Imipenem popPK model review (Zhang 2025)

Model and source

Citation: Zhang P, Zhao Y, Zhu J, Yang Y, Liang G, Wang X, Yu Z. Population pharmacokinetics of imipenem in different populations for individualized dosing: a systematic review. Frontiers in Pharmacology 2025;16:1738055.
Article: https://doi.org/10.3389/fphar.2025.1738055
Supplement (Table S1, study protocols): https://www.frontiersin.org/articles/10.3389/fphar.2025.1738055/full#supplementary-material

This vignette is unusual for the package: its source is a secondary one. Zhang 2025 is a PRISMA-style systematic review that develops no model of its own. It catalogues 18 previously published imipenem population-PK models and tabulates, for each, the structural parameter estimates (Table 2), the inter-individual and residual variability (Table 2), the covariate screening and the final covariate equations (Table 3), and the dosing and sampling protocol (Supplementary Table S1).

The standing policy for a review is to skip it and queue its cited primaries. That policy carries an explicit exception for a review that tabulates the parameters, and this one does: Tables 2 and 3 together are enough to write down every differential equation and initialise every ini() value for most of the 18 constituent models. So the models below are transcribed from the review, each carrying a “re-verify against the primary” note in its description and reference metadata.

Every model file in this family records ALL PARAMETER VALUES ARE SECONDARY in its population$notes. Read the “Assumptions and deviations” section at the foot of this page before using any of them; it reports a measured audit of how faithfully the review reproduces its primaries, and that audit found a real defect.

Coverage of the review’s 18 studies

Four of the review’s 18 studies were already in nlmixr2lib, extracted directly from their primary publications. Those were not re-extracted; they are better provenance than anything this review can supply, and they serve below as the validation set for the review’s own accuracy. Two studies could not be reconstructed from the review alone. Nothing was silently dropped.

Study	Author (review’s citation)	Disposition	Model file
1	Ikawa et al. (2008)	Extracted	Ikawa_2008_imipenem
2	Lamoth et al. (2009)	Already in registry	Lamoth_2009_imipenem
3	Yoshizawa et al. (2013)	Extracted (2 files)	Yoshizawa_2013_imipenem_neonates / _children
4	Couffignal et al. (2014)	Already in registry	Couffignal_2014_imipenem
5	Li and Xie (2019)	Blocked: sieving coefficient unreported	NA
6	Dong et al. (2019)	Extracted	Dong_2019_imipenem
7	Li et al. (2020)	Blocked: dialysate-compartment role + Qd units	NA
8	Chen et al. (2020b)	Extracted	Chen_2020_imipenem
9	de Velde et al. (2020)	Extracted (NONMEM arm)	deVelde_2020_imipenem
10	Nguyen et al. (2021)	Extracted	Nguyen_2021_imipenem
11	Jaruratanasirikul et al. (2021)	Extracted	Jaruratanasirikul_2021_imipenem
12	Por et al. (2021)	Extracted	Por_2021_imipenem
13	Dao et al. (2022)	Extracted	Dao_2022_imipenem
14	Dinh et al. (2022)	Extracted	Dinh_2022_imipenem
15	Bai et al. (2024)	Already in registry	Bai_2024_imipenem
16	Lafaurie et al. (2023)	Extracted (2 files)	Lafaurie_2023_imipenem_neutropenia / _recovery
17	Truong et al. (2025)	Extracted	Truong_2025_imipenem
18	Wang et al. (2024)	Already in registry	Wang_2024_imipenem

14 model files extracted from 12 studies, 4 already covered from their primaries, 2 blocked. The two blocked studies are:

Li and Xie (2019) – a CRRT cohort whose total clearance is CL_CRRT + CL_body, with CL_CRRT = Sd x CRRTintensity x BW/1000. The sieving coefficient Sd is never given a value by the review, and every patient in that study was on CRRT, so the extracorporeal arm is not optional. The review also cites this study inconsistently – Table 1 and Table 2 say “Li and Xie (2019)” while Supplementary Table S1 says “Li et al., 2018” – so the primary should be identified before re-extraction.
Li et al. (2020) – a three-compartment CRRT model whose third compartment is a dialysate space carrying its own volume (0.012 L), its own intercompartmental clearance and its own assayed effluent concentration. The package’s canonical dialysate compartment is defined as a cumulative collected amount, not a volume-bearing distribution compartment, so the role needs ratifying. Separately, the dialysate-flow covariate Qd is centred on 500 with no unit stated; CRRT flows are prescribed in mL/h while the canonical DFR column is mL/min, and with an exponent of 1.944 a 60-fold unit error becomes a 2800-fold clearance error. Both questions turn on what the primary says, so the primary (Clin Ther 2020;42(8):1564-1577.e8, doi:10.1016/j.clinthera.2020.06.010) has been queued for acquisition and this study deferred until it arrives, rather than shipped on a guessed unit.

Jaruratanasirikul et al. (2021) was previously blocked and is now included. It needs an ABW (adjusted body weight) covariate column, which the register did not carry – WT, IBW, BMI, BSA and FFM were registered, ABW was not – because its central volume is V1 = 13.8 - 0.348 x (ABW - 60). ABW was ratified as a canonical covariate column with this extraction as its founding example; see inst/references/covariate-columns.md. Its two source quirks (a Table 2 central volume of 13.6 L against a Table 3 formula intercept of 13.8 L, and a linear volume term that goes negative above 99.7 kg) are recorded in the per-model deviations below.

Population

The review pools 18 studies from 10 countries published between 2008 and 2025, spanning neonates through the elderly. Sample sizes run from 10 patients (Ikawa 2008) to 247 (Chen 2020); only three studies enrolled more than 100. The study populations are predominantly critically ill adults, with specific cohorts for burns, intraabdominal infection, haematological malignancy, neutropenia, AECOPD, CRRT, ECMO, neonates and children.

Per-model demographics are in each file’s population metadata; the table below collects the cohorts for the 14 models extracted here from the review’s Table 1.

