Chronic Pain Disorders — Claim–Evidence Review (2026-07-25)

Correctness review of 10 chronic-pain entries in kb/disorders/, focused on claim–evidence matching: does each quoted snippet actually support the claim it is attached to, and is the supports / evidence_source classification right?

Entries reviewed

Fibromyalgia, Migraine, Chronic_Pancreatitis, Endometriosis, Irritable_Bowel_Syndrome, Osteoarthritis, erythromelalgia, Ankylosing_Spondylitis, Rheumatoid_Arthritis, Gout — 482 evidence items total.

Mechanical validation: clean

Both automated gates pass on all 10 files, so every issue below is semantic, not something CI would catch.

Check	Result
Snippet-in-source (linkml-reference-validator semantics, replicated against references_cache/)	482/482 pass — every snippet is a verbatim substring of its cached source
linkml-term-validator validate-data --labels	10/10 pass — all ontology IDs/labels correct

The anti-hallucination stack is doing its job. What it cannot see is whether a real, correctly quoted sentence is about the claim it is filed under. That is where all the defects are. (The Total checks: 0 line the reference validator prints is not evidence of a gap — see Method below.)

A. Confirmed defects (wrong, not merely thin)

A1. Rheumatoid_Arthritis: JIA is declared a subtype of RA, and the entry's own evidence says it isn't

has_subtypes[2] is JIA. All four attached evidence items fail to support it, and two explanations state the opposite outright:

• PMID:35087087 PARTIAL — "However, JIA is not a subtype of Rheumatoid Arthritis…"
• PMID:23763801 NO_EVIDENCE — "…does not explicitly define it as a subtype of Rheumatoid…"
• PMID:37700346 PARTIAL — "…does not support that JIA [is a subtype]"
• PMID:8465574 NO_EVIDENCE — snippet is about late-onset RA after age 60, entirely unrelated to JIA

Independent confirmation: MONDO does not place JIA (MONDO:0011429) under rheumatoid arthritis — its ancestors reach MONDO:0005554 rheumatic disorder, not MONDO:0008383. JIA also already has its own entry, kb/disorders/Juvenile_Idiopathic_Arthritis.yaml.

Fix: remove the JIA subtype; if a relationship is wanted, model it as a differential/related disease, not a subtype. Note the subtype name is a foreign-key target — check for subtype: JIA references before removing.

A2. Endometriosis: supports: NO_EVIDENCE on evidence that fully supports its claim

pathophysiology[4] Immune Dysfunction, PMID:16166933, is marked NO_EVIDENCE, yet the snippet is a direct, complete statement of the node's claim:

"Increased number and activation of peritoneal macrophages, decreased T cell and natural killer (NK) cell cytotoxicities are the alterations in cellular immunity and result in inadequate removal of ectopic endometrial cells from the peritoneal cavity."

and its own explanation reads "Immune cell dysfunction leads to impaired clearance of ectopic endometrial cells". Should be SUPPORT.

A3. Endometriosis: supports: SUPPORT on evidence the explanation says does not support

phenotypes[3] Infertility, PMID:16166933, is marked SUPPORT while its explanation says "Snippet supports inflammatory lesion growth mechanisms but does not directly support infertility**". The snippet is about cytokine-driven implantation and angiogenesis. Should be PARTIAL/NO_EVIDENCE, or replaced with an infertility source.

A4. Chronic_Pancreatitis: pancreatic-cancer paper cited for acinar injury, self-contradictory support value

pathophysiology[0] Recurrent Acinar Cell Injury, PMID:23622130 ("A starring role for stellate cells in the pancreatic cancer microenvironment"), supports: NO_EVIDENCE, snippet:

"Pancreatic ductal adenocarcinoma is a devastating disease, and patient outcomes have not improved in decades."

The snippet is about PDAC prognosis — nothing about acinar cells, trypsin activation, or chronic pancreatitis — yet the explanation claims it "supports the inflammatory injury cascade." The same PDAC paper carries the Pancreatic Fibrosis node (PARTIAL), where the snippet is explicitly about "pancreatic cancer development" while the explanation asserts it describes stellate-cell fibrosis "in chronic pancreatitis."

