Hybrid H4
Triage × Risk-Factor CLIOH
from one covariate-structured elicitation, get both a live prioritization score (Triage §4) and the implicit risk-factor weights (Risk-Factor §7) as a second deliverable.
- Evidence standard
- split — Triage deliverable is external (outcomes); Risk-Factor deliverable is internal-now → external-later
- Panel
- On-call attendings + senior residents who make the call (the eventual users)
- Validation
- Prospective outcome correlation (Triage); empirical weight validation as data accrues (Risk-Factor); drift monitoring
What it's for
A single covariate-structured Bradley–Terry elicitation over real case profiles produces (a) a real-time urgency/prioritization score (the Triage deliverable) and (b) the fitted feature coefficients read as risk-factor weights (the Risk-Factor deliverable).
Covariate-structured BT already estimates how each case feature moves the worth — that is the risk-factor weighting. Running a separate Risk-Factor study would re-elicit the same information. H4 recognizes that one structured elicitation yields two products: an operational score and an interpretable weight set, the latter being a publishable prediction-rule prior.
"What's the urgency order for live cases (operational), and what are the implicit weights on the features driving it (interpretable / prediction-rule)?"
When to use it
Use when you're building a live triage tool on real cases and want the interpretable risk-factor weights as a second, lower-marginal-cost deliverable. Don't when you only need one (use the parent), when you're ranking triage criteria rather than scoring real cases (that's still Discovery §1), or when the feature set isn't operationally available at the moment of the call.
What you get
- Triage: a real-time urgency score / EHR-dashboard widget (two cases → recommended order + CI + audit trail).
- Risk-Factor: the implicit feature weights with CIs — an interpretable, publishable prediction-rule prior, refinable against outcomes.
A real example
CORTICES OR/ER-Priority on real cases (candidate). Today OR-Priority (k=17) and ER-Priority (k=16) rank urgency criteria (Discovery §1). Deployed as covariate-structured BT on real OR-board/ED case profiles, the same elicitation yields a live triage score and the recovered urgency-factor weights — the Interurban-2026 by-design Strength-Spread slate is the natural demonstrator. Reusable artifact: a covariate-structured case model serving both operational and interpretive ends.
Validation
Triage: prospective outcome correlation; head-to-head vs incumbent protocol; throughput/LOS metrics; ongoing calibration (dynamic BT). Risk-Factor: empirical discrimination/calibration of the weights as data accrues. Pre-register both. No clinical deployment of the triage widget without separate IRB + prospective validation + SaMD-grade QMS (binding, from Triage §11).
Common pitfalls
(a) Treating the risk weights as empirically validated before outcomes confirm them (they're an expert prior at first). (b) Deploying the triage score clinically before prospective validation + SaMD QMS. (c) Collinearity distorting both deliverables. (d) Feature drift between elicitation and deployment (monitor with dynamic BT). (e) Confusing the input with Discovery — if you're ranking criteria, not scoring real cases, you don't yet have H4. (f) Conflating the two timelines — the operational score and the validated weights mature on different schedules; report each honestly.
How it works — show me the method
Comparison structure. Covariate-structured / regression BT (Springall 1973; Cattelan 2012): worth(case) = f(case features), from bare-vote, binary forced-choice pairwise "which case goes first?", NO ties (ADR-CLIOH-03); BIBD over case pairs; ATRD Round 2 (ADR-CLIOH-04); single-step back button (ADR-CLIOH-02). The fitted model is the triage score; its coefficients are the risk weights. Longitudinal → dynamic/time-varying BT (Varin & Firth 2013) for drift.
Panel. On-call attendings + senior residents who actually make the call (10–20+), including the eventual users for face validity and adoption.
Ground truth. The two deliverables sit on different truth loci and timelines. The Triage score is external — validate against LOS, mortality, complications, time-to-OR. The Risk-Factor weights are internal now (the expert-derived prior) → external later (validate empirically as outcome data accrues). Don't claim the weights are empirically validated until they are; don't deploy the triage score clinically until prospectively validated.
Statistical backbone. Covariate-structured BT via choix (bootstrap CIs, Firth). The coefficient vector = risk-factor weights (interpretable as a logistic prediction-rule prior, per Risk-Factor §7 prior-then-update logic). Dynamic BT for drift. Bayesian-hierarchical BT warranted where the covariate count is high relative to data, or for the prior-then-posterior update on the weights. Watch collinearity among features (distorts both the score and the weights).