Model	Country	N	Age	Weight (kg)	Patient group
Ikawa_2008_imipenem	Japan	10	43.7 +/- 14.9 y	56.7 +/- 10.5	Intraabdominal infections
Yoshizawa_2013_imipenem_neonates	Japan	60	10.5 +/- 8.3 d	2.93 +/- 0.7	Neonates
Yoshizawa_2013_imipenem_children	Japan	39	9.61 +/- 3.16 y	29.5 +/- 10.9	Children
Dong_2019_imipenem	China	56	4.86 +/- 2.33 y	18.65 +/- 6.90	Children, haematological malignancy
Chen_2020_imipenem	China	247	67 (20-97) y	65 (37.5-110)	Critically ill, +/- ECMO
deVelde_2020_imipenem	Netherlands	26	51 (39-54) y	75 (66-85)	Critically ill
Nguyen_2021_imipenem	Vietnam	44	65 (60-72) y	50 (47-55)	AECOPD
Jaruratanasirikul_2021_imipenem	Thailand	50	56.2 (41-66.6) y	62.9 (52.8-70)	Critically ill, +/- ECMO
Por_2021_imipenem	USA	23	51-55 y	89.6-105.1	Burns, +/- CVVH
Dao_2022_imipenem	Switzerland	82	GA 26.9 wk, PNA 21 d	1.16 (0.5-4.1)	Neonates (extremely preterm)
Dinh_2022_imipenem	Vietnam	24	57.5 +/- 19.9 y	51.3 +/- 8.6	Critically ill
Lafaurie_2023_imipenem_neutropenia	France	16	37 (18.3-78.3) y	65.5 (48-101)	Neutropenic adults (during)
Lafaurie_2023_imipenem_recovery	France	16	37 (18.3-78.3) y	65.5 (48-101)	Neutropenic adults (recovered)
Truong_2025_imipenem	Netherlands	151	63 (51-72) y	70 (61.2-82)	Critically ill and non-critically ill

Programmatic access, e.g.:

str(readModelDb("Dao_2022_imipenem")()$population[c("n_subjects", "ga_range", "weight_median")])
#> List of 3
#>  $ n_subjects   : int 82
#>  $ ga_range     : chr "24.2-41.3 weeks (median 26.9)"
#>  $ weight_median: chr "1.16 kg (range 0.5-4.1)"

Source trace

Every ini() value in every file carries an in-file comment naming its origin. Because all 14 models share one secondary source, the trace collapses to a single rule, stated here once and repeated per parameter in the files:

Quantity	Source location in Zhang 2025
Structural typical values (CL, V1, V2, V3, Q, Q2, Q3, Ke, Kcp, Kpc)	Table 2, “Population typical value” column, the row for that study
Inter-individual variability	Table 2, “Inter-individual variability (IIV)” column
Residual variability (proportional, additive)	Table 2, “Residual variability (RV)” column
Covariate reference values and exponents	Table 3, “Formulation” column
Which covariates were screened but not retained	Table 3, “Covariates screened” vs “Covariates incorporated”
Covariate selection method and thresholds	Table 3, “Covariate analysis method” column
Demographics (N, age, weight, sex, country, design)	Table 1
Dose, sampling schedule, bioanalytical assay	Supplementary Table S1
Software and model-evaluation method	Table 2, last two columns

The covariate equations transcribed from Table 3 are:

Model	Covariate model (Zhang 2025 Table 3, verbatim)
`Ikawa_2008_imipenem`	none reported (“NR”)
`Yoshizawa_2013_imipenem_*`	per-kg parameters; no other covariate reported (“NR”)
`Dong_2019_imipenem`	`CL = 8.6 x (BW/18)^0.75 x (age/4.69)^0.265 x (CLcr/214)^0.509`; `V1 = 7.2 x (BW/18)`; `V2 = 6.51 x (BW/18)`; `Q = 0.996 x (BW/18)^0.75`
`Chen_2020_imipenem`	`CL = 8.88 x (CLcr/59.1)^0.295 x (BW/65.0)^0.306 x e^1.16 [ECMO] x e^eta`
`deVelde_2020_imipenem`	`Ke = 0.637 x ((eGFR CKD-EPI-abs)/119)^0.655 x e^eta`
`Nguyen_2021_imipenem`	`CL = 7.88 x (CLcr/75.54)^0.532 x e^eta`
`Jaruratanasirikul_2021_imipenem`	`CL = 13.3 + 0.112 x (eGFR CKD-EPI - 89)`; `V1 = 13.8 - 0.348 x (ABW - 60)`
`Por_2021_imipenem`	`Vc = 32.67 x (BW/99.5)^0.74 x (ALB/2.7)^-1.17 x e^eta`; `Vp = 41.23 x (ALB/2.7)^-3.68`; no CVVH: `CL = 15.31 x (CLcr/145.83)^0.46 x (BW/99.5)^0.33 x e^eta`; with CVVH: `CL = 13.78 x (BW/99.5)^0.75 x e^eta + 1.56`
`Dao_2022_imipenem`	`CL = 0.21 x (BW/1.16)^0.75 x (1 + 0.22 x (PNA-21)/21) x (1 + 1.31 x (GA-26.9)/26.9) x (46.6/SCr)^0.2`; `V = 0.73 x (BW/1.16)^0.75`
`Dinh_2022_imipenem`	`CL = 4.79 x e^(0.00642 x CLcr)`
`Lafaurie_2023_imipenem_*`	none reported (“NR”)
`Truong_2025_imipenem`	`CL = 14.6 x (CLcr CG/87.6)^0.462`

Reference subjects

Each model is exercised below at its own reference covariates – the centring values printed inside its Table 3 formula – so that every covariate term evaluates to exactly 1 and the model’s typical clearance and volume must equal the values printed in Table 2. That is what makes the checks in the next section genuine transcription gates: if a reference constant, an exponent, a unit conversion or the covariate algebra is wrong, the reproduced clearance will not match the published one.

