Skip to contents

Model and source

mod_meta <- nlmixr2est::nlmixr(readModelDb("Xu_2024_polymyxinB"))$meta
#> ℹ parameter labels from comments will be replaced by 'label()'
  • Citation: Xu M, Chen N, Yu YW, Pan XY, Li T. Pharmacokinetic Changes and Influencing Factors of Polymyxin B in Different ECMO Modes. Infect Drug Resist. 2024;17:5815-5825. doi:10.2147/IDR.S486169
  • Description: Two-compartment IV infusion population PK model for polymyxin B (summed PMB1 + PMB2) in 11 critically ill adults on venoarterial or venovenous ECMO with carbapenem-resistant Gram-negative infections (Xu 2024). Clearance is 3.75 L/h at an ECMO blood flow of 3.48 L/min and falls linearly by 32.1% per L/min of additional flow, CL = 3.75 * (1 - 0.321 * (Q_ECMO - 3.48)). The central volume is 20.41 L in a patient receiving CRRT and 52.5% smaller (9.69 L) without CRRT. Peripheral volume 9.86 L and intercompartmental clearance 3.82 L/h carry no covariates. Exponential IIV on CL (omega^2 = 0.038) and on the central volume (omega^2 = 0.062); additive residual error 0.361 mg/L.
  • Article (DOI): https://doi.org/10.2147/IDR.S486169

This vignette validates the packaged Xu_2024_polymyxinB model, a two-compartment population PK model for polymyxin B (PMB1 + PMB2 summed) developed in Phoenix NLME from 132 steady-state plasma samples in 11 critically ill adults supported by venoarterial (8) or venovenous (3) extracorporeal membrane oxygenation (ECMO). Two covariates were retained: the ECMO blood flow rate, which enters clearance as a linear centred term, and continuous renal replacement therapy (CRRT), which enters the central volume.

