Pathograph Review: Non-Mendelian Cardiac Disease (2026-09-02)

Review of every non-Mendelian cardiac entry in kb/disorders/ and of the causal graph each one builds. Covers scope definition, three systemic defects found, per-entry repairs, and the follow-up work the review deliberately did not attempt.

Scope

"Non-Mendelian cardiac disease" is not a slot in the schema, so the set was derived rather than looked up. An entry is in scope when its subject is the heart, the great vessels, or the pulmonary circulation acting on the right heart, and its curated etiology is not single-gene Mendelian. That admits six etiologic classes:

Class	Entries
Acquired: infectious, immune-mediated, toxic, stress-induced	16
Complex, degenerative, or haemodynamic	12
Multifactorial or sporadic congenital	9
Acquired or idiopathic conduction disease	2
Maternally inherited mitochondrial	2
Idiopathic, cause unknown	2
Total	43

Of 150 cardiac-adjacent entries, 107 were excluded as Mendelian. Three in-scope entries were missed by a naive keyword sweep because they are named for pulmonary arterial hypertension rather than for the heart.

The etiology axis is frequently unfilled

The set had to be derived partly by reading prose because the machine-readable signal is missing or unbound:

43 cardiac-adjacent entries carry no inheritance: block at all, so Mendelian and non-Mendelian cannot be told apart from the data. Several are unambiguously Mendelian (hypertrophic cardiomyopathy, dilated cardiomyopathy, Brugada syndrome, short QT syndrome, left ventricular noncompaction).
Six carry free text with no HPO term bound, which CLAUDE.md already names as the common gap. Two of those are explicit non-Mendelian assertions: Cardiac_Sarcoidosis ("Multifactorial susceptibility") and Double_Outlet_Right_Ventricle ("Multifactorial and sporadic").

Note that HPO has no "Multifactorial inheritance" term. HP:0001426, which carried that label historically, is now "Non-Mendelian inheritance", and the 23-member InheritanceTerm enum contains no multifactorial concept at all. The nearest available terms are HP:0001426 (Non-Mendelian inheritance), HP:0010982 (Polygenic inheritance) and HP:0003745 (Sporadic). A curator reaching for "multifactorial" will not find it, which is a plausible reason these two blocks were left as free text.

Method

Graph diagnostics — dismech.graph.build_causal_graph per entry, recording node and edge counts, orphan targets, isolated nodes, and weakly connected components.
The project's own QC metric — dismech.qc_plugins.causal_inlink_coverage, which counts a phenotype as connected when at least one edge with predicate causes, leads_to, triggers or exacerbates targets it. Baselines were computed from git HEAD so before/after figures use one definition.
Gates — check_causal_targets, check_entity_refs, check_duplicate_yaml_keys, check_enum_values, and linkml-validate against the Disease class.
Content review — six parallel reviewers, one per cardiac subdomain, reading each chain for direction, missing intermediates, and module conformance.

Finding 1: the phenotype layer is not wired into the graph

This is the dominant defect and it is much worse in this set than in the knowledge base as a whole.

Set	Phenotypes reached by a causal edge
Non-Mendelian cardiac (40 entries with phenotypes)	68/414 (16.4%)
Whole knowledge base (2,506 entries)	13,943/29,021 (48.0%)

Twenty-one of the forty sat at exactly zero. The mechanism layer is not the problem: only 6% of pathophysiology nodes were unconnected, against 51% of all nodes. The deficit is specifically the mechanism-to-phenotype edge, so clinical manifestations were curated as well-evidenced lists that never entered the causal graph and never reached the rendered pathograph.

A plausible reason the acquired diseases fare worse than the Mendelian ones: a Mendelian entry has a single obvious chain from gene to lesion, whereas an acquired disease presents a broad symptom list with no single parent mechanism, so the wiring step is a judgement call rather than a transcription.

Finding 2: `Pathophysiology.consequences` is a dead schema slot

consequences (multivalued) and consequence (scalar) are real slots on Pathophysiology that validate cleanly and are ignored by every consumer: the graph builder in src/dismech/graph.py, the Jinja templates, all thirteen modules under src/dismech/export/, and the claim extractor, which lists consequence among the structural slots it skips. Content written there is invisible everywhere.

This is worse than a dangling bare-name target. A dangling target at least renders a red phantom node and is caught by check_causal_targets; a consequences entry produces nothing at all, and no gate looks for it.

Twenty-seven nodes across the knowledge base use it. The largest casualty is outside this review's scope: Alzheimer_Disease states its entire causal architecture this way across 22 nodes. Within scope, Pulmonary_hypertension used it for the right-ventricular-hypertrophy to right-heart-failure step.