# Reference covariates, published typical CL (L/h) and published central
# volume Vc (L) for each model, all at the reference subject.
#
# The published values are typed in as literal constants read off Zhang
# 2025 Table 2 / Table 3 -- deliberately NOT read back out of the model
# object, which would make the comparison vacuous.
cfg <- list(
  Ikawa_2008_imipenem = list(
    cov = list(), cl_pub = 9.42, vc_pub = 4.66, dose = 500),
  Yoshizawa_2013_imipenem_neonates = list(
    # per-kg model: CL = (CLr + CLnr) * WT, V = 0.466 * WT, at the cohort mean 2.93 kg
    cov = list(WT = 2.93), cl_pub = (0.0783 + 0.138) * 2.93, vc_pub = 0.466 * 2.93, dose = 2.93 * 20),
  Yoshizawa_2013_imipenem_children = list(
    cov = list(WT = 29.5), cl_pub = (0.187 + 0.0711) * 29.5, vc_pub = 0.203 * 29.5, dose = 29.5 * 15),
  Dong_2019_imipenem = list(
    cov = list(WT = 18, AGE = 4.69, CRCL = 214), cl_pub = 8.6, vc_pub = 7.2, dose = 18 * 20),
  Chen_2020_imipenem = list(
    cov = list(CRCL = 59.1, WT = 65.0, ECMO_STATUS = 0), cl_pub = 8.88, vc_pub = 20.5, dose = 500),
  deVelde_2020_imipenem = list(
    # micro-constant model: CL is a derived quantity, kel * Vc = 0.637 * 29.6
    cov = list(CRCL = 119), cl_pub = 0.637 * 29.6, vc_pub = 29.6, dose = 500),
  Nguyen_2021_imipenem = list(
    cov = list(CRCL = 75.54), cl_pub = 7.88, vc_pub = 15.1, dose = 500),
  Jaruratanasirikul_2021_imipenem = list(
    # Both covariate terms are LINEAR deviations, so the reference subject
    # is the one at which both deviations are zero: eGFR CKD-EPI = 89
    # mL/min/1.73 m^2 and ABW = 60 kg. vc_pub is the Table 3 FORMULA
    # INTERCEPT (13.8 L), not the Table 2 value (13.6 L) -- see the Errata.
    cov = list(CRCL = 89, ABW = 60), cl_pub = 13.3, vc_pub = 13.8, dose = 1000),
  Por_2021_imipenem = list(
    # ALB supplied in canonical SI g/L; the model converts to the g/dL the
    # source was calibrated on, so 27 g/L == the reference 2.7 g/dL.
    cov = list(WT = 99.5, ALB = 27, CRCL = 145.83, RRT_CRRT_STATUS = 0),
    cl_pub = 15.31, vc_pub = 32.67, dose = 500),
  Dao_2022_imipenem = list(
    # PNA supplied in canonical months; 21 days = 21/30.4375 months.
    cov = list(WT = 1.16, PNA = 21 / 30.4375, GA = 26.9, CREAT = 46.6),
    cl_pub = 0.21, vc_pub = 0.73, dose = 1.16 * 20),
  Dinh_2022_imipenem = list(
    # UNCENTRED exponential covariate: the published 4.79 L/h is CL at CRCL = 0.
    cov = list(CRCL = 0), cl_pub = 4.79, vc_pub = 11.1, dose = 500),
  Lafaurie_2023_imipenem_neutropenia = list(
    cov = list(), cl_pub = 14.3, vc_pub = 20.7, dose = 1000),
  Lafaurie_2023_imipenem_recovery = list(
    cov = list(), cl_pub = 10.9, vc_pub = 14.5, dose = 1000),
  Truong_2025_imipenem = list(
    cov = list(CRCL = 87.6), cl_pub = 14.6, vc_pub = 28.7, dose = 1000)
)
length(cfg)
#> [1] 14

Simulation

A single intravenous bolus into central is used for the transcription gates. A bolus makes both checks closed-form and exact: immediately after the dose the entire amount is in the central compartment, so Cmax = Dose / Vc; and over an infinite interval a linear system must clear the whole dose, so AUC(0, inf) = Dose / CL. Neither identity involves any quantity read back out of the model.

All random effects are zeroed with rxode2::zeroRe(), so these are deterministic typical-value solves. Nothing below depends on a drawn cohort, which also means nothing below can pass on one machine and fail on another because of solver-thread RNG partitioning.

# Observation grid: dense early so Cmax is captured at the bolus peak,
# long-tailed so the terminal phase is well characterised for AUCinf.
obs_times <- unique(c(
  0,
  seq(0.05, 2, by = 0.05),
  seq(2.25, 12, by = 0.25),
  seq(13, 24, by = 1),
  seq(26, 120, by = 2)
))

solve_one <- function(nm) {
  spec <- cfg[[nm]]
  mod  <- rxode2::zeroRe(readModelDb(nm))

  ev <- rxode2::et(amt = spec$dose, cmt = "central", time = 0) |>
    rxode2::et(obs_times)
  ev <- as.data.frame(ev)
  ev$id <- 1L
  # Covariates are added to the materialized data frame, never assigned
  # onto the rxEt object (rxode2 silently drops those assignments).
  for (cv in names(spec$cov)) ev[[cv]] <- spec$cov[[cv]]

  out <- rxode2::rxSolve(mod, events = ev)
  data.frame(model = nm, time = out$time, Cc = out$Cc)
}

sim <- dplyr::bind_rows(lapply(names(cfg), solve_one))
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc', 'etalvp', 'etalvp2'
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalcl_renal', 'etalcl_nonren'
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalcl_renal', 'etalcl_nonren', 'etalvc'
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc'
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc'
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalkel'
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc'
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc', 'etalvp'
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc'
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalcl'
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc', 'etalvp', 'etalq'
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc'
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc'
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalcl'
dplyr::glimpse(sim)
#> Rows: 1,974
#> Columns: 3
#> $ model <chr> "Ikawa_2008_imipenem", "Ikawa_2008_imipenem", "Ikawa_2008_imipen…
#> $ time  <dbl> 0.00, 0.05, 0.10, 0.15, 0.20, 0.25, 0.30, 0.35, 0.40, 0.45, 0.50…
#> $ Cc    <dbl> 107.29614, 69.15468, 50.41882, 40.69728, 35.20715, 31.74291, 29.…

Figure: typical-value profiles after a single IV bolus

Replicates, qualitatively, the clearance comparison of Zhang 2025 Figure 2: the wide between-study spread the review’s abstract reports is visible as a wide spread of terminal slopes.

PKNCA validation

sim_nca <- sim |>
  dplyr::filter(!is.na(Cc)) |>
  dplyr::select(id = model, time, Cc) |>
  dplyr::mutate(treatment = "IV bolus")

dose_df <- tibble::tibble(
  id        = names(cfg),
  time      = 0,
  amt       = vapply(cfg, function(x) x$dose, numeric(1)),
  treatment = "IV bolus"
)

conc_obj <- PKNCA::PKNCAconc(as.data.frame(sim_nca), Cc ~ time | treatment + id,
                             concu = "mg/L", timeu = "h")
dose_obj <- PKNCA::PKNCAdose(as.data.frame(dose_df), amt ~ time | treatment + id,
                             doseu = "mg")

intervals <- data.frame(
  start       = 0,
  end         = Inf,
  cmax        = TRUE,
  tmax        = TRUE,
  aucinf.obs  = TRUE,
  aucpext.obs = TRUE,
  half.life   = TRUE
)

res <- PKNCA::pk.nca(PKNCA::PKNCAdata(conc_obj, dose_obj, intervals = intervals))
#> Warning: aucpext is typically only calculated when aucinf is greater than auclast.
#> aucpext is typically only calculated when aucinf is greater than auclast.
#> aucpext is typically only calculated when aucinf is greater than auclast.
#> aucpext is typically only calculated when aucinf is greater than auclast.
#> aucpext is typically only calculated when aucinf is greater than auclast.
#> aucpext is typically only calculated when aucinf is greater than auclast.
#> aucpext is typically only calculated when aucinf is greater than auclast.
nca <- as.data.frame(res)
head(nca)
#> # A tibble: 6 × 8
#>   treatment id                 start   end PPTESTCD  PPORRES exclude PPORRESU
#>   <chr>     <chr>              <dbl> <dbl> <chr>       <dbl> <chr>   <chr>   
#> 1 IV bolus  Chen_2020_imipenem     0   Inf auclast   5.63e+1 <NA>    h*mg/L  
#> 2 IV bolus  Chen_2020_imipenem     0   Inf cmax      2.44e+1 <NA>    mg/L    
#> 3 IV bolus  Chen_2020_imipenem     0   Inf tmax      0       <NA>    h       
#> 4 IV bolus  Chen_2020_imipenem     0   Inf tlast     1.2 e+2 <NA>    h       
#> 5 IV bolus  Chen_2020_imipenem     0   Inf clast.obs 3.64e-8 <NA>    mg/L    
#> 6 IV bolus  Chen_2020_imipenem     0   Inf lambda.z  1.52e-1 <NA>    1/h