mod <- readModelDb("Xu_2024_polymyxinB")
mod
#> function() {
#>   description <- "Two-compartment IV infusion population PK model for polymyxin B (summed PMB1 + PMB2) in 11 critically ill adults on venoarterial or venovenous ECMO with carbapenem-resistant Gram-negative infections (Xu 2024). Clearance is 3.75 L/h at an ECMO blood flow of 3.48 L/min and falls linearly by 32.1% per L/min of additional flow, CL = 3.75 * (1 - 0.321 * (Q_ECMO - 3.48)). The central volume is 20.41 L in a patient receiving CRRT and 52.5% smaller (9.69 L) without CRRT. Peripheral volume 9.86 L and intercompartmental clearance 3.82 L/h carry no covariates. Exponential IIV on CL (omega^2 = 0.038) and on the central volume (omega^2 = 0.062); additive residual error 0.361 mg/L."
#>   reference <- "Xu M, Chen N, Yu YW, Pan XY, Li T. Pharmacokinetic Changes and Influencing Factors of Polymyxin B in Different ECMO Modes. Infect Drug Resist. 2024;17:5815-5825. doi:10.2147/IDR.S486169"
#>   vignette <- "Xu_2024_polymyxinB"
#>   units <- list(time = "h", dosing = "mg", concentration = "mg/L")
#> 
#>   # Total polymyxin B (PMB1 + PMB2 summed) in plasma by LC-MS/MS (Xu 2024
#>   # Drug Assay section); the peripheral compartment is the model's tissue
#>   # distribution space.
#>   compartmentData <- list(
#>     central = list(analyte = "polymyxinB", units = "mg", specimen = "plasma", verified = TRUE),
#>     peripheral1 = list(analyte = "polymyxinB", units = "mg", specimen = "tissue", verified = TRUE)
#>   )
#> 
#>   covariateData <- list(
#>     Q_ECMO = list(
#>       description = "ECMO circuit blood flow rate",
#>       units = "L/min",
#>       type = "continuous",
#>       reference_category = NULL,
#>       notes = "Source covariate 'speadECMO' in the final-model CL equation, defined in the Table 2 footnote as 'spead of ECMO (L/min)'; the L/min unit identifies it as the delivered blood flow rather than the pump speed (Table 1 reports pump speed separately in RPM, median 3500). Table 1 median (IQR) 3.60 (3.00, 3.65) L/min. Enters CL as the linear centred form 1 + dCLdspeedECMO * (Q_ECMO - 3.48), with 3.48 L/min the centring value printed in the equation (not the Table 1 median of 3.60). Treated as a per-subject value at the steady-state sampling occasion; the paper does not describe flow as time-varying. Because the effect is linear with a negative slope (-0.321 per L/min), predicted CL reaches zero at Q_ECMO = 3.48 + 1/0.321 = 6.60 L/min; do not extrapolate far outside the observed 3.0-3.65 L/min interquartile range.",
#>       source_name = "speadECMO"
#>     ),
#>     RRT_CRRT_STATUS = list(
#>       description = "Subject-level binary indicator for continuous renal replacement therapy during the PK sampling period",
#>       units = "(binary)",
#>       type = "binary",
#>       reference_category = 0,
#>       notes = "Source covariate CRRT. Table 1: 8/11 (72.73%) received CRRT, all of it continuous venovenous haemodialysis (CVVHD; Ultraflux AV1000S) for anuria, connected in the ECMO circuit (outflow at the post-oxygenator, inflow at the pre-oxygenator line). The paper estimates the effect in the Phoenix 'dVdCRRT0' parameterization, V = tvV * (1 + dVdCRRT0 * (if no CRRT)), so the typical central volume of 20.41 L is the CRRT value and the -0.525 coefficient applies to patients WITHOUT CRRT. Encoded here as (1 + e_nocrrt_vc * (1 - RRT_CRRT_STATUS)). Direction: CRRT is associated with a larger central volume, as the Discussion states ('patients receiving CRRT-ECMO had significantly increased the central compartment').",
#>       source_name = "CRRT"
#>     )
#>   )
#> 
#>   covariatesDataExcluded <- list(
#>     AGE = list(
#>       description = "Subject age",
#>       units = "years",
#>       type = "continuous",
#>       notes = "The Results text says 'CRRT and age were identified as the covariates for V', and the Table 2 abbreviation list defines 'dVdage, fixed parameter coefficient of age (year) to V', but neither the printed final-model V equation nor the Table 2 parameter rows contain an age term or estimate. The printed equation is taken as the final model, so age is not carried. Table 1 median (IQR) 56 (48, 64) years; the Results prose gives 60 (53.50-62.50)."
#>     )
#>   )
#> 
#>   population <- list(
#>     species = "human",
#>     n_subjects = 11L,
#>     n_studies = 1L,
#>     age_range = "Adults (>= 18 years); Table 1 median (IQR) 56 (48, 64) years (the Results prose reports 60, 53.50-62.50)",