Fix: drop the NO_EVIDENCE item; replace the fibrosis citation with a chronic-pancreatitis stellate-cell source (Apte et al. and successors).

A5. Irritable_Bowel_Syndrome: prevalence quote used as evidence for microbiome dysbiosis

pathophysiology[5] Microbiome Dysbiosis, PMID:37048642, supports: SUPPORT:

"It has a prevalence of 10 to 25% in the United States and has a high disease burden…"

Epidemiology cannot support a dysbiosis mechanism, and the explanation concedes as much ("emphasizing the importance of understanding…"). The same PMID is already quoted correctly elsewhere in the file (line 109) with a sentence that names "gut microbiota composition" — that snippet belongs here.

A6. Migraine: three causative gene assignments rest on a snippet that names no gene

CACNA1A, ATP1A2, and SCN1A are each association: Causative / supports: SUPPORT, all citing the identical PMID:38508838 sentence:

"Some distinct, rare, familial migraine subtypes are caused by pathogenic variants in genes involved in ion transport and neurotransmitter release…"

No gene is named. The gene identities come from the curator's explanations, not the source. The facts are correct, but "causative gene" is the highest-stakes claim class in the KB and here it is unsourced. TRPM8 has the mirror problem: its snippet supports expression in pain circuitry, while the LRP1 entry's snippet is the one that actually names rs10166942 (2q37.1, TRPM8) as a GWAS hit.

A7. Migraine: "central sensitization" node evidenced by peripheral sensitization data

pathophysiology[2] Central Sensitization claims "sensitization of trigeminal nucleus and thalamic neurons" — i.e. second-order CNS neurons. Both evidence items are PMID:37487740, about meningeal primary afferent mechanical sensitization, which is peripheral. The explanation bridges the gap by calling afferent sensitization "a key component of central sensitization." The entry has correct central-sensitization evidence available elsewhere (PMID:30982963, cutaneous allodynia → TNC neurons).

A8. Migraine: sleep-disturbance trigger evidenced by a quote that never mentions sleep

environmental[1] Sleep Disturbance, PMID:1525797:

"One or more precipitating factor was present in 61% with MA and in 90% with MO."

A count of "some trigger exists" is used to support a specific trigger.

A9. Rheumatoid_Arthritis: 19 NO_EVIDENCE items retained, 9 with explanations that argue for the claim

This is a systematic pattern, not isolated. Examples where the snippet is about the wrong analyte or wrong disease while the explanation asserts support:

Claim	Ref	Snippet is about	Explanation says
ESR elevated	PMID:19788068	homocysteine, ADA, MDA — ESR not mentioned	"ESR … is elevated"
CRP elevated	PMID:29256110	serum substance P	"markers like CRP would typically be elevated"
Symmetric polyarthritis	PMID:11358413	HLA-DRB1 shared epitope	"thus supporting the statement"
Fatigue	PMID:26803313	fatigue in all chronic inflammatory disease	"supports … in rheumatoid arthritis"
RF positive	PMID:37475055	RF-positive JIA	"supports the presence of RF"
Morning stiffness	PMID:24461540	polymyalgia rheumatica	"…which includes RA"

Two Morning Stiffness items (PMID:30936222, PMID:25437284) have empty snippets (snippet: ''). These render as evidence rows on the disorder page.

Fix: delete NO_EVIDENCE items whose snippet is off-topic; where a paper genuinely doesn't support the claim, that belongs in notes, not an evidence block.

A10. Ankylosing_Spondylitis: an entire pathophysiology node evidenced by a five-word fragment

pathophysiology[6] TNF-Mediated Inflammation — a node asserting TNF's pivotal recruiting role and the rationale for TNF inhibitors — is supported by:

"IL-22, and tumor necrosis factor α (TNF-α)."

A mid-sentence list fragment with no subject, verb, or disease context. Two neighbours are similar: Mechanical Stress at Entheses cites "genetic predisposition, environmental factors (infections and mechanical stress), or innate and acquired immune mechanisms." for a node claiming microtrauma, cartilage-peptide release and neovascularization; the IL-23/IL-17 Axis node's first item is an abbreviation-expansion list ("PsA, Psoriasis, Psoriatic Arthritis; AS, Ankylosing Spondylitis; …").