The schema itself records the ambivalence, carrying todos: [unify consequences and consequence] on the slot. No GitHub issue covers this.

Finding 3: two thirds of genetic nodes never reach the mechanism graph

build_causal_graph links a genetic[] item to a pathophysiology node only by matching gene keys, drawn from the mechanism node's own gene/genes descriptors. A susceptibility gene named only in a node's prose therefore stays isolated.

Across the in-scope entries, 75 of 116 genetic nodes (65%) were isolated for this reason. Worst affected were Congenital_Heart_Disease (13 of 13), Pulmonary_hypertension (9 of 9), Coronary_Artery_Congenital_Malformation (6 of 6), and Tetralogy_of_Fallot (5 of 5).

Closing this requires adding HGNC-bound gene descriptors to mechanism nodes, which is an ontology-binding change rather than a wiring change, so it was left for follow-up.

Finding 4: module conformance is largely absent, including where a module exists

Twenty-four of the forty entries declared no conforms_to on any node. The sharpest case is that Pulmonary_hypertension does not conform to the pulmonary_vascular_remodeling module, although nine other entries do, three of which (Eisenmenger_Syndrome, Ventricular_Septal_Defect, Rheumatic_Heart_Disease) are downstream consequences of the very condition the general entry describes.

Similar gaps, each where a sibling entry already conforms: Cor_Pulmonale to pulmonary_vascular_remodeling; Heart_Failure, Takotsubo_Cardiomyopathy and Peripartum_Cardiomyopathy to cardiomyopathy_maladaptive_remodeling.

Conformance is also uneven within the pulmonary arterial hypertension family: Idiopathic_Pulmonary_Arterial_Hypertension declares it on five nodes, Heritable_Pulmonary_Arterial_Hypertension on one, and Drug_or_Toxin-Induced_Pulmonary_Arterial_Hypertension on none, despite all three walking the same chain.

Coverage gap: acquired long QT has no entry

CLAUDE.md notes that the cardiac_ion_channel_repolarization module scopes itself to inherited arrhythmia syndromes in structurally normal hearts, and that drug-induced long QT would need a separate module. There is currently no entry for acquired or drug-induced long QT syndrome either, though the knowledge base curates other drug-toxicity entities including Anthracycline_Induced_Cardiomyopathy and Drug_or_Toxin-Induced_Pulmonary_Arterial_Hypertension.

Repairs made

Every entry in scope was passed through a wiring repair: missing mechanism-to-phenotype downstream edges, phenotype sequelae, reports_on readouts for investigation and test-result phenotypes, treatment target_mechanisms, environmental influences_mechanisms with an explicit environmental_effect, model modeled_mechanisms, and biomarker readouts.

Measure	Before	After
Phenotypes reached by a causal edge	68/414 (16.4%)	353/414 (85.3%)
Phenotypes touching the graph at all	—	408/414 (98.6%)
Entries with no wired phenotype	21	0
Edges across the 40 entries	568	1,126
Grandfathered dangling targets, KB-wide	126	123

The largest individual recoveries were Kawasaki_Disease (1/40 to 40/40), Congenital_Heart_Disease (6/20 to 20/20), Coronary_Artery_Congenital_Malformation (0/13 to 13/13), Histiocytoid_Cardiomyopathy (0/15 to 14/15) and Eisenmenger_Syndrome (1/17 to 13/17).

Three dangling bare-name targets were repaired rather than baselined. In Idiopathic_Spontaneous_Coronary_Artery_Dissection both were the same defect, one edge bundling four phenotypes, now split into separate edges naming the real phenotypes with the flow-limitation and regional-ischemia steps carried as intermediate_mechanisms. In Posterior_Myocardial_Infarction the target was retargeted to the ischemic cell-death node, which is what the curated claim says.

The metric understates correct curation

PhenotypeConnectivityPlugin counts only causes, leads_to, triggers and exacerbates. A readout edge does not count, by deliberate design. But an ECG sign, an echo measurement or a troponin value should be modelled as a reports_on readout, not as a consequence, so an entry rich in investigation phenotypes cannot reach 100% by curating correctly. After this pass, 55 phenotypes across the set are wired as readouts and score as unconnected.

Posterior_Myocardial_Infarction is the clearest case: it reads 1/5 while being fully connected, because four of its five phenotypes are ECG signs and a troponin level. The metric is not wrong, but it should not be read as a curation target.

A correction to an earlier assumption

Unlinked treatments looked like a defect and are not. Restricted to therapeutic actions, 105 of 206 treatments in scope carried target_mechanisms at baseline, against 48% knowledge-base-wide, so the set was normal. The apparent gap came from counting monitoring, screening and counselling actions, and 0 of 222 such actions carry target_mechanisms anywhere in the knowledge base. That is a consistent convention: the graph has no monitoring predicate, and linking surveillance to a mechanism would assert that surveillance acts on it.