#>     weight_range = "Median (IQR) 60 (57.05, 66.00) kg; BSA 1.69 (1.68, 1.79) m^2 (Table 1)",
#>     sex_female_pct = 27.27,
#>     race_ethnicity = "Not reported (single-centre Chinese ICU cohort)",
#>     disease_state = "Heart or respiratory failure requiring ECMO (8 VA-ECMO, 3 VV-ECMO; SOFA >= 2) with infection by carbapenem-resistant A. baumannii (7), K. pneumoniae (6) or P. aeruginosa (3), polymyxin B MIC 0.5 mg/L. Pneumonia 9/11, bloodstream infection 4/11. Median APACHE II 14, SOFA 10. ECMO duration median 16.18 days.",
#>     dose_range = "Loading dose 200-300 mg on day 1, then 100-150 mg (2.5-3 mg/kg/day) at the treating physician's discretion, each dose in 50 mL saline as a 1-h infusion. Sampling at steady state (after more than four doses) at 1, 1.5, 2, 4, 8 and 12 h after the start of infusion, pre- and post-oxygenator.",
#>     regions = "China (First Affiliated Hospital of Zhejiang University School of Medicine; single centre)",
#>     renal_function = "8/11 on CVVHD for anuria. Table 1 CrCL median (IQR) 79.70 (33.33, 94.89) mL/min; serum creatinine 137.27 (71.00, 184.50), labelled mg/dL but clearly umol/L.",
#>     ecmo = "Rotaflow centrifugal pump with polymethylpentene oxygenator (MAQUET), PVC tubing, 600 mL saline prime. Pump speed median 3500 (3000, 3770) RPM; blood flow 3.60 (3.00, 3.65) L/min.",
#>     notes = "Prospective study, January-December 2021. 132 plasma samples (66 pre- and 66 post-oxygenator; no significant difference between sites, Figure 1). Estimation: Phoenix NLME 8.1, FOCE; pcVPC with 1000 replicates and 1000-sample bootstrap."
#>   )
#> 
#>   ini({
#>     # Structural parameters: Xu 2024 Table 2, 'Final Model Estimate, Mean'
#>     # column. The reference patient receives CRRT and has an ECMO blood flow
#>     # of 3.48 L/min (final-model CL and V equations).
#>     lcl <- log(3.75); label("Clearance at Q_ECMO = 3.48 L/min (L/h)") # Table 2: tvCL = 3.75 L/h (CV 6.98%; bootstrap median 3.74, 95% CI 1.59-4.34)
#>     lvc <- log(20.41); label("Central volume of distribution with CRRT (L)") # Table 2: tvV = 20.41 L (CV 17.1%; bootstrap median 20.47, 95% CI 12.82-27.89)
#>     lvp <- log(9.86); label("Peripheral volume of distribution (L)") # Table 2: tvVp = 9.86 L (CV 53.4%; bootstrap median 10.34, 95% CI 4.06-24.34)
#>     lq <- log(3.82); label("Intercompartmental clearance (L/h)") # Table 2: tvQ = 3.82 L/h (CV 42.1%; bootstrap median 4.19, 95% CI 1.83-10.49)
#> 
#>     # Covariate effects (final-model equations printed in the Results)
#>     e_q_ecmo_cl <- -0.321; label("Fractional change in CL per L/min of ECMO blood flow above 3.48 L/min (1/(L/min))") # Table 2: dCLdspeedECMO = -0.321 (bootstrap median -0.315, 95% CI -0.710 to 0.00120); CL equation: tvCL * (1 + (speadECMO - 3.48 L/min) * dCLdspeadECMO)
#>     e_nocrrt_vc <- -0.525; label("Fractional change in central volume without CRRT (unitless)") # Table 2: dVdCRRT0 = -0.525 (bootstrap median -0.553, 95% CI -0.769 to -0.207); V equation: tvV * (1 + dVdCRRTO * (if no CRRT))
#> 
#>     # Between-subject variability: Table 2 rows 'omega2 V' and 'omega2 CL',
#>     # footnoted as variances; exponential IIV per Methods. No covariance
#>     # term is reported.
#>     etalcl ~ 0.038 # Table 2: omega2 CL = 0.038 (bootstrap median 0.049)
#>     etalvc ~ 0.062 # Table 2: omega2 V = 0.062 (bootstrap median 0.058)
#> 
#>     # Residual variability: Table 2 'stdev0' = 0.361. The Results select 'the
#>     # two-compartment model with an additive option', and the Figure 3 pcVPC
#>     # lower prediction band dips below zero, consistent with an additive
#>     # (not proportional) error in mg/L.
#>     addSd <- 0.361; label("Additive residual error (mg/L)") # Table 2: stdev0 = 0.361 (CV 12.0%; bootstrap median 0.351, 95% CI 0.256-0.447)
#>   })
#>   model({
#>     # Individual parameters (Xu 2024 final-model equations)
#>     cl <- exp(lcl + etalcl) * (1 + e_q_ecmo_cl * (Q_ECMO - 3.48))
#>     vc <- exp(lvc + etalvc) * (1 + e_nocrrt_vc * (1 - RRT_CRRT_STATUS))
#>     vp <- exp(lvp)
#>     q <- exp(lq)
#> 
#>     kel <- cl / vc
#>     k12 <- q / vc
#>     k21 <- q / vp
#> 
#>     # Two-compartment disposition; polymyxin B given as a 1-h IV infusion
#>     # into the central compartment. Dose in mg and volume in L give mg/L.
#>     d/dt(central) <- -kel * central - k12 * central + k21 * peripheral1
#>     d/dt(peripheral1) <- k12 * central - k21 * peripheral1
#> 
#>     Cc <- central / vc
#>     Cc ~ add(addSd)
#>   })
#> }
#> <environment: 0x55564a1df4a0>