Gate 1 – clearance identity `AUC(0, inf) x CL = Dose`

For a linear system the whole dose is eventually cleared, so Dose / AUC(0, inf) recovers total clearance exactly. Comparing that against the clearance printed in Zhang 2025 Table 2 tests the covariate reference constants, the covariate algebra, the unit conversions (albumin g/L to g/dL, postnatal age months to days) and the micro-constant-to-clearance mapping all at once.

aucinf <- nca |>
  dplyr::filter(PPTESTCD == "aucinf.obs") |>
  dplyr::select(model = id, aucinf = PPORRES)
pext <- nca |>
  dplyr::filter(PPTESTCD == "aucpext.obs") |>
  dplyr::select(model = id, aucpext = PPORRES)

cl_check <- aucinf |>
  dplyr::left_join(pext, by = "model") |>
  dplyr::mutate(
    dose     = vapply(cfg[model], function(x) x$dose,   numeric(1)),
    cl_pub   = vapply(cfg[model], function(x) x$cl_pub, numeric(1)),
    cl_nca   = dose / aucinf,
    pct_diff = 100 * (cl_nca - cl_pub) / cl_pub
  ) |>
  dplyr::arrange(model)

knitr::kable(
  cl_check |>
    dplyr::transmute(
      Model            = model,
      `Published CL (L/h)` = round(cl_pub, 4),
      `NCA Dose/AUCinf (L/h)` = round(cl_nca, 4),
      `% difference`   = round(pct_diff, 3),
      `AUC extrapolated (%)` = round(aucpext, 3)
    ),
  caption = "Clearance recovered from the simulated profile against the value printed in Zhang 2025 Table 2."
)

Clearance recovered from the simulated profile against the value printed in Zhang 2025 Table 2.
Model	Published CL (L/h)	NCA Dose/AUCinf (L/h)	% difference
Chen_2020_imipenem	8.8800	8.8792	-0.009
Dao_2022_imipenem	0.2100	0.2100	0.000
Dinh_2022_imipenem	4.7900	4.7896	-0.008
Dong_2019_imipenem	8.6000	8.5987	-0.015
Ikawa_2008_imipenem	9.4200	9.4004	-0.208
Jaruratanasirikul_2021_imipenem	13.3000	13.2970	-0.022
Lafaurie_2023_imipenem_neutropenia	14.3000	14.3000	0.000
Lafaurie_2023_imipenem_recovery	10.9000	10.9000	0.000
Nguyen_2021_imipenem	7.8800	7.8800	0.000
Por_2021_imipenem	15.3100	15.3076	-0.016
Truong_2025_imipenem	14.6000	14.5986	-0.010
Yoshizawa_2013_imipenem_children	7.6140	7.6130	-0.013
Yoshizawa_2013_imipenem_neonates	0.6338	0.6338	0.000
deVelde_2020_imipenem	18.8552	18.8525	-0.014


# The identity is exact in exact arithmetic; the only error is numerical
# (ODE solver tolerance plus trapezoidal AUC on a finite grid plus the
# log-linear extrapolation of the tail). 1% is far outside that numerical
# error and far inside any real transcription defect: a mis-typed
# clearance digit, a wrong covariate reference or a dropped unit
# conversion moves this by tens of percent.
stopifnot(all(abs(cl_check$pct_diff) < 1))
# Guard against the tail being so poorly characterised that aucinf is
# mostly extrapolation, which would make the gate above meaningless.
stopifnot(all(cl_check$aucpext < 5))

Gate 2 – central-volume identity `Cmax = Dose / Vc`

Immediately after an IV bolus the entire dose sits in the central compartment, so the peak concentration is Dose / Vc regardless of how many peripheral compartments the model has.

vc_check <- nca |>
  dplyr::filter(PPTESTCD == "cmax") |>
  dplyr::select(model = id, cmax = PPORRES) |>
  dplyr::mutate(
    dose     = vapply(cfg[model], function(x) x$dose,   numeric(1)),
    vc_pub   = vapply(cfg[model], function(x) x$vc_pub, numeric(1)),
    vc_nca   = dose / cmax,
    pct_diff = 100 * (vc_nca - vc_pub) / vc_pub
  ) |>
  dplyr::arrange(model)

knitr::kable(
  vc_check |>
    dplyr::transmute(
      Model                 = model,
      `Published Vc (L)`    = round(vc_pub, 4),
      `NCA Dose/Cmax (L)`   = round(vc_nca, 4),
      `% difference`        = round(pct_diff, 3)
    ),
  caption = "Central volume recovered from the simulated peak against the value printed in Zhang 2025 Table 2."
)

Central volume recovered from the simulated peak against the value printed in Zhang 2025 Table 2.
Model	Published Vc (L)	NCA Dose/Cmax (L)
Chen_2020_imipenem	20.5000	20.5000
Dao_2022_imipenem	0.7300	0.7300
Dinh_2022_imipenem	11.1000	11.1000
Dong_2019_imipenem	7.2000	7.2000
Ikawa_2008_imipenem	4.6600	4.6600
Jaruratanasirikul_2021_imipenem	13.8000	13.8000
Lafaurie_2023_imipenem_neutropenia	20.7000	20.7000
Lafaurie_2023_imipenem_recovery	14.5000	14.5000
Nguyen_2021_imipenem	15.1000	15.1000
Por_2021_imipenem	32.6700	32.6700
Truong_2025_imipenem	28.7000	28.7000
Yoshizawa_2013_imipenem_children	5.9885	5.9885
Yoshizawa_2013_imipenem_neonates	1.3654	1.3654
deVelde_2020_imipenem	29.6000	29.6000


# 2% rather than 1% here: the peak is read off a finite observation grid,
# so a model with very fast initial distribution (Ikawa 2008, whose
# combined outflow rate constant from central exceeds 10 /h) can lose a
# fraction of a percent between the dose and the first recorded sample.
# A mis-transcribed volume moves this by tens of percent.
stopifnot(all(abs(vc_check$pct_diff) < 2))

Gate 3 – the review’s own cross-study ranges

Zhang 2025 states several numeric ranges in its abstract and Discussion that its authors computed across the constituent studies. Reproducing them from the packaged models checks the set as a whole, independently of any single file. These bounds are quoted from the review, not chosen from an observed run.

