Archetype 11
Phenotype / Classification Boundary CLIOH
locate the decision boundary between two or more discrete clinical phenotypes — not rank a single severity dimension.
- Evidence standard
- external
- What you compare
- Borderline real cases sampled densely at the boundary
- Panel
- Disease experts across the relevant subspecialties
- Output
- 2-D/3-D decision boundary + algorithmic phenotype classifier
- Validation
- Outcome differentiation across phenotype clusters
What it's for
A CLIOH study that uses pairwise comparisons specifically to locate the boundary between two or more discrete clinical phenotypes, then validates that boundary against external data (treatment response, genetics, long-term outcome).
"Where does FAI end and acetabular dysplasia begin, in the hips that look like both?"
When to use it
(a) A multi-phenotype disease space; (b) the overlap region is clinically consequential (different treatments); (c) genotype / outcome / treatment-response data exist to validate boundaries; (d) existing classifications collapse a 2-D space into 1-D arbitrarily (e.g., a single LCEA threshold).
When not to
(a) Single severity dimension → Classification §3; (b) no outcome differentiation between phenotypes (then the boundary is cosmetic); (c) variables not cases → Discovery §1.
What you get
A 2-D/3-D decision-boundary plot (cases on the measurement axes, color-coded by panel-inferred phenotype, smoothed boundary) + an algorithmic phenotype classifier.
A real example
- CLIOH_Acetabular — pairwise classification of borderline FAI-vs-dysplasia hips with outcome linkage. The defining lab instance of this archetype.
- Methodological cousin: the i-ROP vascular-severity-scale validation (Ophthalmology 2022, PMID 35157950) shows expert pairwise → continuous-scale → external-classification validation in a different domain (this is also the template for Hybrid H3, Classification × AI-Calibration).
Validation
Outcome differentiation across the derived phenotype clusters (do the two sides actually respond differently to treatment / have different natural history?). Replication at an independent CORTICES site. Comparison vs the existing single-threshold classification by outcome-prediction AUC.
Common pitfalls
(a) Forcing categorical structure onto a genuinely continuous phenomenon; (b) anchoring on existing radiographic thresholds (defeats the purpose); (c) inadequate boundary-region sampling; (d) too few outcome events to validate cluster separation.
How it works — show me the method
Item selection. Real cases sampled densely at the suspected boundary (not uniformly across the whole space). Boundary-region undersampling is the cardinal sin here. 60–150 cases typical.
What you compare. Cases carry multidimensional measurements (e.g., LCEA, alpha angle, femoral/acetabular version). The point is to recover where, in that multidimensional space, experts switch phenotype calls.
Comparison structure. Categorical / triadic plus pairwise. Bare-vote, no-tie default still applies (ADR-CLIOH-03); the "which phenotype" judgment is the categorical layer, the "more X-like" judgment is the pairwise layer. BIBD + ATRD on contested boundary cases.
Panel. Disease experts spanning the phenotype subspecialties (e.g., hip-preservation surgeons on both the FAI and dysplasia sides). N=15–30. Cross-subspecialty balance matters because the boundary is exactly where subspecialty priors diverge.
External ground truth. Required. Link borderline cases to treatment response, genetic data, or long-term outcomes. Without it you have expert opinion about a boundary, not a validated boundary.
Statistical backbone. Mixture-model BT / latent-class CJ, or a probit BT on 2+ latent dimensions to recover the boundary surface. Frequentist primary where feasible; Bayesian hierarchical here is more often warranted (small boundary samples, multidimensional latent structure) — this archetype is a strong candidate for the "Bayesian mandatory for boundary/hybrid" path (ADR-CLIOH-07 draft). Validate cluster separation against outcomes (ANOVA/ROC).
For researchers — reporting, IRB & grant language
Reporting standards. STARD + COSMIN. Pre-register on OSF. LitGuard + DLRP.
IRB / ethics. Human-subjects (real cases + outcome linkage); de-identified case bank under data-use agreement; CORTICES central IRB. No PHI in the instrument; aggregate-only public display.
Grant-application language.
"Existing radiographic classifications of borderline hip morphology collapse an inherently two-dimensional space (acetabular coverage × femoral morphology) onto a single arbitrary threshold, yielding poor interobserver reliability at exactly the cases where treatment choice is most consequential. Phenotype-Boundary CLIOH applies multidimensional Bradley–Terry / latent-class methods to expert pairwise judgments of borderline cases, recovering an outcome-validated decision boundary between femoroacetabular impingement and acetabular dysplasia."