Neuroimmune Entry Review: Claim–Evidence Correctness Audit

Date: 2026-07-25 Scope: 10 neuroimmune disorder entries in kb/disorders/ Focus: correctness of claim↔evidence matches (does the cited snippet actually support the claim it is attached to?) Status: all findings below have been fixed — see Remediation for what changed. Findings are kept in their as-discovered form as the record of why.

Entries reviewed

Entry	Lines	Evidence items	Verdict
`Multiple_Sclerosis`	1576	156	Needs work — 14 non-supporting items, 1 retired nosology claim
`Stiff_Person_Syndrome`	1599	95	Minor — 1 unsupported + internally inconsistent genetic claim
`Guillain_Barre_Syndrome`	795	37	Minor — 1 materially reversed claim–evidence match
`Chronic_Inflammatory_Demyelinating_Polyneuropathy`	467	20	Minor — 1 meaning-altering truncation, 1 source misclass
`Neurosarcoidosis`	1668	67	Minor — 2 content-free snippets used as SUPPORT; module gap
`Neuromyelitis_Optica_Spectrum_Disorder_with_Anti-AQP4_Antibodies`	978	40	Minor — MOA-as-mechanism, 1 overread sequence claim
`Myasthenia_Gravis`	1234	72	Good — 1 `evidence_source` misclass
`Anti-NMDA_Receptor_Encephalitis`	1316	72	Good
`Acute_Disseminated_Encephalomyelitis`	2046	111	Good
`Rasmussen_Encephalitis`	390	7	Structurally sound but thin (4/7 pathophysiology nodes uncited)

Method

Three passes:

Snippet-substring check against references_cache/, ellipsis-aware and whitespace-collapsing.

Correction (post-review). The first pass of this check also folded Unicode punctuation (curly→straight quotes, en/em dash→hyphen) and case. linkml-reference-validator does not — it requires an exact substring. That leniency masked two real failures, one of which this audit itself introduced (PMID:32404428, Neurosarcoidosis: ASCII "definite" vs the source's “definite”; caught by ai4c-reviewer on PR #6962) and one pre-existing (PMID:28757204, MS Vitamin D: snippet capitalized We showed where the source reads we showed). Both are now fixed, and the numbers below come from a strict re-check that collapses whitespace only. Repo convention is to preserve curly quotes verbatim in snippets rather than substitute ASCII equivalents. 2. Automated semantic checks: reference_title drift vs cached title (wrong-PMID detector), supports: enum vs explanation wording (self-contradiction detector), and snippet species/setting vs evidence_source. 3. Manual read of every pathophysiology node and its evidence in all 10 entries, plus targeted reads of flagged sections.

Note on tooling: linkml-reference-validator reports Total checks: 0 in a network-restricted environment (it fails to fetch full text and silently performs no snippet checks). Do not treat a passing run in that setting as evidence that snippets were verified.

Findings

Ordered most to least severe.

1. MS: "Progressive-Relapsing" course is retired nosology with zero supporting evidence

kb/disorders/Multiple_Sclerosis.yaml — progression[3] (phase: Progressive-Relapsing) carries five evidence items, and all five are tagged supports: NO_EVIDENCE with explanations that explicitly say the reference does not address PRMS:

PMID:37068931 — "does not mention progressive-relapsing multiple sclerosis (PRMS)"
PMID:17884680 — "does not address progressive-relapsing multiple sclerosis (PRMS)"
PMID:31397221 — "does not mention progressive-relapsing multiple sclerosis"
PMID:11207871 — "does not specifically address progressive-relapsing multiple sclerosis (PRMS)"
PMID:24722325 — "does not specifically mention progressive-relapsing multiple sclerosis"

Separately, PRMS was eliminated as a clinical course descriptor by the 2013 Lublin revision (Lublin et al., Neurology 2014;83:278–286); those patients are now classified as PPMS with activity. So the entry asserts a category that is both retired and unsupported by anything it cites.

Recommendation: remove the Progressive-Relapsing progression phase (and its five non-evidence items), or retain it only as an explicitly historical descriptor citing the Lublin revision that retired it.

