← Transparent Dosing·Literature Reproducibility

Literature Reproducibility

Every Vancomyzer release runs through a library of published vancomycin cases. If our calculator drifts from the cited literature beyond a pre-declared tolerance, the build fails before it ships. The result is on this page — auditable, deterministic, and refreshed on every deploy.

Summary

5 / 5 engine tests within tolerance· 1 reference band

Median |AUC₂₄ error|: 0.8%
Max |AUC₂₄ error|: 25.0%

Abstract + Results — structural parameters for the typical individual (35y, 70kg, SCr 0.83 mg/dL)

✓ Within tolerance

What it tests: Implementation correctness: our engine should reproduce the Colin 2019 typical-adult clearance (4.10 L/h) and the AUC₂₄ it implies at 1 g q12h.

This is a 'reference prior reproduction' test — most popPK papers don't publish per-patient individual cases with full demographics and AUC, so we test against the model's typical-individual output instead. If our AUC₂₄ at this regimen lands within 8% of 488 mg·h/L, the Colin 2019 prior is implemented faithfully.

Patient: 35y M · 70 kg · SCr 0.83 mg/dL·Regimen: 1000 mg q12h × 4 doses

	Published	Vancomyzer	Δ
AUC₂₄ (mg·h/L)	487.8	487.5	-0.1% ✓

Page 8, Section 4 Discussion: verbatim text describing predicted CL for a 60y/65kg/SCr 0.97 mg/dL patient

✓ Within tolerance

What it tests: Implementation correctness of FDecline × FSCR composition at age 60 with SCr 0.97 mg/dL — verifies the engine's age-decline function (50% CL by 61.6y) and SCr covariate compose correctly against Colin 2019's published per-patient clearance.

Verifies our engine's composition of the Colin 2019 age-decline (FDecline, 50% at 61.6y) and SCr (FSCR) covariates on top of size-allometric scaling. At 60y/65kg/SCr 0.97, Colin published CL = 2.55 L/h verbatim — at 750 mg q12h, AUC₂₄ should land near 588 mg·h/L.

Patient: 60y M · 65 kg · SCr 0.97 mg/dL·Regimen: 750 mg q12h × 5 doses

	Published	Vancomyzer	Δ
AUC₂₄ (mg·h/L)	588.2	587.9	-0.0% ✓

Table 2 final-model parameter estimates (CL = 5.72 × (TBW/70)^0.535) + Results section on dosing simulation for the 130 kg subgroup

✓ Within tolerance

What it tests: Obesity-model activation: at BMI ~47, our Vancomyzer Obesity Model should produce a CL close to the Smit 2020 covariate prediction (7.93 L/h at 130 kg).

Visible by design: our Vancomyzer Obesity Model composes Smit 2020 with Zhang 2024 + Janmahasatian FFM, so it produces a higher CL (~25% AUC lower) than pure Smit at 130 kg/normal renal function. We publish that drift here transparently. The wide tolerance bounds catastrophic regression; the modest realized delta IS the documented design.

Patient: 35y F · 130 kg · SCr 0.8 mg/dL · 165 cm·Regimen: 2275 mg q12h × 6 doses

	Published	Vancomyzer	Δ
AUC₂₄ (mg·h/L)	574.0	430.3	-25.0% ✓

Results — n=31 BMI ≥40 adults; median TBW 147.9 kg, BMI 49.5; measured 24-h urine ClCr 124.8 mL/min/1.73m²; median dose 4000 mg/day; measured cohort-median AUC₂₄ 582.9 mg·h/L (IQR 513.8–726.2); NONMEM popPK CL 6.54 L/h, V 0.51 L/kg

✓ Within tolerance

What it tests: Real-patient cohort obesity test: cohort-median AUC₂₄ of 583 mg·h/L (IQR 514–726) from 31 prospectively sampled BMI ≥40 adults. Our Vancomyzer Obesity Model should land inside the published IQR.

Real prospective cohort: 31 BMI ≥40 adults sampled at steady state with measured 24-h urine ClCr. Median AUC₂₄ 583 mg·h/L (published IQR 514–726). Our obesity-model branch should land inside the published IQR for a cohort-median patient at 2 g q12h.

Patient: 50y F · 147.9 kg · SCr 0.9 mg/dL · 173 cm·Regimen: 2000 mg q12h × 6 doses

	Published	Vancomyzer	Δ
AUC₂₄ (mg·h/L)	582.9	528.7	-9.3% ✓

Results — n=12 adult obese patients (median age 61, BMI 45, ClCr 86 mL/min). Full-data AUC₂₄ across 4 priors: 437–489 mg·h/L. Peak + trough (AUC_PT) approximated the full-data AUC best; midpoint+trough (AUC_MT) and trough-only (AUC_T) tended to overestimate.

✓ Within tolerance

What it tests: Sparse-sampling Bayesian AUC₂₄ recovery in obesity: fed a peak + trough from a cohort-typical patient, our posterior AUC should land inside the paper's published 437–489 mg·h/L band that four popPK priors produced with full data.

