Independent Engine Cross-Check
A second, independently-developed pharmacokinetic engine should reach the same answer ours does. We checked: given identical priors and identical measured levels, does Vancomyzer’s Bayesian engine reach the same posterior PK parameters as Tucuxi — a separate, independently-built model-informed precision dosing program developed by a different academic team (the REDS institute at HEIG-VD, Switzerland)? Across 200 simulated patients, the two engines agree to under 1% (median) on every parameter.
Engine-to-engine agreement · n = 200
Relative difference between the two engines’ posterior estimates (vs the mean of the two), per PK parameter. Both engines were given the same population prior and the same two measured levels.
| Parameter | median |Δ| | mean signed | p90 |Δ| | p95 |Δ| | max |Δ| |
|---|---|---|---|---|---|
| Clearance (CL) | 0.81% | +0.97% | 3.48% | 5.65% | 18.10% |
| Central volume (V₁) | 0.85% | +0.62% | 2.23% | 3.00% | 32.76% |
| Inter-comp. clearance (Q) | 0.66% | -0.01% | 1.76% | 2.93% | 21.34% |
| Peripheral volume (V₂) | 0.23% | +0.36% | 1.58% | 3.83% | 13.77% |
All four parameters agree to under 1% at the median, with negligible systematic offset (largest is CL at +0.97%). The wider individual maxima come from a handful of augmented-renal-clearance patients — see below.
Accuracy vs the known synthetic truth
Because the patients are simulated, the “true” PK is known. Median absolute error of each engine’s posterior against that truth, with the unfitted prior shown for reference (lower is better).
| Parameter | prior (no fit) | Vancomyzer | Tucuxi |
|---|---|---|---|
| Clearance (CL) | 31.7% | 10.2% | 10.3% |
| Central volume (V₁) | 28.9% | 26.0% | 26.3% |
| Inter-comp. clearance (Q) | 53.2% | 53.0% | 52.6% |
| Peripheral volume (V₂) | 26.2% | 25.4% | 25.2% |
The two engines are statistically interchangeable — within 0.1–0.4 points of each other on every parameter. An honest nuance: only clearance is materially improved by a two-level fit (31.7% → ~10.2%). The inter-compartmental and peripheral-volume terms (Q, V₂) are barely moved — a peak-plus-trough at steady state simply doesn’t constrain them, so both engines correctly leave them near the prior. That the two independent engines behave identically on the under-constrained parameters is itself corroboration, not a defect.
Where the engines diverge most
The largest disagreements cluster at creatinine-clearance extremes, where the steady-state trough falls toward assay noise and two levels under-constrain the fit. Top three per parameter (patient id, CrCl, Δ):
Clearance (CL)
Central volume (V₁)
The single largest divergence on both parameters is the same patient (p196, CrCl 177 mL/min — augmented renal clearance). This is expected for sparse-data Bayesian estimation; it affects only the distribution tail, not the median.
Methodology
- Generate 200 synthetic ICU patients (seed 42, the same generator as the Predictive Performance page), with “true” PK drawn from the Goti-2018 population model.
- Simulate a steady-state peak and trough for each patient, with realistic assay noise.
- Compute Vancomyzer’s per-patient Colin-2019 starting estimate (prior), then run our Bayesian engine on the two levels to obtain the individualized estimates for clearance and volumes (CL, V₁, Q, V₂).
- Set up the Tucuxi program with the same vancomycin model (two-compartment, IV infusion), the same starting estimate, and matched variability and assay-error settings, then give it the same dosing history and the same two levels, and run its Bayesian fit.
- Compare the two programs’ individualized estimates, parameter by parameter. The comparison confirms every one of the 200 patients is accounted for, and reports nothing unless all are present.
Comparator: Tucuxi — a free, openly-available dosing program. Tucuxi is developed by the REDS institute at HEIG-VD, Switzerland (Prof. Yann Thoma), and is described in a peer-reviewed publication.
Scope & honest limitations
This compares the two programs' calculations, not the Colin model itself
The published Colin-2019 vancomycin model is not openly available in a form Tucuxi can load (it comes only with Tucuxi's own desktop application). So we entered the Colin model into Tucuxi ourselves. This comparison therefore confirms that two independently-built programs reach the same answer from the same model — it is not a separate confirmation of the Colin equations. Those are checked on the Literature Reproducibility page.
Both programs start from the same estimate
Tucuxi was given Vancomyzer's exact starting estimate for each patient. That is deliberate: it isolates the dose calculation itself as the only thing being compared.
Simulated patients
The patients and their vancomycin levels are computer-generated, not real. No patient data is involved. Real-world performance is addressed separately by the Predictive Performance analysis.
Two-level sampling
A peak plus a trough pins down clearance well, but does not fully pin down the distribution-volume terms. The agreement on clearance and central volume is the key result; agreement on the other two reflects both programs staying near the shared starting estimate.
What this establishes: a separate, independently-built Bayesian dosing program, given identical starting estimates and data, reproduces Vancomyzer’s individualized clearance and volume estimates to within ~1% (median, n=200) and matches its accuracy against a known answer to within 0.4 points. Vancomyzer’s calculations are corroborated by an independent program.