Uncertainty and provenance in PK/PD evidence
A smooth PTA curve can look like certainty. Often it is a stack of borrowed numbers wearing a clean plot. The scientific job is not only to compute a probability. It is to show which assumptions, from which sources, under which scenario rules, produced that probability.
This lesson is about evidence hygiene. It is the difference between a reusable simulation and a one-off slide.
Provenance habits for antimicrobial PK/PD simulation, aligned with component-level evidence thinking.
Label every input like it might be audited
Useful labels look boring. That is why they work.
| Label | Meaning |
|---|---|
| Directly reported | Used as published for this purpose |
| Transformed | Unit change, free/total conversion, algebra |
| Donor / inherited | Borrowed from a related but non-identical source |
| Assumed | Filled so the model can run; must stay visible |
| Scenario-generated | Created by the simulation design itself |
Missing data should stay missing until a compatible donor is justified. Silent median fills create false completeness.
A practical test
If a coauthor asks “where did this fu come from?” and you need five minutes to remember, the provenance trail is already too weak for a methods section.Uncertainty comes in layers
- Parameter uncertainty: IIV, residual error, covariate effects.
- Structural uncertainty: compartment choice, infusion constraints, correlation.
- Evidence uncertainty: wrong population, wrong matrix, stale MIC ecology.
- Decision-rule uncertainty: which threshold counts as “enough.”
Monte Carlo on layer 1 alone can look precise while layer 3 quietly dominates.
Pretty sampling does not repair a shaky evidence base.
Source list for CL, V, fu, targets, MIC weights. Free/total notes. Population statement. Seed and N. One sensitivity check on the most fragile input.
Everything in the minimum set, plus scenario compatibility notes for donor evidence, correlation assumptions, convergence checks, and alternative MIC priors with dates.
Compatibility is a scientific decision
ICU continuous-infusion parameters are not automatically interchangeable with healthy-volunteer bolus data, CSF studies, or in vitro kill curves. Those sources can still help as donors or anchors. They should not be pooled as if they were one population. Write the compatibility rule in human language: what was allowed to enter the primary scenario, and what stayed outside with a label.
What good provenance feels like in a meeting
Someone asks why ICU PTA used a clearance relationship from a continuous-infusion critically ill study and an MIC table from a regional surveillance year. You can answer in one breath: primary PK from source A for scenario match, MIC weights from source B dated to that year, fu transformed from total using source C, inhibitor threshold from source D with a compatibility note. No scavenger hunt through forgotten spreadsheets.
If you cannot give that breath-length answer, stop polishing the figure and rebuild the input table first. Most “mysterious” PTA sensitivity traces back to one unlabeled donor.
Keep a one-page input ledger beside the code: parameter, value or distribution, source, transformation, scenario role. It is unglamorous. It is also what makes the project reusable six months later.
Takeaway
Reproducible PK/PD is less about fancier random draws and more about refusing to launder uncertain evidence into false certainty. If the trail of sources is clear, other people can challenge your result productively. That is a feature.
Previous: Joint targets for BL/BLI pairs · Next: HGT and MIC ecology · Then: Effect site versus plasma.