Carreno 2017 showed peak + trough Bayesian fits recover the full-sampling AUC₂₄ in obese adults to within ~10%. Our engine fed the same two-point profile should land in the 437–489 mg·h/L band the paper published across four popPK priors.

Patient: 61y M · 130 kg · SCr 1 mg/dL · 170 cm·Regimen: 1250 mg q12h × 5 doses·Levels: 25 mcg/mL @ 2h, 10 mcg/mL @ 11.5h

	Published	Vancomyzer	Δ
AUC₂₄ (mg·h/L)	463.0	459.2	-0.8% ✓

Abstract + Table 1 (cohort demographics) + Section 3.2 Main Results (per-model AUC₂₄ means)

◇ Reference band

What it shows: Industry-context: three popPK models inside one Bayesian platform (Tucuxi) produce AUC₂₄ estimates that differ by ~20% on the same 188 ICU adults — the choice of model alone changes dosing decisions in 1 of 3 cases.

Same 188 ICU patients, same 466 measured concentrations, three different popPK priors. The cohort-mean AUC₂₄ spans 469 (Goti) to 562 (Colin) — a ~20% inter-model range. Three-way agreement on AUC dosing category was only 48%. Platform choice is itself a clinical decision.

Cohort: 188 adult ICU patients · 466 AUC₂₄ estimations · mean age 58 ± 17 y · 63% male · APACHE III 62 ± 22 · 39% mechanically ventilated · 35% on vasopressors · Royal Prince Alfred Hospital, Sydney, 2019-2020

Cohort-mean AUC₂₄ (mg·h/L) per popPK model

Goti 2018 (via Tucuxi)

469 ± 148

Colin 2019 (via Tucuxi)

Vancomyzer uses this prior

562 ± 172

Thomson 2009 (via Tucuxi)

517 ± 164

200500800

Where Vancomyzer sits: Vancomyzer's default adult prior is Colin 2019 — at the cohort level, our engine should produce AUC₂₄ distributions centered near 562 mg·h/L (the Colin band). We do NOT match the Goti or Thomson bands; those are shown to surface the inter-model variance clinicians face when choosing a Bayesian platform.

Honest limitations

These are reproducibility tests against published model parameters, not real-world clinical validation. Most popPK papers don't publish per-patient individual cases with full demographics + dose + AUC — they publish covariate equations and population-typical predictions. So these cases test "does our engine reproduce the cited model's typical-individual output?", not "does our engine match a published real patient's observed AUC?".
Cases drawn from our derivation papers (Colin 2019) confirm implementation correctness — by construction the typical-individual output should match within a few percent.
The Smit 2020 case is published with a by-design drift: our Vancomyzer Obesity Model composes Smit + Zhang 2024 + Janmahasatian FFM and produces a higher CL (~25% lower AUC) than pure Smit at 130 kg. We show that drift transparently rather than tuning the test until it looks clean.
No pediatric, dialysis, or post-transplant cases in this set. The platform's prior is an adult population model; we don't attempt to validate scenarios outside its derivation cohort.
Cases we attempted but could not extract, or dropped after attempting (now with full text in hand for several): Rybak/ASHP 2020 (read full text — the executive summary contains narrative recommendations only, not a per-patient worked example to reproduce; verbatim cap citations now in code comments instead); Pai 2014 (read full text — Table 1 reports aggregate AUC ratios across n=47 sparse-Bayesian validation cohort but no per-patient demographics + AUC, so we cite Pai 2014 as the methodology source for our Carreno sparse-Bayesian case rather than its own card); Patanwala 2022 multi-platform ICU comparison (read full text — publishes cohort-mean AUC per platform (Goti 469, Colin 562, Thomson 517 across 188 ICU adults) but no per-patient data; the cohort-mean format doesn't fit the per-patient test schema — deferred until we add a "reference band" card type); Turner 2018 (read full text — Tables 1-3 are aggregate medians + IQRs across 19 ICU patients per platform, no per-patient breakdown; deferred for the same reference-band card type as Patanwala); Drennan 2024 trough-only Bayesian (no such paper exists in PubMed); Neely 2014 cohort trough (test was circular — our Bayesian fitter trivially matched its own input observation, and the framing conflicted with Vancomyzer's AUC-targeted positioning even though Neely's own paper argued AUC over trough); Shingde 2020 single-sample Bayesian (the candidate's "published AUC" was analytically derived rather than extracted from the paper, which would have made the test circular). These dropouts are documented openly here rather than swept under the rug.
A "within tolerance" result does not mean the recommendation is correct for any individual patient. Every clinical recommendation remains the responsibility of the licensed clinician at the bedside.
When a case shows drift beyond its tolerance, we display it anyway with an amber badge and explanation. The page is intended to surface honest engine behavior, not to be a curated success story. The build itself hard-fails on undocumented drift, which is why this page is up to date on every deploy.