2. GBS: complement evidence cites a study whose headline result is negative

kb/disorders/Guillain_Barre_Syndrome.yaml:215 — pathophysiology node "Complement-Mediated Nerve Damage", PMID:10355667.

The cited paper is titled "Anti-ganglioside antibodies can bind peripheral nerve nodes of Ranvier and activate the complement cascade without inducing acute conduction block in vitro". Its conclusion is that the node of Ranvier is "relatively resistant to acute antiganglioside antibody mediated injury" and that "this in vitro sciatic nerve model appears of limited use".

The dismech explanation reads: "…establishing the complement-mediated mechanism of nerve damage." The paper establishes complement fixation and explicitly fails to demonstrate nerve damage — the explanation inverts the study's conclusion. supports: PARTIAL is the right tag; the explanation is not.

Compounding this: the study is mouse sciatic nerve, ex vivo, but evidence_source is unset (defaults to HUMAN_CLINICAL). Should be MODEL_ORGANISM or IN_VITRO.

Recommendation: rewrite the explanation to state what the study actually shows (antibody binding + complement fixation at nodes of Ranvier, without acute electrophysiological deterioration), and set evidence_source.

3. MS: 14 `NO_EVIDENCE` items, 5 with explanations that claim the opposite

Multiple_Sclerosis.yaml is the only entry of the 10 carrying a large NO_EVIDENCE population (14; SPS has 2, the other eight have none). Five pair supports: NO_EVIDENCE with an explanation asserting support — a direct internal contradiction:

Line	Reference	Location	Explanation says
591	`PMID:23153835`	phenotypes > Gait and Balance Issues	"This reference supports the statement…"
998	`PMID:22919874`	biochemical > Oligoclonal Bands	"This reference supports the detection of oligoclonal bands…"
1448	`PMID:15575796`	treatments > Symptomatic Treatments	"This reference supports the statement…"
1515	`PMID:10073279`	treatments > Corticosteroids	"…indirectly supports the use of other treatments like corticosteroids"
620	`PMID:10101582`	phenotypes > Muscle Weakness	"…supporting the frequency of muscle-related issues"

PMID:10073279 is the worst: an interferon beta efficacy paper attached to the Corticosteroids treatment, justified as "indirectly supports the use of other treatments". An interferon paper is not evidence about corticosteroids.

The remaining nine NO_EVIDENCE items are honestly labelled (their explanations correctly say the reference does not support the claim) but add noise without adding evidence.

Recommendation: delete these items rather than re-tagging them. Per CLAUDE.md §"When Evidence Cannot Be Verified", the fix for an unsupportable claim is to move it to notes or drop the evidence block — not to keep a citation that does not support it.

4. Neurosarcoidosis: content-free snippets used as `SUPPORT`

kb/disorders/Neurosarcoidosis.yaml — PMID:30167654 (Neurosarcoidosis Consortium consensus criteria) is cited seven times. Two of those are quoting sentences that carry no claim content:

Pathophysiology "Nervous System Granulomatous Inflammation" — snippet: "Diverse disease presentations and lack of specificity of relevant diagnostic tests contribute to diagnostic uncertainty." This is about diagnostic uncertainty; it says nothing about granulomatous inflammation of neural tissue, which is the node's claim.
Pathophysiology "Neurologic Dysfunction" — snippet: "The work of this collaboration included a review of the manifestations of neurosarcoidosis and the establishment of an approach to the diagnosis of this disorder." This is a scope sentence about the paper itself, not a finding. It cannot support a node claiming cranial neuropathy / visual pathway / meningeal / peripheral nerve involvement.

Both are tagged SUPPORT. The same PMID is used correctly elsewhere in the file (e.g. the DIAGNOSTIC_CRITERIA definition at line 127), so this is snippet selection, not a wrong PMID.

Additional gap: Neurosarcoidosis declares no conforms_to, despite kb/modules/granuloma_formation.yaml naming sarcoidosis as a target conformer. Its pathophysiology (antigen presentation → Th1/Th17 → granulomatous inflammation) maps directly onto the module chain.