Population

tibble::tribble(
  ~Characteristic,             ~Value,
  "Subjects",                  "11 (132 plasma samples, pre- and post-oxygenator)",
  "Sex",                       "8 male (72.73%), 3 female (27.27%)",
  "Age (years)",               "56 (48, 64)",
  "Weight (kg)",               "60 (57.05, 66.00)",
  "BSA (m^2)",                 "1.69 (1.68, 1.79)",
  "CrCL (mL/min)",             "79.70 (33.33, 94.89)",
  "Albumin (g/L)",             "34.00 (30.55, 36.20)",
  "CRRT",                      "8 (72.73%), all CVVHD",
  "ECMO mode (VA / VV)",       "8 / 3",
  "ECMO pump speed (RPM)",     "3500 (3000, 3770)",
  "ECMO blood flow (L/min)",   "3.60 (3.00, 3.65)",
  "ECMO duration (days)",      "16.18 (8.50, 23.00)",
  "APACHE II / SOFA",          "14 (9.50, 18.00) / 10 (4.00, 11.50)"
) |>
  knitr::kable(caption = "Xu 2024 Table 1; continuous values are median (IQR).")
Xu 2024 Table 1; continuous values are median (IQR).
Characteristic Value
Subjects 11 (132 plasma samples, pre- and post-oxygenator)
Sex 8 male (72.73%), 3 female (27.27%)
Age (years) 56 (48, 64)
Weight (kg) 60 (57.05, 66.00)
BSA (m^2) 1.69 (1.68, 1.79)
CrCL (mL/min) 79.70 (33.33, 94.89)
Albumin (g/L) 34.00 (30.55, 36.20)
CRRT 8 (72.73%), all CVVHD
ECMO mode (VA / VV) 8 / 3
ECMO pump speed (RPM) 3500 (3000, 3770)
ECMO blood flow (L/min) 3.60 (3.00, 3.65)
ECMO duration (days) 16.18 (8.50, 23.00)
APACHE II / SOFA 14 (9.50, 18.00) / 10 (4.00, 11.50)

All patients had carbapenem-resistant A. baumannii, K. pneumoniae or P. aeruginosa infection (polymyxin B MIC 0.5 mg/L). They received a 200-300 mg loading dose on day 1, then 100-150 mg (2.5-3 mg/kg/day), each dose as a 1-hour infusion. Samples were drawn at steady state (after more than four doses) at 1, 1.5, 2, 4, 8 and 12 h after the start of an infusion. Pre- and post-oxygenator concentrations did not differ (Figure 1), so the two sampling sites are pooled in the model.

