Phenotype Associations vs. Detection Algorithms

Two knowledge layers in a dismech entry describe a disease's phenotypes, and they answer opposite questions. This note contrasts them so curators keep the two distinct — a distinction that recurs for every disease that carries both a phenotypes: block and a definitions: block (a computable case definition). Familial hypercholesterolemia (FH) is the worked example throughout.

The core distinction: direction of inference

	HPOA-style phenotype associations (phenotypes:)	Detection algorithms (definitions: with definition_type: PHENOTYPE_ALGORITHM)
Encodes	P(phenotype | disease) — given the disease, what features appear, and how often	P(disease | data) — given a patient's data, is this the disease
Direction	Forward (disease → feature); descriptive	Backward (features → disease); discriminative / ascertainment
Question	What does the disease look like?	Which patients probably have the disease?
Quantity attached	Frequency within disease (FrequencyEnum: OBLIGATE, VERY_FREQUENT, …)	Discriminative weight / likelihood: point scores, model coefficients, ML feature importance, PPV/AUC
Feature vocabulary	HP ontology terms only	HP-mappable features plus numeric lab cutoffs, ICD codes, medication records, data trajectories, healthcare-utilization signals
Values	Qualitative HP terms	Operational thresholds (e.g. LDL-C ≥190 mg/dL)
Exclusion logic	None — a union of observed features	Explicit (e.g. rule out secondary causes)
Provenance	Literature / expert curation, aggregated over the disease	Real EHR/claims cohorts, statistically validated

The key asymmetry: frequency ≠ discriminative weight

The single most important point is that a feature's frequency in the disease and its value for detecting the disease are different axes, and both representations are correct on their own axis. Detection algorithms effectively re-weight the phenotype spectrum by specificity / likelihood ratio, not by prevalence-within-disease.

FH makes this vivid (weights below are from the Dutch Lipid Clinic Network score, which the FH definitions: block operationalizes):

Feature	Frequency in FH (phenotypes:)	Weight for detecting FH (DLCN)
Hypercholesterolemia (raw high cholesterol)	OBLIGATE (~100%)	Low — common in the general population, weakly discriminative
Very high LDL-C (e.g. ≥ ~325 mg/dL)	— (a threshold, not an HP term)	8 / max
Tendon xanthoma	FREQUENT / OCCASIONAL	6 — near-pathognomonic
Pathogenic LDLR/APOB/PCSK9 variant	(modeled in genetic:, not a phenotype)	8

Hypercholesterolemia is obligate yet poorly discriminative; tendon xanthoma is only occasional yet highly discriminative. A representation optimized for one axis is the wrong tool for the other.

What each layer holds that the other cannot

• Only the phenotype associations have the long tail of pleiotropic / rare features — for FH (HoFH), optic neuropathy, hepatic/renal/myocardial steatosis, arthralgia. Detection algorithms drop these because they add little discriminative power.
• Only the detection algorithms have numeric thresholds, exclusion criteria, family-history / pedigree variables, non-phenotypic administrative signals (diagnosis codes, medication records), and — for ML models such as FIND FH — emergent, clinically-uninterpretable features. None of these are representable as an HPOA annotation.
• The shared core (the intersection) is exactly the discriminative phenotypes: for FH, high LDL-C, tendon xanthoma, corneal arcus, premature CAD, family history. Point-score algorithms that operationalize published clinical criteria (DLCN, Simon Broome) are the bridge — they are HPOA-style phenotypes plus numeric thresholds and family history, assembled into a rule. Statistical models (FAMCAT) depart from that by re-weighting; ML models (FIND FH) depart the most.

Where the pathograph fits

Neither HPOA associations nor detection algorithms explain why a feature is discriminative. The dismech pathograph does: it is the causal layer that connects the descriptive and the diagnostic. In FH, tendon xanthoma and corneal arcus both trace to a single Extra-arterial Cholesterol Deposition node driven by the obligate LDL elevation — so their shared cause is exactly what makes them jointly informative for detection. An entry that carries all three layers holds three complementary views of one disease:

1. Mechanism — the pathograph (gene defect → ↑LDL → atherosclerosis / extra-arterial deposition → phenotypes);
2. Phenotype spectrum — the phenotypes: block (P(phenotype | disease) + frequency + evidence);
3. Computable detection — the definitions: block (P(disease | data), with thresholds, weights, and validation metrics).

Curation guidance

• Do not import a detection algorithm's variable list into phenotypes:. LDL-C thresholds, ICD codes, and family-history variables are not phenotype associations; they belong in a definitions: block.
• Do not infer frequency from an algorithm's weight, or a weight from frequency. They are different quantities; each needs its own evidence (see frequency-evidence-guidelines).
• Record detection algorithms as definitions: entries with definition_type: PHENOTYPE_ALGORITHM, a derivation_basis, and a validation_status carrying the reported metric (PPV/AUC). See the FH entry (kb/disorders/Familial_Hypercholesterolemia.yaml) for worked DLCN/PheKB, FAMCAT, and FIND FH examples, and hypothesis-based-phenotype-algorithms for the epistemic-grounding fields.
• Keep the discriminative core mechanistically anchored. When a phenotype is both frequent-and-discriminative, it should be reachable in the pathograph so its diagnostic value has a causal explanation.