# Typical clearance at each model's own reference subject, from Gate 1.
cl_typ <- setNames(cl_check$cl_nca, cl_check$model)

# Zhang 2025 Discussion: "Substantial interindividual variability in PK
# parameters was observed across studies utilizing a two-compartment
# model (CL: 4.79-15.31 L/h; V1: 7.2-32.67 L; V2: 2.9-41.23 L;
# Q: 0.996-24.3 L/h)."
#
# deVelde 2020 is a two-compartment model but is EXCLUDED from this
# range, because it is the one study in the review parameterised in
# micro-constants: Table 2 reports Ke, Kcp, Kpc and Vc and never a
# clearance, so the review's authors had no CL to include. Its implied
# clearance is Ke * Vc = 18.9 L/h, which is reported as a deviation row
# below rather than forced into the range.
two_cmt_here <- c("Chen_2020_imipenem",
                  "Dinh_2022_imipenem", "Dong_2019_imipenem",
                  "Por_2021_imipenem", "Truong_2025_imipenem",
                  "Yoshizawa_2013_imipenem_children")

claims <- tibble::tribble(
  ~Claim, ~`Review's value`, ~Achieved, ~Pass,
  "Lowest adult CL across the review is Dinh 2022 at 4.79 L/h",
  4.79, round(unname(cl_typ["Dinh_2022_imipenem"]), 3),
  isTRUE(all.equal(unname(cl_typ["Dinh_2022_imipenem"]), 4.79, tolerance = 1e-3)),

  "Highest CL among two-compartment models is Por 2021 at 15.31 L/h",
  15.31, round(unname(cl_typ["Por_2021_imipenem"]), 3),
  isTRUE(all.equal(unname(cl_typ["Por_2021_imipenem"]), 15.31, tolerance = 1e-3)),

  "Every two-compartment model's CL lies within 4.79-15.31 L/h",
  NA_real_, round(max(cl_typ[intersect(two_cmt_here, names(cl_typ))]), 3),
  all(cl_typ[intersect(two_cmt_here, names(cl_typ))] >= 4.79 - 1e-3 &
      cl_typ[intersect(two_cmt_here, names(cl_typ))] <= 15.31 + 1e-3),

  "Children's V1 is 0.203 L/kg (Yoshizawa 2013), the review's lower V1 bound",
  0.203, round(unname(cfg$Yoshizawa_2013_imipenem_children$vc_pub) / 29.5, 4),
  isTRUE(all.equal(unname(cfg$Yoshizawa_2013_imipenem_children$vc_pub) / 29.5, 0.203, tolerance = 1e-3)),

  "Burn patients' V1 is 32.67 L (Por 2021), the review's upper V1 bound",
  32.67, round(unname(vc_check$vc_nca[vc_check$model == "Por_2021_imipenem"]), 3),
  isTRUE(all.equal(unname(vc_check$vc_nca[vc_check$model == "Por_2021_imipenem"]), 32.67, tolerance = 2e-3))
)

knitr::kable(claims, caption = "Quantitative claims made by Zhang 2025, reproduced from the packaged models.")

Quantitative claims made by Zhang 2025, reproduced from the packaged models.
Claim	Review’s value	Achieved	Pass
Lowest adult CL across the review is Dinh 2022 at 4.79 L/h	4.790	4.790	TRUE
Highest CL among two-compartment models is Por 2021 at 15.31 L/h	15.310	15.308	TRUE
Every two-compartment model’s CL lies within 4.79-15.31 L/h	NA	15.308	TRUE
Children’s V1 is 0.203 L/kg (Yoshizawa 2013), the review’s lower V1 bound	0.203	0.203	TRUE
Burn patients’ V1 is 32.67 L (Por 2021), the review’s upper V1 bound	32.670	32.670	TRUE

stopifnot(all(claims$Pass))

# Deviation, reported rather than gated: the implied clearance of the one
# micro-constant-parameterised model exceeds every clearance the review
# tabulates.
cl_develde <- unname(cl_typ["deVelde_2020_imipenem"])
knitr::kable(
  tibble::tibble(
    Finding = paste(
      "deVelde 2020 implied CL = Ke x Vc, outside the review's",
      "two-compartment CL range (4.79-15.31 L/h) AND above the",
      "abstract's stated adult maximum of 16.2 L/h"
    ),
    `Implied CL (L/h)` = round(cl_develde, 2)
  ),
  caption = "Known deviation from the review's own summary ranges (not gated; see the text below)."
)

Known deviation from the review’s own summary ranges (not gated; see the text below).
Finding	Implied CL (L/h)
deVelde 2020 implied CL = Ke x Vc, outside the review’s two-compartment CL range (4.79-15.31 L/h) AND above the abstract’s stated adult maximum of 16.2 L/h	18.85

stopifnot(cl_develde > 15.31)  # the deviation is real and reproducible, not marginal

The deviation row is not a transcription defect in this package; it is a property of the review. Zhang 2025’s abstract says imipenem clearance “varied from 4.79 to 16.2 L/h in adults” and its Discussion gives 4.79-15.31 L/h for two-compartment models. Both ranges are computed from the clearances printed in Table 2 – and de Velde 2020 has no clearance in Table 2, because it is the only study in the review parameterised in micro-constants (Ke, Kcp, Kpc, Vc). Its implied clearance, Ke x Vc = 0.637 x 29.6, is 18.9 L/h, higher than every clearance the review tabulates. The review’s headline spread is therefore narrower than its own constituent models support. This is recorded here rather than gated away, and is a further reason to re-verify each model against its primary.

Two of the review’s headline ranges cannot be closed from the models extracted here, because their endpoints belong to studies that were already in the registry or are blocked:

The abstract’s adult clearance range 4.79-16.2 L/h takes its upper endpoint from Lamoth 2009, which is Lamoth_2009_imipenem (extracted from its primary and validated in its own vignette).
The two-compartment Q range 0.996-24.3 L/h takes its upper endpoint from Jaruratanasirikul 2021, which is blocked on the ABW canonical. The lower endpoint, 0.996 L/h, is Dong_2019_imipenem and is reproduced above.