A11. Ankylosing_Spondylitis: duplicate uveitis phenotypes with conflicting frequencies

Uveitis (FREQUENT, notes "25-40%", evidence "pooled prevalence … 25.8%") and Anterior uveitis (FREQUENT, evidence "occurs in up to 50% of the patients") are two entries for the same manifestation carrying incompatible numbers. Both also appear as separate downstream targets of the TNF node. Collapse to one.

A12. Rheumatoid_Arthritis: a duplicated interstitial-lung phenotype

Interstitial pneumonitis duplicates the existing Interstitial Lung Disease phenotype — same manifestation, two entries, both sourced from PMID:33609792.

(Correction to the first draft of this review: Myocardial infarction vs Accelerated Atherosclerosis and Pleuritis vs Pleural Effusion were also called duplicates here. On closer reading they are distinct entities — an event versus the process producing it, and serosal inflammation versus its fluid consequence — and have been kept. Only the interstitial pair was a true duplicate.)

A13. Osteoarthritis: guideline snippet truncated past the clause that carries its polarity

treatments Duloxetine, PMID:31908149:

"topical capsaicin for knee OA, acetaminophen, duloxetine, and tramadol."

The governing clause in the ACR guideline abstract is "Conditional recommendations were made for …". As quoted, the snippet cannot distinguish a recommendation from a recommendation against. Every other treatment in this file quotes the "Strong recommendations were made for…" clause in full. Extend the snippet. (Same file: this item is evidence_source: OTHER while the other five items from the same guideline are HUMAN_CLINICAL.)

B. Evidence-source misclassification (animal/in-vitro data typed as human)

evidence_source is absent on 339 of 482 items (70%), so they default to HUMAN_CLINICAL. Mostly harmless for review articles, but it is actively wrong where the snippet is explicitly non-human — and CLAUDE.md requires model-organism evidence to stay distinguishable:

File	Ref	Snippet	Should be
Migraine	PMID:37495957 (×3, incl. Headache phenotype)	"In both WT and FHM1 mutant mice, CSDs induced headache-related behaviour…"	MODEL_ORGANISM
Migraine	PMID:39080518	"Studies in rodents have demonstrated…"	MODEL_ORGANISM
Irritable_Bowel_Syndrome	PMID:17241857 (×3)	"…excite rat nociceptive visceral sensory nerves"; "Ca²⁺ in dorsal root ganglia neurons"	IN_VITRO/MODEL_ORGANISM
Fibromyalgia	PMID:21684692 (×2, incl. Chronic Widespread Pain phenotype)	"GAD65 knockout mice … supraspinal hyperalgesia"	MODEL_ORGANISM

Migraine is internally inconsistent about this: the same PMID:37495957 Panx1 snippet is correctly tagged MODEL_ORGANISM on the CSD→Trigeminovascular downstream edge but untagged on the parent node.

Conversely, Osteoarthritis PMID:40621694 tags a general conclusion sentence MODEL_ORGANISM while the sibling item from the same human single-cell abstract is HUMAN_CLINICAL.

C. Ontology bindings that undersell the claim

Term validation passes (the labels are correct), but three bindings are needlessly coarse and a better HPO term exists:

File	Phenotype	Current	Better
Migraine	Visual Aura (notes: "Scintillating scotoma, fortification spectra")	HP:0000505 Visual impairment	HP:0010822 Scintillating scotoma
Endometriosis	Painful Bowel Movements (preferred_term: Dyschezia)	HP:0002027 Abdominal pain	HP:6000222 Painful defecation
Fibromyalgia	—	Glutamate biochemical context: Insula levels on MRS cites a GAD conceptual-model paper that reports no MRS data	cite an insula-MRS study (e.g. Harris et al.) or drop the MRS context

D. Uncited claim blocks

D1. Genetics — 49 of 70 gene entries carry no evidence at all

File	genes	no evidence	no gene_term
Rheumatoid_Arthritis	23	18	23
Ankylosing_Spondylitis	16	13	16
Fibromyalgia	9	9	9
Endometriosis	4	4	4
Gout	3	3	0
Irritable_Bowel_Syndrome	2	2	2

11 genes appear in both RA and AS with byte-identical notes strings (BACH2, TNFAIP3, STAT3, IL10, CD28, EGR2, ETS1, IRF8, SATB1, IKZF1, PRDM1), all association: GWAS, none cited in either file — a copied block rather than per-disease curation.