Source trace

tibble::tribble(
  ~Quantity,                    ~`Model file`,     ~Value,          ~`Source location`,
  "Clearance (Q_ECMO 3.48)",    "lcl",             "3.75 L/h",      "Table 2 tvCL; CL equation",
  "Central volume (CRRT)",      "lvc",             "20.41 L",       "Table 2 tvV; V equation",
  "Peripheral volume",          "lvp",             "9.86 L",        "Table 2 tvVp",
  "Intercompartmental CL",      "lq",              "3.82 L/h",      "Table 2 tvQ",
  "ECMO flow slope on CL",      "e_q_ecmo_cl",     "-0.321 per L/min", "Table 2 dCLdspeedECMO; CL equation",
  "No-CRRT effect on V",        "e_nocrrt_vc",     "-0.525",        "Table 2 dVdCRRT0; V equation",
  "IIV on CL (variance)",       "etalcl",          "0.038",         "Table 2 omega2 CL",
  "IIV on V (variance)",        "etalvc",          "0.062",         "Table 2 omega2 V",
  "Additive residual SD",       "addSd",           "0.361 mg/L",    "Table 2 stdev0; Results 'additive option'",
  "Structural model",           "d/dt(central), d/dt(peripheral1)", "2-cmt", "Results, OFV 109.9 vs 126.6 for 1-cmt",
  "Flow centring value",        "3.48",            "3.48 L/min",    "CL equation",
  "Infusion duration",          "dur = 1",         "1 h",           "Methods, Drug Regimens"
) |>
  knitr::kable(caption = "Source trace for every ini() value and structural choice.")
Source trace for every ini() value and structural choice.
Quantity Model file Value Source location
Clearance (Q_ECMO 3.48) lcl 3.75 L/h Table 2 tvCL; CL equation
Central volume (CRRT) lvc 20.41 L Table 2 tvV; V equation
Peripheral volume lvp 9.86 L Table 2 tvVp
Intercompartmental CL lq 3.82 L/h Table 2 tvQ
ECMO flow slope on CL e_q_ecmo_cl -0.321 per L/min Table 2 dCLdspeedECMO; CL equation
No-CRRT effect on V e_nocrrt_vc -0.525 Table 2 dVdCRRT0; V equation
IIV on CL (variance) etalcl 0.038 Table 2 omega2 CL
IIV on V (variance) etalvc 0.062 Table 2 omega2 V
Additive residual SD addSd 0.361 mg/L Table 2 stdev0; Results ‘additive option’
Structural model d/dt(central), d/dt(peripheral1) 2-cmt Results, OFV 109.9 vs 126.6 for 1-cmt
Flow centring value 3.48 3.48 L/min CL equation
Infusion duration dur = 1 1 h Methods, Drug Regimens

The final-model equations printed in the Results are

CL = tvCL x (1 + (speadECMO - 3.48 L/min) x dCLdspeadECMO) x exp(nCL)
V  = tvV  x (1 + dVdCRRTO x (if no CRRT)) x exp(nV)

In this Phoenix parameterization the typical central volume of 20.41 L belongs to a patient on CRRT and the -0.525 coefficient applies to patients without CRRT, so CRRT is associated with a larger central volume, as the Discussion states.

Typical parameter values by covariate

typ <- tidyr::expand_grid(
  RRT_CRRT_STATUS = c(1, 0),
  Q_ECMO = c(3.0, 3.48, 3.6, 4.0, 5.0)
) |>
  mutate(
    `CL (L/h)` = 3.75 * (1 - 0.321 * (Q_ECMO - 3.48)),
    `V (L)` = 20.41 * (1 - 0.525 * (1 - RRT_CRRT_STATUS)),
    CRRT = ifelse(RRT_CRRT_STATUS == 1, "CRRT", "no CRRT")
  )
typ |>
  select(CRRT, `ECMO flow (L/min)` = Q_ECMO, `CL (L/h)`, `V (L)`) |>
  knitr::kable(digits = 2, caption = "Typical CL and central volume across the covariate range.")
Typical CL and central volume across the covariate range.
CRRT ECMO flow (L/min) CL (L/h) V (L)
CRRT 3.00 4.33 20.41
CRRT 3.48 3.75 20.41
CRRT 3.60 3.61 20.41
CRRT 4.00 3.12 20.41
CRRT 5.00 1.92 20.41
no CRRT 3.00 4.33 9.69
no CRRT 3.48 3.75 9.69
no CRRT 3.60 3.61 9.69
no CRRT 4.00 3.12 9.69
no CRRT 5.00 1.92 9.69

The packaged model must reproduce these closed-form typical values exactly. zeroRe() removes the random effects, so the comparison is pure arithmetic.