5. CIDP: truncation that changes a diagnostic definition

Chronic_Inflammatory_Demyelinating_Polyneuropathy.yaml:177 — PMID:38330421, phenotypes > Sensory Loss.

Snippet: "Sensory CIDP was diagnosed when two inclusion criteria are met: 1) acquired, chronic progressive or relapsing symmetrical or asymmetrical sensory polyneuropathy that had progressed for >2 months."

Abstract: "…progressed for >2 months; and 2) definite electrophysiological and/or biopsy evidence of demyelinating neuropathy."

The snippet announces "two inclusion criteria" and then supplies only one, replacing the semicolon with a period. A reader would take the clinical criterion as sufficient when the source requires electrophysiological/biopsy confirmation. This also fails substring validation.

Also at line 111 (PMID:36645654, Macrophage-Mediated Myelin Stripping): snippet truncated at "…downregulation of macrophage activation." where the abstract reads "…downregulation of macrophage activation or co-stimulatory and adhesion molecules." Milder, but the truncation makes macrophage downregulation look like the terminal mechanism — precisely the claim being supported.

At line 74 (PMID:36346134): snippet says "cytotoxic effects in vitro" but evidence_source is unset → should be IN_VITRO.

6. NMOSD: drug design intent used as mechanistic evidence

Neuromyelitis_Optica_Spectrum_Disorder_with_Anti-AQP4_Antibodies.yaml — pathophysiology node "AQP4-Reactive B Cell Autoantibody Production":

PMID:36933107 snippet: "which is designed to suppress autoantibody production by blocking the interleukin-6 (IL-6) receptor" — explanation calls this "directly supports autoantibody production as an upstream therapeutic target". A statement of what a drug is designed to do is not evidence that the mechanism operates. Reasonable as PARTIAL with a hedged explanation; not "directly supports".
PMID:31495497 snippet is a mid-sentence fragment ("of inebilizumab, an anti-CD19…") — a trial-efficacy result used to infer an upstream mechanism.

Separately, node "Secondary Demyelination and Neuronal Injury" cites DOI:10.4103/nrr.nrr-d-23-01325 with the explanation "Supports the downstream sequence from astrocytopathy to demyelination and neuronal loss." The snippet is an unordered list of features that animal models reproduce ("aquaporin-4 loss, astrocytopathy, granulocyte and macrophage infiltration, complement activation, demyelination, and neuronal loss") — it carries no causal ordering. The MODEL_ORGANISM tagging here is correct.

7. SPS: HLA claim unsupported and names the wrong gene

Stiff_Person_Syndrome.yaml — genetic > HLA-DRB1:

notes claims "The DQB1*0201 allele is present in approximately 70% of SPS patients" — but the entry is keyed to HLA-DRB1. DQB1 is a different gene (HLA-DQB1). Gene/allele mismatch within one record.
Its only evidence item (PMID:35084720) is tagged NO_EVIDENCE, and the snippet is about clinical overlap across GAD-spectrum disorders — it mentions no HLA allele. The explanation concedes the inference: "…implies common genetic susceptibility factors including HLA associations." Inference, not evidence.

Recommendation: either cite a real HLA association study for the 70% DQB1*0201 figure and re-key the record to HLA-DQB1, or move the claim to notes without an evidence block.

8. `evidence_source` misclassifications (4 total)

All four are non-human evidence defaulting to HUMAN_CLINICAL:

File	Line	Reference	Snippet setting	Should be
`Guillain_Barre_Syndrome`	215	`PMID:10355667`	mouse sciatic nerve, "in vitro"	`MODEL_ORGANISM`/`IN_VITRO`
`Myasthenia_Gravis`	256	`PMID:29266249`	"Studies in animals… EAMG"	`MODEL_ORGANISM`
`Multiple_Sclerosis`	1127	`PMID:35963325`	"preclinical MS model… (EAE)"	`MODEL_ORGANISM`
`Chronic_Inflammatory_Demyelinating_Polyneuropathy`	74	`PMID:36346134`	"cytotoxic effects in vitro"	`IN_VITRO`

Per CLAUDE.md, model-organism evidence should not be the sole support for human phenotype claims — untagged, these read as human clinical evidence. Note the MG and MS cases are the only animal-derived support for their respective nodes' specific sub-claims.