Separately, RA misuses the association slot for function rather than gene–disease relationship: TNF → "Central proinflammatory cytokine", IL6R → "Mediates IL-6 signaling". Neither file sets relationship_type anywhere.

D2. Pathophysiology nodes with zero evidence

Rheumatoid_Arthritis has five: NF-κB Activation, Neutrophil Extracellular Trap Formation, B Cell and Plasma Cell Responses, Mucosal Origins and Dysbiosis, Epigenetic Dysregulation of T Cell Function. Endometriosis has Adhesion Formation.

D3. Frequency bands

Per docs/frequency-evidence-guidelines.md, a frequency: value is a separate quantitative claim. Most bands in this set have no supporting quote. Two are contradicted by their own evidence:

• Rheumatoid_Arthritis Cervical Spine Instability = OCCASIONAL (5–29%), evidence says RA "can compromise the cervical spine in up to 80% of the cases."
• Migraine Visual Aura = OCCASIONAL, but the file's own evidence gives aura in "one-third of patients" with visual aura in "90% of subjects with MA" → FREQUENT.

Fibromyalgia subtype shares are also mutually inconsistent: FM-CS "50–60%", FM-SFN "40–55%", pathophysiology node "approximately 50%", none cited.

D4. Missing prevalence blocks

8 of 10 entries have none (Osteoarthritis and Rheumatoid_Arthritis are the exceptions) — including Migraine and Irritable_Bowel_Syndrome, which already quote the numbers inside evidence snippets (MA 5% / MO 8% lifetime; IBS 10–25% US) but never record them structurally.

E. Module-conformance gap

kb/modules/cellular_senescence.yaml (line 34) and CLAUDE.md both advertise "Worked conformers: Osteoarthritis (senescent chondrocytes; PMID:28436958)", but kb/disorders/Osteoarthritis.yaml has no conforms_to on its Chondrocyte Senescence node — its only conformance is to osteoarthritis_cartilage_degradation. The documentation overstates the KB. Either add conforms_to: "cellular_senescence#Senescent Cell Accumulation" or correct the module docs.

Osteoarthritis also under-links its module: Subchondral Bone Remodeling and the catabolic chondrocyte state map onto osteoarthritis_cartilage_degradation nodes that are left unconnected.

Per-entry summary

Entry	Evidence items	Verdict
Gout	40	Strongest. Clean causal chain, reference ranges with interpretation bands, computational model. Gaps: Hyperuricemia and Inflammasome Activation nodes have no snippet naming urate elevation or NLRP3/IL-1β respectively; Nephrolithiasis edge and phenotype uncited.
Endometriosis	33	Strong. Best pain-mechanism modelling in the set (neuroangiogenesis → peripheral → central sensitization, with a hypothesis group and a negative P2X3 trial cited as a qualifier). Two support-value errors (A2, A3).
erythromelalgia	15	Clean. Every item verifiable and appropriately hedged. Marking the 2004 "single common pathogenetic mechanism — microvascular arteriovenous shunting" claim SUPPORT is generous given Nav1.7; PARTIAL fits better.
Osteoarthritis	26	Strong. Only file with evidence_source on 100% of items. Issues: A13, E.
Migraine	72	Good mechanism, weak genetics. Issues: A6, A7, A8, B, C, D3, D4.
Irritable_Bowel_Syndrome	33	Good. Issues: A5, B. Also geo:GSE36701 is titled "…rectal mucosa…" but described and tissue-bound as jejunum (UBERON:0002115) — the jejunal study is GSE14841. Both microbiome datasets set organism.preferred_term: human gut metagenome against NCBITaxon:9606, and type 16S amplicon runs as data_type: WGS.
Fibromyalgia	25	Mixed. Descending Pain Modulation Dysfunction claims reduced serotonin/norepinephrine but both citations are about glutamate/GABA; Serotonin: Decreased is uncited. Plus B, C, D1, D3.
Chronic_Pancreatitis	32	Mixed. A4 is the main defect. Two biological_processes (Extracellular Matrix Organization, Digestive System Process) have no term: binding though GO:0030198/GO:0022600 exist. Explanations cite numbers absent from their snippets ("35-62% prevalence", "68.9% osteopenia", "diabetes HR 2.3"). PRSS1 is Causative on a snippet that says "associated with the risk of".
Ankylosing_Spondylitis	40	Weak sourcing. A10, A11, D1. HLA-B27 … Present in 90-95% of patients is uncited. No prevalence block despite PMID:32712723 (already cached for RA) giving AS at 0.20–0.25%.
Rheumatoid_Arthritis	166	Most content, most defects. A1 (nosology), A9 (19 NO_EVIDENCE, 2 empty snippets), A12, D1, D2. Breadth is excellent; evidence hygiene has not kept up with it.