Gate 4 – renal-function gradient

Eight of the 14 models carry a renal-function covariate on clearance, all with a positive coefficient. Clearance must therefore increase monotonically with creatinine clearance in every one of them. A sign error, or a reference constant placed in the wrong position of a ratio, breaks this. Note that the eight are not all on the same renal scale – deVelde_2020_imipenem uses an absolute CKD-EPI eGFR in mL/min while Jaruratanasirikul_2021_imipenem uses the BSA-normalised CKD-EPI eGFR in mL/min/1.73 m^2 – so the grid below is a within-model sweep, not a cross-model comparison of clearance at a common renal function.

renal_models <- c("Chen_2020_imipenem", "deVelde_2020_imipenem",
                  "Dinh_2022_imipenem", "Dong_2019_imipenem",
                  "Jaruratanasirikul_2021_imipenem",
                  "Nguyen_2021_imipenem", "Por_2021_imipenem",
                  "Truong_2025_imipenem")
crcl_grid <- c(15, 30, 60, 90, 120, 150)

renal_cl <- function(nm) {
  spec <- cfg[[nm]]
  mod  <- rxode2::zeroRe(readModelDb(nm))
  vapply(crcl_grid, function(cr) {
    ev <- rxode2::et(amt = spec$dose, cmt = "central", time = 0) |>
      rxode2::et(obs_times)
    ev <- as.data.frame(ev)
    ev$id <- 1L
    cov <- spec$cov
    cov$CRCL <- cr
    for (cv in names(cov)) ev[[cv]] <- cov[[cv]]
    out <- rxode2::rxSolve(mod, events = ev)
    # Dose / AUC recovers CL without reading any model internal.
    spec$dose / PKNCA::pk.calc.auc(out$Cc, out$time, interval = c(0, Inf),
                                  auc.type = "AUCinf",
                                  lambda.z = PKNCA::pk.calc.half.life(
                                    out$Cc, out$time)$lambda.z,
                                  clast = PKNCA::pk.calc.clast.obs(out$Cc, out$time))
  }, numeric(1))
}

renal_tab <- vapply(renal_models, renal_cl, numeric(length(crcl_grid)))
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc'
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalkel'
#> ℹ omega/sigma items treated as zero: 'etalkel'
#> ℹ omega/sigma items treated as zero: 'etalkel'
#> ℹ omega/sigma items treated as zero: 'etalkel'
#> ℹ omega/sigma items treated as zero: 'etalkel'
#> ℹ omega/sigma items treated as zero: 'etalkel'
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc', 'etalvp', 'etalq'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc', 'etalvp', 'etalq'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc', 'etalvp', 'etalq'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc', 'etalvp', 'etalq'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc', 'etalvp', 'etalq'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc', 'etalvp', 'etalq'
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc'
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc', 'etalvp'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc', 'etalvp'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc', 'etalvp'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc', 'etalvp'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc', 'etalvp'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc', 'etalvp'
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc'
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc'
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalcl'
#> ℹ omega/sigma items treated as zero: 'etalcl'
#> ℹ omega/sigma items treated as zero: 'etalcl'
#> ℹ omega/sigma items treated as zero: 'etalcl'
#> ℹ omega/sigma items treated as zero: 'etalcl'
#> ℹ omega/sigma items treated as zero: 'etalcl'
rownames(renal_tab) <- paste0("CrCL ", crcl_grid, " mL/min")
knitr::kable(round(t(renal_tab), 3),
             caption = "Typical clearance (L/h) across creatinine clearance, all other covariates at their reference values.")

Typical clearance (L/h) across creatinine clearance, all other covariates at their reference values.
	CrCL 15 mL/min	CrCL 30 mL/min	CrCL 60 mL/min	CrCL 90 mL/min	CrCL 120 mL/min	CrCL 150 mL/min
Chen_2020_imipenem	5.925	7.270	8.919	10.052	10.942	11.687
deVelde_2020_imipenem	4.856	7.646	12.039	15.701	18.956	21.939
Dinh_2022_imipenem	5.274	5.807	7.040	8.535	10.348	12.545
Dong_2019_imipenem	2.223	3.163	4.501	5.533	6.406	7.176
Jaruratanasirikul_2021_imipenem	5.012	6.691	10.050	13.409	16.767	20.125
Nguyen_2021_imipenem	3.334	4.821	6.971	8.650	10.080	11.351
Por_2021_imipenem	5.377	7.397	10.174	12.260	13.995	15.507
Truong_2025_imipenem	6.460	8.898	12.257	14.782	16.883	18.717


# Strictly increasing in every model. This is a deterministic solve of a
# monotone algebraic expression, not a noisy cohort statistic, so exact
# monotonicity is the right assertion here (contrast the cohort-derived
# quantities warned about in the failure-pattern catalogue).
stopifnot(all(apply(renal_tab, 2, function(x) all(diff(x) > 0))))

Gate 5 – steady-state exposure and the review’s PK/PD index

Zhang 2025 states that imipenem is time-dependent and that the relevant index is %fT>MIC. The table below reports steady-state exposure for the adult models on a common regimen, 1000 mg as a 0.5 h infusion every 8 h, with %T>MIC computed on total concentration at an MIC of 2 mg/L (the EUCAST Pseudomonas aeruginosa breakpoint). This is descriptive rather than a gate against a published table – the review reports no per-model NCA values, so there is nothing to compare against – but the spread is the review’s central finding, made quantitative.

adults_ss <- c("Chen_2020_imipenem", "deVelde_2020_imipenem",
               "Dinh_2022_imipenem", "Ikawa_2008_imipenem",
               "Jaruratanasirikul_2021_imipenem",
               "Lafaurie_2023_imipenem_neutropenia",
               "Lafaurie_2023_imipenem_recovery",
               "Nguyen_2021_imipenem", "Por_2021_imipenem",
               "Truong_2025_imipenem")
tau <- 8
mic <- 2

ss_one <- function(nm) {
  spec <- cfg[[nm]]
  mod  <- rxode2::zeroRe(readModelDb(nm))
  ev <- rxode2::et(amt = 1000, cmt = "central", dur = 0.5, time = 0,
                   ii = tau, addl = 14) |>
    rxode2::et(seq(0, 15 * tau, by = 0.02))
  ev <- as.data.frame(ev)
  ev$id <- 1L
  for (cv in names(spec$cov)) ev[[cv]] <- spec$cov[[cv]]
  out <- rxode2::rxSolve(mod, events = ev)
  # Last complete dosing interval = steady state.
  ss <- out[out$time >= 14 * tau & out$time <= 15 * tau, ]
  data.frame(
    Model        = nm,
    `Cmax (mg/L)` = round(max(ss$Cc), 2),
    `Cmin (mg/L)` = round(min(ss$Cc), 3),
    `T>MIC (%)`   = round(100 * mean(ss$Cc > mic), 1),
    check.names   = FALSE
  )
}

ss_tab <- dplyr::bind_rows(lapply(adults_ss, ss_one))
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc'
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalkel'
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc', 'etalvp', 'etalq'
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc', 'etalvp', 'etalvp2'
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc', 'etalvp'
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc'
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc'
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc'
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc'
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalcl'
knitr::kable(ss_tab,
             caption = "Steady-state exposure on 1000 mg q8h (0.5 h infusion), typical subject, MIC 2 mg/L.")