mod_typ <- zeroRe(mod)
#> ℹ parameter labels from comments will be replaced by 'label()'
typ_ev <- typ |>
  mutate(id = row_number(), time = 0, amt = 0, evid = 0, cmt = "central") |>
  select(id, time, amt, evid, cmt, Q_ECMO, RRT_CRRT_STATUS)
typ_sim <- rxSolve(mod_typ, typ_ev, returnType = "data.frame", addDosing = FALSE)
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc'
#> Warning: multi-subject simulation without without 'omega'
stopifnot(
  max(abs(typ_sim$cl - typ$`CL (L/h)`)) < 1e-8,
  max(abs(typ_sim$vc - typ$`V (L)`)) < 1e-8
)
flow_grid <- tibble(Q_ECMO = seq(2.5, 5, by = 0.05)) |>
  mutate(cl = 3.75 * (1 - 0.321 * (Q_ECMO - 3.48)))
ggplot(flow_grid, aes(Q_ECMO, cl)) +
  annotate("rect", xmin = 3.0, xmax = 3.65, ymin = -Inf, ymax = Inf, alpha = 0.15) +
  geom_line() +
  geom_point(data = tibble(Q_ECMO = 3.48, cl = 3.75), colour = "red") +
  labs(x = "ECMO blood flow (L/min)", y = "Typical CL (L/h)")
Typical clearance against ECMO blood flow (Xu 2024 CL equation). The shaded band is the Table 1 interquartile range of flow (3.00-3.65 L/min).

Typical clearance against ECMO blood flow (Xu 2024 CL equation). The shaded band is the Table 1 interquartile range of flow (3.00-3.65 L/min).

Virtual cohort

Covariates are drawn to resemble Table 1: CRRT with probability 8/11, and ECMO blood flow from a normal distribution with mean 3.4 L/min and SD 0.45 L/min truncated to 2.5-4.5 L/min (Table 1 reports only the median and IQR). Two maintenance regimens bracket the reported dosing: 50 mg and 75 mg every 12 h (100 and 150 mg/day), each preceded by a 200 mg loading dose and given as 1-h infusions. Ten maintenance doses bring every subject to steady state, and the final interval is sampled densely.

set.seed(20241227)
rxSetSeed(20241227)
n_per_arm <- 200L
tau <- 12
n_maint <- 10L
arms <- tibble(treatment = c("50 mg q12h", "75 mg q12h"), maint = c(50, 75))

cov_df <- tidyr::expand_grid(arm_i = seq_len(nrow(arms)), j = seq_len(n_per_arm)) |>
  mutate(
    id = row_number(),
    treatment = arms$treatment[arm_i],
    maint = arms$maint[arm_i],
    RRT_CRRT_STATUS = rbinom(n(), 1, 8 / 11),
    Q_ECMO = pmin(pmax(rnorm(n(), 3.4, 0.45), 2.5), 4.5)
  )

dose_times <- c(0, tau * seq_len(n_maint))
t_last <- max(dose_times)
obs_times <- t_last + c(0, seq(0.25, tau, by = 0.25))

dose_rows <- cov_df |>
  tidyr::expand_grid(time = dose_times) |>
  mutate(
    amt = ifelse(time == 0, 200, maint),
    evid = 1L,
    dur = 1
  )
obs_rows <- cov_df |>
  tidyr::expand_grid(time = obs_times) |>
  mutate(amt = 0, evid = 0L, dur = 0)
events <- bind_rows(dose_rows, obs_rows) |>
  mutate(cmt = "central") |>
  arrange(id, time, desc(evid)) |>
  select(id, time, amt, evid, dur, cmt, treatment, Q_ECMO, RRT_CRRT_STATUS)

sim <- rxSolve(mod, events, returnType = "data.frame", addDosing = FALSE,
               keep = c("treatment")) |>
  mutate(tad = time - t_last)
#> ℹ parameter labels from comments will be replaced by 'label()'

Replicate Figure 3 (prediction-corrected VPC)

Figure 3 of Xu 2024 is a prediction-corrected VPC of the steady-state interval. The points below are the observed median of that figure, digitised by the maintainers at the paper’s sampling times; they are approximate and are shown for orientation only. The paper does not report the individual maintenance doses, so both simulated regimens are drawn.