9. Terminal-punctuation truncations (cosmetic, 12 items)

Twelve snippets fail substring validation only because the curator truncated mid-sentence and appended a period where the source has a comma or semicolon. Examples: PMID:32388832 (×2, MS), PMID:24314688 (also swaps the source's curly double quotes around "multiple" for straight single quotes), PMID:32560364, PMID:37059571, PMID:31971066, PMID:32408148, PMID:29452342 (MS); PMID:37869140 (MG); PMID:37108447 (GBS); PMID:36645654, PMID:38330421 (CIDP).

Ten are harmless. Two — PMID:38330421 and PMID:36645654, both CIDP — change meaning and are written up separately in finding 5.

10. Weak-attribution and coverage notes (no action required)

MS prevalence, Europe (PMID:37059571): a review of Chinese and Asian MS epidemiology is cited for a European prevalence of 115/100,000. The figure appears in the abstract, but as a comparator for "countries with predominantly white populations" — not a European estimate. A primary European epidemiological source would be better.
MS oligoclonal bands (PMID:32408148): a 6-patient Baló's concentric sclerosis series used for OCB in MS. Correctly tagged PARTIAL.
Rasmussen encephalitis: 4 of 7 pathophysiology nodes (Microglial Activation, Cortical Hyperexcitability, Drug-Resistant Focal Seizures, Progressive Neurological Decline) carry no evidence at all. Nothing incorrect — just thin. The two conforms_to declarations against epilepsy_excitation_inhibition_imbalance are correctly formed.
CIDP declares no conforms_to despite peripheral_axonal_degeneration#Distal Axonal Degeneration and Demyelination being an obvious fit.

What checked out clean

No wrong-PMID / named-entity-confusion errors. Every reference_title in all 10 entries matches its cached PubMed title. No fabricated PMIDs: the two high-numbered 2026 references (PMID:42093930 anti-GQ1b spectrum review, PMID:41750202 ADEM review) are both real and correctly quoted.
Anti-NMDA_Receptor_Encephalitis — the strongest of the ten. Clean separation of IN_VITRO (Dalmau rat hippocampal culture work) from HUMAN_CLINICAL (paired CSF/serum), and every snippet is a faithful quote carrying the claim it supports.
Acute_Disseminated_Encephalomyelitis — 111 evidence items, zero substring failures, appropriate PARTIAL/OTHER tagging, and mechanistic nodes correctly framed as proposed hypotheses rather than established fact.
Myasthenia_Gravis, Neuromyelitis_Optica_…AQP4, Neurosarcoidosis — 100% SUPPORT with no NO_EVIDENCE padding.

Remediation

All findings were fixed in the same branch. Six entries changed; the four clean entries (Anti-NMDA_Receptor_Encephalitis, Acute_Disseminated_Encephalomyelitis, Neuromyelitis_Optica_Spectrum_Disorder_with_Anti-AQP4_Antibodies, Rasmussen_Encephalitis) were left untouched. A history record accompanies each changed entry under history/disorders/<SLUG>/.