Steady-state exposure on 1000 mg q8h (0.5 h infusion), typical subject, MIC 2 mg/L.
Model	Cmax (mg/L)	Cmin (mg/L)	T>MIC (%)
Chen_2020_imipenem	45.63	2.937	100.0
deVelde_2020_imipenem	28.92	1.149	67.3
Dinh_2022_imipenem	75.43	9.146	100.0
Ikawa_2008_imipenem	79.55	0.784	76.3
Jaruratanasirikul_2021_imipenem	44.19	1.418	87.8
Lafaurie_2023_imipenem_neutropenia	41.01	0.231	60.6
Lafaurie_2023_imipenem_recovery	57.63	0.205	61.8
Nguyen_2021_imipenem	59.20	1.182	87.0
Por_2021_imipenem	27.88	2.778	100.0
Truong_2025_imipenem	31.50	1.582	85.8


# Sanity, not tuning: on 3 g/day every adult model must keep a typical
# subject above a 2 mg/L MIC for some but not all of the interval, and
# must produce a physically ordered profile. A model whose clearance or
# volume was transcribed wrong by an order of magnitude fails this.
stopifnot(all(ss_tab$`Cmax (mg/L)` > ss_tab$`Cmin (mg/L)`))
stopifnot(all(ss_tab$`Cmax (mg/L)` > mic))
stopifnot(all(ss_tab$`T>MIC (%)` >= 20, ss_tab$`T>MIC (%)` <= 100))

Assumptions and deviations

Transcription fidelity of the secondary source – measured, not assumed

Four of the review’s 18 studies were already in nlmixr2lib, extracted from their primary publications: Lamoth_2009_imipenem, Couffignal_2014_imipenem, Bai_2024_imipenem and Wang_2024_imipenem. They form a validation set for the review’s accuracy, and the result decides how much weight the other 14 models can carry.

Structural parameters and covariate equations: 4 of 4 agree exactly. The review’s Table 2 typical values and Table 3 covariate formulations match the corresponding primary-sourced model files digit for digit – including Lamoth’s additive split clearance CL = 10.7 + 4.79 x eGFR/100 with V = 33.5 x BW/70, Couffignal’s three-exponent V1 = 20.4 x (BW/77)^1.3 x (ALB/18)^-1.1, Bai’s CL = 11.357 x (CrCl/99.896)^0.473 and Wang’s CL = 13.1 x (CLCR/71)^0.263. The review transcribes structure and covariate models reliably.

Inter-individual variability: the column is NOT on a single scale, and at least one entry is wrong. The review prints IIV as a bare percentage under the heading “Inter-individual variability (IIV)” without ever stating the convention, and it reproduces each primary’s own reported quantity without harmonising. Checked against the four primary-sourced files:

Study	Review Table 2	What the primary actually reports	Convention
Lamoth 2009	`CL: 17%`	“17% +/- 6% proportional IIV”	apparent CV
Couffignal 2014	`CL: 38%`, `V1: 31%`	Monolix `omega CL (%)`	omega SD as a percent
Bai 2024	`CL: 35.748%`	Table 2 column headed “Iiv (CV%)”	apparent CV
Wang 2024	`CL: 8.32%`	Table 2 row `omega^2 CL = 0.0832`	a VARIANCE, mis-rendered as a percent

The Wang 2024 row is a genuine defect in the review: 0.0832 is a variance, corresponding to an apparent CV of 29.5%, and the review prints it as “8.32%” – understating that model’s between-subject variability roughly three-and-a-half-fold. It was caught only because that study’s primary is independently in this package.

What this package does about it. Every model transcribed from this review adopts one uniform, documented reading: the printed percentage is an apparent CV of a log-normal random effect, so omega^2 = log(1 + CV^2). That reading is:

exact for a source reporting NONMEM-style CV% (Lamoth, Bai above, and by inference most of the NONMEM studies in the review);
modestly low for a source reporting an omega SD as a percent (the Monolix studies – Couffignal, and hence probably also Dinh 2022, Lafaurie 2023 and Nguyen 2021). At CV 30% the two readings differ by 4% in variance; the gap widens with CV, so the largest exposure is Dinh_2022_imipenem’s 110% CV on Q, where the readings differ by more than 50%;
badly wrong for a row that is really a variance – but the only known such row, Wang 2024, is already extracted from its primary and is therefore not among the 14 models here.

The alternative – blocking all 14 models until every primary is obtained – was rejected because the per-study answer requires exactly the primaries that the “re-verify against the primary” notes already flag. Treat every IIV magnitude in this family as provisional. The structural parameters, covariate equations and residual errors are on much firmer ground.