Recommended fix order

1. A1 — remove the JIA subtype from RA (nosologically wrong, self-contradicted, MONDO-contradicted).
2. A2, A3, A4, A5 — support-value errors and wrong-disease citations; each is a 1–5 line edit.
3. A9 — sweep RA's 19 NO_EVIDENCE items: delete the off-topic ones, move real caveats to notes, and fix the two empty snippets.
4. A6, A7, A8, A13 — re-source or re-quote; all have a correct source already present in the same file or one just fetch-reference away.
5. B — set evidence_source: MODEL_ORGANISM / IN_VITRO on the ~10 explicitly non-human snippets.
6. C, E — swap the two HPO bindings; resolve the cellular_senescence conformance claim.
7. D — larger curation projects: cite or drop the uncited gene blocks, add the missing prevalence records, audit frequency bands.

Method

• Snippet fidelity was checked by replicating SupportingTextValidator.normalize_text and _split_query from the installed linkml-reference-validator against references_cache/ (punctuation-stripping, ...-splitting, order-independent substring match) to get a per-snippet accounting. Note that linkml-reference-validator validate data reports "Total checks: 0" on a clean file because that counter reports issues found, not snippets examined — the affirmative signal is the Snippets checked: N/N verified against cached references line, available standalone via just count-verified-snippets <file> (see CLAUDE.md, "Reading the reference-validation summary", and issue #7252). An earlier draft of this report read the zero as evidence that validation was running vacuously; that was wrong, and no validator change is needed.
• Term validation used scripts/run_term_validator.sh validate-data … --labels.
• Alternative HPO terms were located with runoak -i sqlite:obo:hp search; MONDO placement with runoak -i sqlite:obo:mondo ancestors.
• Claim–evidence matching was done by reading each entry in full.

Status: fixes applied

All findings in sections A–C and E were applied in the follow-up commit on this branch, plus the two contradicted frequency bands and the three prevalence blocks whose numbers were already present in cached snippets (D3, D4 partial). Section D1/D2 (uncited gene blocks and evidence-free pathophysiology nodes) remains open — those need per-disease literature curation rather than an edit.

Subsequent review rounds on PR #7476 corrected three things in the fixes themselves: the evidence_source: OTHER tags on the new prevalence records were retagged HUMAN_CLINICAL (pooled human epidemiology is clinical evidence regardless of publication format); curation meta-commentary was removed from user-facing description/notes/context fields, which render on the public disorder pages; and the mechanistic biology this review had deleted from the Ankylosing_Spondylitis "Mechanical Stress at Entheses" node was restored with a real supporting citation rather than left excised — a better resolution than removal. Part of D1 is now closed: the five Migraine gene entries carry HGNC gene_term bindings, and the mixed human/mouse Fibromyalgia snippet was split so each item carries one evidence_source.

Post-fix validation: schema 10/10 No issues found; term validation 10/10 passed; full tests/test_data.py suite 16,782 passed; every snippet verifying via just count-verified-snippets; reference-cache frontmatter contract OK. Three references were newly fetched via the project's fetch path (PMID:15753610, PMID:32430436, PMID:33547227); no cache file was hand-edited.