fig3_obs <- tibble(
  tad = c(1.25, 2, 4, 7),
  Cc = c(3.1, 2.3, 1.75, 1.0)
)
vpc <- sim |>
  group_by(treatment, tad) |>
  summarise(
    p05 = quantile(Cc, 0.05),
    p50 = median(Cc),
    p95 = quantile(Cc, 0.95),
    .groups = "drop"
  )
ggplot(vpc, aes(tad, p50, colour = treatment, fill = treatment)) +
  geom_ribbon(aes(ymin = p05, ymax = p95), alpha = 0.15, colour = NA) +
  geom_line() +
  geom_point(data = fig3_obs, aes(tad, Cc), inherit.aes = FALSE) +
  labs(x = "Time after dose (h)", y = "Polymyxin B concentration (mg/L)",
       colour = NULL, fill = NULL)
Simulated steady-state concentrations (median and 5th-95th percentile band) for the two maintenance regimens. Black points: observed median read from Xu 2024 Figure 3. Replicates Figure 3 of Xu 2024.

Simulated steady-state concentrations (median and 5th-95th percentile band) for the two maintenance regimens. Black points: observed median read from Xu 2024 Figure 3. Replicates Figure 3 of Xu 2024.

The digitised observed median falls between the two simulated regimens, which is consistent with the 100-150 mg/day dosing the Methods describe.

PKNCA validation

Steady-state NCA over the final 12-h interval, by regimen.

sim_nca <- sim |>
  filter(!is.na(Cc)) |>
  mutate(time = tad) |>
  select(id, time, Cc, treatment)
dose_nca <- events |>
  filter(evid == 1, time == t_last) |>
  mutate(time = 0) |>
  select(id, time, amt, treatment)

conc_obj <- PKNCA::PKNCAconc(sim_nca, Cc ~ time | treatment + id,
                             concu = "mg/L", timeu = "h")
dose_obj <- PKNCA::PKNCAdose(dose_nca, amt ~ time | treatment + id,
                             doseu = "mg")
intervals <- data.frame(
  start = 0, end = tau,
  cmax = TRUE, tmax = TRUE, cmin = TRUE, auclast = TRUE, cav = TRUE
)
nca_res <- PKNCA::pk.nca(PKNCA::PKNCAdata(conc_obj, dose_obj, intervals = intervals))
summary(nca_res)
#>  Interval Start Interval End  treatment   N AUClast (h*mg/L) Cmax (mg/L)
#>               0           12 50 mg q12h 200      12.7 [24.0] 2.78 [26.4]
#>               0           12 75 mg q12h 200      19.6 [25.6] 4.23 [27.0]
#>   Cmin (mg/L)          Tmax (h)  Cav (mg/L)
#>  0.388 [56.7] 1.00 [1.00, 1.00] 1.06 [24.0]
#>  0.615 [59.9] 1.00 [1.00, 1.00] 1.63 [25.6]
#> 
#> Caption: AUClast, Cmax, Cmin, Cav: geometric mean and geometric coefficient of variation; Tmax: median and range; N: number of subjects

Mass balance at steady state

At steady state the AUC over one interval equals dose / CL for every subject. Both sides use the same drawn parameters, so the difference is only the trapezoidal error of the 0.25-h grid.

cl_by_id <- sim |> distinct(id, cl)
auc_chk <- as.data.frame(nca_res) |>
  filter(PPTESTCD == "auclast") |>
  select(id, treatment, auclast = PPORRES) |>
  left_join(cl_by_id, by = "id") |>
  left_join(dose_nca |> select(id, amt), by = "id") |>
  mutate(pct_diff = 100 * (auclast - amt / cl) / (amt / cl))
summary(auc_chk$pct_diff)
#>     Min.  1st Qu.   Median     Mean  3rd Qu.     Max. 
#> -0.22316 -0.04201 -0.02004 -0.03162 -0.01085  0.36296
stopifnot(max(abs(auc_chk$pct_diff)) < 1)