Per-model deviations and warnings

Chen_2020_imipenem – residual error is implausibly small. The review reports 6.2% proportional plus 0.003 mg/L additive for a sparse routine-TDM dataset of 580 samples in 247 critically ill patients. It is by a wide margin the smallest residual error among the 18 studies. Transcribed verbatim, but a user simulating residual-error-bearing observations from this model should treat the noise level as unverified.
Chen_2020_imipenem – the ECMO coefficient is very large. e^1.16 = 3.19, i.e. ECMO more than triples clearance, for a mechanism the review’s own Discussion attributes mainly to an increased volume of distribution via circuit adsorption. No standard error is reported. Re-verify before using this model for ECMO dosing.
Por_2021_imipenem – the albumin exponent on peripheral volume is extreme. -3.68 means a halving of albumin multiplies Vp by 12.8. For a 23-patient cohort this is almost certainly poorly identified. The review reports neither a standard error nor the observed albumin range, so the fitted domain is unknown.
Por_2021_imipenem – albumin unit conversion. The reference 2.7 is g/dL (a US study; 2.7 g/L would be incompatible with life). The canonical ALB column is SI g/L, so model() converts inline with alb_gdL <- ALB * 0.1, per the register’s instruction for models calibrated on g/dL.
Dao_2022_imipenem – postnatal age unit reparameterisation. The source centres PNA on 21 days; the canonical PNA column is months, so model() recovers days with PNA * 30.4375 before forming the centred term, leaving the printed constants untouched. Same pattern as the registered Zhao 2018 and Bardhi 2026 entries.
Dao_2022_imipenem – volume exponent 0.75. The conventional allometric pairing is 0.75 on clearance and 1 on volume; this model uses 0.75 on both, as printed. Reproduced verbatim rather than “corrected”, but a transcription slip in the secondary source is a live alternative.
Dao_2022_imipenem – the two age covariates are centred LINEAR terms, not power terms. The gestational-age factor reaches zero at 6.4 weeks and goes negative below that; the postnatal-age factor reaches zero at -74.5 days. Both are outside any physical domain, but the model will not error if given nonsense – it will return a nonsense clearance.
Dinh_2022_imipenem – uncentred exponential covariate. The published typical clearance of 4.79 L/h is the value at CrCL = 0 mL/min, not at any cohort-typical renal function. At a normal 100 mL/min the model gives 9.11 L/h. The review’s abstract quotes 4.79 L/h as the lowest adult clearance it found, which is therefore an extrapolated intercept rather than any patient’s clearance.
Dinh_2022_imipenem – unit typo in the source. Table 2 prints V1: 11.1 h. Encoded as 11.1 L; a volume cannot be in hours and the sibling entries in the same cell are in L and L/h.
Yoshizawa_2013_imipenem_* – footnote erratum. The Table 2 footnote glosses the clearance symbols as “CLr, clearance of plasma drug; CLnr, clearance of urine drug”, which is incoherent. Read as renal and non-renal clearance. Supporting evidence: the renal arm dominates in children (0.187 vs 0.0711 L/h/kg) while the non-renal arm dominates in neonates (0.138 vs 0.0783), which is the direction renal maturation predicts. Only the sum enters the plasma model, so this affects the interpretation of the two etas, not the predictions.
Yoshizawa_2013_imipenem_* – linear per-kg scaling. All parameters are reported per kilogram, i.e. weight exponent 1 on clearance as well as volume, unlike the allometric 0.75 used by the other paediatric models here.
deVelde_2020_imipenem – only the NONMEM arm is encoded. The study also fitted the data non-parametrically in Pmetrics. That arm’s covariate model is printed in terms of per-subject posterior medians, which are not population parameters; and a non-parametric population is a discrete support-point distribution, not the log-normal random effect ini() expresses. It cannot be faithfully encoded even in principle. The review’s Pmetrics “Gamma: 3.4” is an error-polynomial multiplier, not a residual SD.
deVelde_2020_imipenem – absolute, not BSA-normalised, eGFR. The covariate is CKD-EPI multiplied by BSA, in mL/min. Supplying a standard mL/min/1.73 m^2 eGFR understates the ratio for anyone larger than 1.73 m^2.
Dong_2019_imipenem – allometric exponents left unwrapped. The 0.75 on clearance and 1 on volume may well have been fixed rather than estimated, but the review reports no standard errors and does not say, so they are not wrapped in fixed(). The same applies to Dao_2022_imipenem.
Truong_2025_imipenem – dose not reported. The review records “NR” for this study’s dosage and says so explicitly in its Results. The steady-state table above therefore assumes a regimen for this model rather than reproducing the study’s own.
Truong_2025_imipenem – all-male cohort. Table 1 records the sex split as 151/0. Unusual enough to warrant re-checking against the primary; it would make sex unidentifiable by construction rather than merely non-significant.
Jaruratanasirikul_2021_imipenem – the source disagrees with itself about the central volume. Table 2 prints V1: 13.6 L, while the Table 3 covariate formula reads V1 = 13.8 - 0.348 x (ABW - 60), i.e. an intercept of 13.8 L at the reference ABW of 60 kg. The formula intercept is encoded, because it is the quantity the covariate term is conditioned on: pairing the Table 2 value with the Table 3 slope would shift every predicted volume by a constant 0.2 L and break the reference-subject identity that Gate 2 tests. The discrepancy is 1.5%. Clearance has no such conflict – Table 2 and the Table 3 intercept agree exactly at 13.3 L/h – which is mild evidence that the 13.6 is the slip.
Jaruratanasirikul_2021_imipenem – the volume term has a bounded domain and the model will not tell you when you leave it. V1 is linear in ABW with a negative slope, so it falls to zero at ABW = 60 + 13.8/0.348 = 99.7 kg and is negative above that. The study’s own cohort is light (median total body weight 62.9 kg, IQR 52.8-70.0), so the term was fitted far from the zero-crossing; do not simulate this model at obese adjusted body weights. The clearance term is linear too but with a positive slope, so it has no reachable zero-crossing. This is the same failure shape as the two centred linear age terms in Dao_2022_imipenem noted above.
Jaruratanasirikul_2021_imipenem – ABW is a newly ratified canonical, and its formula is not in the source. ABW (adjusted body weight) was added to inst/references/covariate-columns.md with this extraction as its founding example. Neither the review nor its Table 3 footnote – which glosses only “ABW, adjusted body weight” – states which ideal-body-weight formula or which interpolation factor the primary used. The conventional IBW + 0.4 x (TBW - IBW) is recorded in the register and in the model file as the usual convention, explicitly not as a value read from this source. A user driving this model must supply ABW on whatever scale the primary used, which means the primary is needed before the covariate can be reproduced exactly.
Jaruratanasirikul_2021_imipenem – the most permissive covariate selection thresholds in the review. Forward inclusion at dOFV > 3.84 (p < 0.05) and backward elimination at dOFV > 6.64 (p < 0.01), against the 10.828 (p < 0.001) backward threshold used by, for example, Nguyen 2021. Both retained effects should be read with that in mind. The study’s headline negative result – that ECMO support, ECMO type, ECMO flow rate and ECMO duration were all screened and none retained – is correspondingly stronger, since it survived the easiest threshold to pass.
Lafaurie_2023_imipenem_* – two files, not one. The review reports the neutropenic and recovered phases with separate structural values, separate IIV and separate residual error. A per-phase residual error is the signature of two independent fits rather than one model carrying a phase covariate, so they are shipped as two files.

What is not validated here

No published NCA comparison. The review reports no Cmax, Tmax, AUC or half-life for any constituent study, so nlmixr2lib::ncaComparisonTable() has nothing to compare against. Gates 1 and 2 are closed-form identities instead; Gate 3 reproduces the review’s own cross-study claims; Gate 5 is descriptive.
No observed data and no VPC. No study’s individual concentrations are available, and the review reproduces no goodness-of-fit or VPC figure in a digitisable form.
The gates test the packaged encoding, not the review’s arithmetic. Gates 1 and 2 confirm that each model’s ODE system, covariate algebra and unit conversions deliver the clearance and volume printed in Table 2 at the reference subject. They cannot confirm that Table 2 itself faithfully reproduces the primary – only re-reading each primary can, which is what the “re-verify against the primary” notes ask for. The fidelity audit above is the closest available substitute, and it found the IIV-scale defect described there.