Deliberate non-fixes

These were found and left alone, with the reason recorded rather than the change forced:

Node-name collisions where one name is both a pathophysiology node and a phenotype, which the flat namespace merges (issue #9896). Found in Pulmonary_hypertension (Right Ventricular Hypertrophy), Tetralogy_of_Fallot (two), Pericarditis (Cardiac tamponade), Rheumatic_Heart_Disease and Endomyocardial_Fibrosis (Congestive heart failure), and Myocardial_Infarction (Elevated cardiac troponin, phenotype and biochemical entry). Renaming a curated node to satisfy the graph model would be the wrong direction of fix.
Isolated genetic nodes, for the architectural reason in Finding 3.
Curator decisions recorded in notes. Ventricular_Septal_Defect leaves two phenotypes unwired with a stated reason. Eisenmenger_Syndrome prostanoid therapy has an obvious mechanism target contradicted by the entry's own REFUTE evidence that pulmonary resistance did not improve. Myocarditis records that no influences_mechanisms link is asserted for exercise. Histiocytoid_Cardiomyopathy says the relationship between its two lesions is unresolved.
Pulmonary_hypertension's remaining dangling target. Right Heart Failure names no node, and no phenotype is a right-heart-failure node. Retargeting to Edema or Fatigue would downgrade a claim whose evidence is explicitly about right-sided failure. The correct fix is a new pathophysiology node.
A two-cycle in Coronary_Vasospasm. The edge from smooth-muscle hyperreactivity to the nitric-oxide defect carries a description stating the reverse direction, and duplicates the reciprocal edge already curated at the nitric-oxide node. Either the edge is redundant or its description is wrong; a curator should decide.

Entries where wiring is not the fix

Kawasaki_Disease has 40 phenotypes hanging off 2 pathophysiology nodes. It is now fully connected but structurally a two-hub star, not a chain. It has no node for the unknown trigger, no inflammasome node despite curating anakinra and listing NLRP3 among susceptibility genes, no endothelial-activation node distinct from the arteritis, none of the standard three-stage arteritis histopathology, and no thrombosis or stenosis node between the aneurysm and infarction. It also carries two separate phenotypes bound to the same term, HP:0002617.
Coronary_Arterial_Fistulas compresses steal physiology, chamber volume load, ischemia, arrhythmia and heart failure into a single node.
Patent_Ductus_Arteriosus has no node for the diastolic-steal physiology its own intraventricular-haemorrhage and retinopathy evidence turns on, and no separation of functional constriction from anatomic remodelling.
Pulmonary_hypertension has no thromboembolic-obstruction node despite curating both anticoagulation and balloon pulmonary angioplasty as treatments for chronic thromboembolic disease, and still uses the deprecated free-text percentage field for prevalence.
Heritable_Pulmonary_Arterial_Hypertension, though Mendelian and out of scope, was found to be in the pre-repair state: 32 nodes, 8 edges, no phenotype reached by any causal edge. It needs the same pass.

Follow-up worth filing

Decide what Pathophysiology.consequences and consequence are for. Either wire them into the graph or retire them and migrate the 27 nodes, starting with Alzheimer_Disease. Today they are a trap that validates.
Give genetic susceptibility loci a route into the pathograph that does not require a gene descriptor on a mechanism node.
Add module conformance where a module already exists, beginning with Pulmonary_hypertension and Cor_Pulmonale against pulmonary_vascular_remodeling. Cor_Pulmonale and Pulmonary_hypertension currently curate the same pulmonary-hypertension-to-right-ventricle biology twice under disjoint node names with no shared module and no cross-reference.
Bind an HPO term on the six free-text inheritance blocks, and consider whether an entry with no inheritance: block at all should be flagged.
Curate an acquired or drug-induced long QT entry, which would need a module of its own since cardiac_ion_channel_repolarization scopes itself to inherited syndromes.

Validation

just validate-disorders over all 38 changed entries: schema, ontology terms and references all pass, with 2,186 of 2,350 snippets verified against the cached references and 164 skipped by prefix. The four ungated whole-knowledge-base gates (check_causal_targets, check_entity_refs, check_duplicate_yaml_keys, check_enum_values) and check_snippet_grading are clean. A history record was added for each changed entry; just validate-history-all passes over all 8,010 records.

No PMID, snippet or reference title was invented anywhere in this pass. An evidence block appears on a new edge only where an item already present in the same file supports that specific edge; every other new edge carries description prose alone.