. This is the same named-entity-confusion risk the project's evidence policy warns about, caught by reading the titles.

Thin-abstract reality: the pregnancy/childbirth guideline (PMID:34687549) and the physiotherapy appraisal (PMID:35717492) have no quotable abstract body, so they could not supply snippet-verified evidence and were left for a curator with full-text access — a reminder that a guideline hit is not automatically a usable citation.

Known limitations (for the curator)

• Common-disease skew. "Search-driven, keep hits" ranking by count surfaces high-prevalence conditions with rich society guidelines. This is the intended outcome for a first batch, but the highest-value gap-mining targets may be mid-tier disorders (a few good guidelines, likely under-annotated) — the FA case itself was a rare disease. The full 464-disorder ranking supports re-slicing on that basis.
• Subtype collapse to a broad MONDO label. Several dismech entries share one umbrella search term and therefore an identical guideline set: MSI_High_Colorectal_Cancer + Metastatic_Colorectal_Cancer (both "colorectal cancer"), Hepatocellular_Carcinoma + Metastatic_HCC, Renal_Cell_Carcinoma + Metastatic_Renal_Cell_Carcinoma. The guidelines are for the parent disease, not the specific subtype. Dedupe by search_name before mining.
• Umbrella entries. Lymphoma and Infectious_Disease match a very broad MeSH term; their counts overstate specific care relevance. Infectious_Disease has since been retired as a rung-0 abstraction (design decisions §3e, issue #10115) and is no longer a dismech entry, so its row in the table above records the search only; there is no entry to mine for.
• Candidate list, not evidence. Automatic Term Mapping favors recall. Every citation is a lead — spot-check that the top hits are actually about the intended disease.

Second-generation search — "does the abstract state a recommendation?"

The count-ranked search above is a good prioritization tool but a poor evidence-sourcing tool, and the reason is worth recording: it ranks by how many Practice Guidelines exist, which reliably surfaces the flagship umbrella guideline for a disease — and those abstracts are frequently scope and process metadata (OBJECTIVE / TARGET POPULATION / EVIDENCE / METHODS, panel composition, or a chapter list) with no concrete recommendation in them. An abstract that states no specific recommendation cannot yield a snippet-verified evidence item, however authoritative the guideline is.

Scope — this is not about drugs. Care guidelines cover the whole of clinical care: pharmacotherapy is only one branch. A usable abstract is one that states a specific, actionable recommendation naming an intervention of any modality — drug, surgical/interventional procedure, radiotherapy, device, diet, rehabilitation, monitoring interval — or a diagnostic action (screening, imaging, biopsy, staging, testing). Scoring only drug names encodes a pharmacology bias and wrongly discards surgical, diagnostic and supportive-care guidance. (Worked example: a cervical-cancer screening guideline whose abstract says "screening assays should differentiate between HPV genotypes 16 and 18" is perfectly good evidence for a diagnostic recommendation and names no drug at all.)

uv run python .claude/skills/collect-care-guidelines/scripts/therapy_specific_search.py \
    spec.json out.json     # spec = [{slug, query, terms?[]}, ...]

It runs esearch, fetches each abstract, strips the citation/author/affiliation front matter, and ranks hits by recommendation_sentences — sentences carrying both an intervention/diagnostic term and a recommendation cue (we recommend, should be offered, first-line, …). The looser intervention_sentences count is reported alongside for triage. Optional per-disease terms extend the default modality vocabulary with specific drug or procedure names. Records land in CLINICAL_CARE_GUIDELINES/therapy_specific_searches.jsonl.

Query- and scoring-design lessons (each cost a round to learn):

1. Don't OR guideline*[Title] with intervention terms. It matches studies about guidelines — "Guideline adherence to aspirin prophylaxis…", "The Nationwide Impact of Guidelines for Prophylactic Aspirin…" — not guidelines themselves. Preeclampsia returned nothing but adherence/impact studies until the filter was tightened.
2. Require "Practice Guideline"[Publication Type]; put intervention terms in the scoring, not the query. Terms in [tiab] bias toward trials of that intervention over guidelines about it.
3. Require a recommendation cue, and strip the front matter — bare term matching produces two classic false positives: author affiliations ("Department of Surgery, …" — the ESMO metastatic-colorectal abstract scored 6 bogus "intervention" hits this way) and chapter/TOC listings ("1) Definition; … 5) Surgical management"), neither of which recommends anything.

Recommendation-free abstracts (negative results, recorded so they are not re-litigated). These disorders have many guidelines, but the abstracts state no specific recommendation of any modality — not drug, not procedural, not diagnostic. They need full-text access or a different source type, and should be skipped by abstract-only snippet mining:

Known scorer limitation. A recommendation cue also matches sentences that merely describe recommendations existing ("N recommendations were established…", "the guideline provides recommendations for…"). These inflate recommendation_sentences without offering anything quotable. The score is a triage signal, not a verdict — always read the sampled sentence before committing to a source.