#	Finding	Fix
1	MS `Progressive-Relapsing` phase	Phase removed with its five `NO_EVIDENCE` items. Historical context folded into the `Primary Progressive` notes and backed by two verified quotes from `PMID:24871874` (the 2013 Lublin revision), newly fetched into the cache.
2	GBS `PMID:10355667` inverted explanation	Explanation rewritten to state what the study shows (binding + complement fixation) and what it explicitly does not (nerve injury; no electrophysiological deterioration over 4–6 h). `supports: PARTIAL` kept; `evidence_source: MODEL_ORGANISM` added.
3	MS's 14 `NO_EVIDENCE` items	All removed (5 with the PRMS phase, 9 individually). No claim lost its last support — every affected block retains verified `SUPPORT`/`PARTIAL` evidence, checked programmatically for orphaned `evidence:` lists.
4	Neurosarcoidosis content-free snippets	Both replaced with substantive quotes from `PMID:32404428` (definite-NS neural pathology requirement; the enumerated manifestation list). `PMID:30167654` retained on both nodes, downgraded to `PARTIAL` with explanations stating what it does and does not establish.
5	CIDP `PMID:38330421` truncation	Second inclusion criterion restored. Also restored the truncated `PMID:36645654` clause ("…or co-stimulatory and adhesion molecules") and rewrote its explanation to stop presenting macrophage downregulation as the terminal IVIg mechanism.
6	NMOSD MOA-as-mechanism	Not changed. On re-reading, `supports: SUPPORT` on a drug's stated design rationale is defensible for an "upstream therapeutic target" claim, and the `MODEL_ORGANISM` tagging was already correct. Left as a style note rather than a defect.
7	SPS HLA-DRB1 / DQB1*0201	Record re-keyed to HLA-DQB1 with `gene_term` `hgnc:4944` (OAK-verified). The `NO_EVIDENCE` item was replaced by `PMID:8263140` (Pugliese 1993 — the actual source of the ~70% figure) and `PMID:32152690` (four-digit typing confirming a primary DQ effect), both newly fetched. SPS's other `NO_EVIDENCE` item (Stiff Limb Syndrome subtype) was likewise replaced, with the phenotype-defining quote from `DOI:10.1007/s00415-023-12123-0`, already cited elsewhere in the entry.
8	Four `evidence_source` misclassifications	All set: `MODEL_ORGANISM` (GBS `PMID:10355667`, MG `PMID:29266249`, MS `PMID:35963325`) and `IN_VITRO` (CIDP `PMID:36346134`).
9	12 terminal-punctuation truncations	All repaired; the two meaning-altering ones are covered by finding 5. `PMID:24314688`'s straight-single-quote damage restored to the source's double quotes.
10	Module conformance gaps	`conforms_to` added: Neurosarcoidosis → `granuloma_formation#Th1 and TNF-Driven Macrophage Recruitment and Activation` and `#Organized Granuloma Assembly`; CIDP → `peripheral_axonal_degeneration#Distal Axonal Degeneration and Demyelination`. All references verified to resolve.

Left as noted-but-unchanged: the MS Europe-prevalence attribution and Baló-series OCB citation (finding 10 of the original list) are weak but not incorrect, and Rasmussen's uncited pathophysiology nodes are a coverage gap rather than an error — filling them is new curation, not a correctness fix.

Verification after remediation

Across all 10 entries, 668 evidence items:

Snippet substring check (strict — whitespace-collapse only): 0 failures (was 12 under the original check, plus the 2 that its Unicode/case folding masked; see the correction under Method).
Semantic audit: 0 supports/explanation contradictions (was 5), 0 species/evidence_source mismatches (was 4), 0 title drift.
NO_EVIDENCE items: 0 (was 16).
linkml-validate -C Disease: passes on all six changed entries.
linkml-term-validator --labels: passes on the entries with new/changed terms.
linkml-validate -C HistoryRecord: passes on all six new history records.
pytest tests/test_data.py: the one failure (test_evidence_items_have_references[Babesiosis.yaml], a bare CDC URL reference) is pre-existing and unrelated — Babesiosis.yaml is untouched here and was last modified by PR #6857.

Cross-cutting observation

Defects cluster almost entirely in the older, larger entries (Multiple_Sclerosis, and to a lesser degree Stiff_Person_Syndrome, Chronic_Inflammatory_Demyelinating_Polyneuropathy). The recently curated entries (Anti-NMDA_Receptor_Encephalitis, Acute_Disseminated_Encephalomyelitis, Neuromyelitis_Optica_Spectrum_Disorder_with_Anti-AQP4_Antibodies, Rasmussen_Encephalitis) set evidence_source consistently, avoid NO_EVIDENCE padding, and quote faithfully. The failure mode being cleaned up is recognizable: "cite something adjacent and explain the gap away" — an explanation that argues for relevance rather than a snippet that carries the claim. The supports: enum is doing its job (curators tagged these honestly); the explanations are where the drift lives.