Corollary — don't let the meta heuristic push you onto a secondary source. The limitation above has a mirror-image failure, and it bit on the very next batch after it was written down. The primary ERS bronchiectasis guideline states its own recommendation in reporting voice — "The Task Force recommendations include strong recommendations in favour of airway clearance techniques…" — which reads as meta. A Chinese-language interpretation of that same guideline phrased it more crisply ("The guideline strongly recommends airway clearance techniques…"). Scoring on phrasing alone therefore picked the commentary over the guideline itself; review caught it (#6601).

A guideline's own RECOMMENDATIONS: section frequently reports in the third person — that is the primary source speaking, not a description of someone else's work. Check journal and title for primary-vs-commentary before preferring a cleaner sentence. A bracketed, translated title ([Highlights and interpretation of …]) is a strong tell that you are reading commentary.

rec = 0 is a prompt to read, not a verdict — the scorer produces false negatives too. The two failures above inflate the score; this one suppresses it. The scorer requires both an intervention/diagnostic term and a recommendation cue, so a missing noun silently kills a real hit. Lyme_Disease scored rec=0 and was nearly recorded as a dead end, yet its guideline abstract is full of usable guidance:

"Serology is recommended only in suspected disseminated LB…"

The cue (is recommended) matched; "serology" simply wasn't in the intervention vocabulary. One missing noun, one lost guideline. The vocabulary now covers laboratory/pathology diagnostics (serology, assay, antibody, antigen, culture, histology, cytology, genotyping, sequencing) — but the vocabulary will always be incomplete, so treat rec=0 on a disease whose care you'd expect to be guideline-rich as a signal to open the abstract.

How to tell a true negative from a vocabulary gap: grep the abstract for cue language alone (recommend, should, shall, must, advise, indicated, offer, first-line). If there are no cues at all, no vocabulary could rescue it — a true negative. If cues are present but rec=0, you have a vocabulary gap. The batch-16 negatives (Sickle_Cell_Disease, Lynch_Syndrome) were verified this way and hold; they contain no directive language whatsoever.

Corollary for source selection: prefer the therapy-specific guideline over the flagship (AGA's ascites update over a general cirrhosis guideline; an appropriate-use recommendation over a disease overview). Regional and specialty-society guidelines (SEOM-GEICO, SEOM-GOTEL, AFU, ALEH, Brazilian Psychiatric Association) are frequently more snippet-usable than the big international ones, because their abstracts summarize recommendations rather than describe process.

Evidence policy

This is a discovery artifact. Any citation ultimately used as dismech evidence must still pass the standard snippet-verification workflow (just fetch-reference PMID:…, then just validate-kb-references). Guideline provenance alone does not satisfy the dismech PMID + verified-snippet policy — the same discipline applied when the FA .hpoa (evidence code TAS) terms were each independently re-sourced.

Files

• CLINICAL_CARE_GUIDELINES/guideline_citations.tsv — batch 1: the top-40 citation table (one row per disorder × guideline PMID).
• CLINICAL_CARE_GUIDELINES/guideline_citations_rare_batch.tsv — batch 2: the 10 rare-disease citation table (same columns).
• CLINICAL_CARE_GUIDELINES/guideline_search_all.jsonl — the full 1,564-disorder search result / prioritization ranking.
• CLINICAL_CARE_GUIDELINES/therapy_specific_searches.jsonl — second-generation recommendation-scored searches: one record per disorder with the query, the chosen PMID, its recommendation-sentence count, and the outcome (USED / REJECTED_NO_RECOMMENDATION). Covers enrichment batches 10–14 and records the four recommendation-free disorders so they are not re-searched.
• .claude/skills/collect-care-guidelines/ — the reusable Agent Skill (SKILL.md + scripts/collect_guidelines.py for the count-ranked search, scripts/therapy_specific_search.py for the recommendation-scoring variant).

Next steps

1. Gap assessment (the second half of #4878). For a chosen disorder, diff its guideline-derived phenotype/treatment content against (a) the dismech entry and (b) the HPOA annotation file, FA-style.
2. Mine the rest of batch 2. Epidermolysis Bullosa is done (worked example above); Kawasaki Disease has the most remaining guideline material, then work down the batch-2 table.
3. Widen the type filter to Guideline where Practice Guideline is sparse.
4. Extend the rare slice. 48 rare disorders have guidelines; batch 2 covers
5. The rest are queued in guideline_search_all.jsonl.