Comparison with the published Css,avg

Xu 2024 reports Css,avg (AUC0-tau / tau from WinNonlin) only by microbiological outcome: 2.26 +/- 0.72 mg/L in the clearance group and 1.25 +/- 0.24 mg/L in the uncleared group (Results and Figure 4). Because those groups are defined by outcome rather than by regimen, the side-by-side table below is context rather than a like-for-like NCA comparison.

cav_sim <- as.data.frame(nca_res) |>
  filter(PPTESTCD == "cav") |>
  group_by(treatment) |>
  summarise(
    median = median(PPORRES),
    p05 = quantile(PPORRES, 0.05),
    p95 = quantile(PPORRES, 0.95),
    .groups = "drop"
  )
bind_rows(
  cav_sim |>
    transmute(Source = paste("Simulated,", treatment),
              Value = sprintf("%.2f (%.2f-%.2f)", median, p05, p95)),
  tibble(Source = c("Xu 2024, clearance group", "Xu 2024, uncleared group"),
         Value = c("2.26 +/- 0.72", "1.25 +/- 0.24"))
) |>
  rename(`Css,avg (mg/L)` = Value) |>
  knitr::kable(caption = "Simulated median (5th-95th percentile) Css,avg against the published outcome-group mean +/- SD.")
Simulated median (5th-95th percentile) Css,avg against the published outcome-group mean +/- SD.
Source Css,avg (mg/L)
Simulated, 50 mg q12h 1.07 (0.74-1.47)
Simulated, 75 mg q12h 1.63 (1.13-2.44)
Xu 2024, clearance group 2.26 +/- 0.72
Xu 2024, uncleared group 1.25 +/- 0.24

cav75 <- cav_sim$median[cav_sim$treatment == "75 mg q12h"]
stopifnot(cav75 > 1.25, cav75 < 2.26)

The 150 mg/day regimen gives a median Css,avg between the two published group means, as expected for a cohort dosed at 100-150 mg/day whose individual exposures span both groups.

Assumptions and deviations

  • Age is not carried. The Results text says “CRRT and age were identified as the covariates for V”, and the Table 2 abbreviations define a dVdage coefficient, but neither the printed V equation nor the Table 2 parameter rows contain an age term or estimate. The printed final-model equation is implemented; age is recorded in covariatesDataExcluded.
  • ECMO flow, not pump speed. The CL covariate is printed as speadECMO with units of L/min, which identifies it as the delivered blood flow (Table 1 reports pump speed separately in RPM). It is mapped to Q_ECMO. The centring value 3.48 L/min is the one in the printed equation, not the Table 1 median of 3.60 L/min. The Discussion at one point links ECMO flow to the volume of distribution; the equation and Table 2 put the effect on CL, which is what is implemented.
  • Linear flow effect. With a slope of -0.321 per L/min the typical CL reaches zero at 6.60 L/min. The observed flows were 3.0-3.65 L/min (IQR); do not extrapolate far outside that range.
  • Residual error is additive. Table 2 reports a single stdev0 of 0.361. The Results select “the two-compartment model with an additive option”, and the lower prediction band of the Figure 3 VPC falls below zero, which only an additive error produces. It is encoded as an additive SD in mg/L.
  • IIV. Table 2 reports variances of 0.062 (V) and 0.038 (CL) for an exponential model; no covariance is reported. No IIV is reported on Vp or Q.
  • Dosing in the simulations. The Methods give “100-150 mg (2.5-3 mg/kg/day)” without stating the interval; the 12-h sampling schedule and the Figure 3 concentrations agree with 50-75 mg every 12 h, which is what is simulated. The loading dose is taken as 200 mg.
  • Covariate distribution. Table 1 gives only the median and IQR of ECMO flow; the truncated normal used for the virtual cohort is a maintainers’ choice.
  • Table 1 inconsistencies. Serum creatinine is labelled mg/dL but the values (median 137.27) are umol/L; the Results prose gives CrCL the same numbers as serum creatinine and an age median of 60 (53.50-62.50) against Table 1’s 56 (48, 64). The number of patients with bacterial clearance is 7 in the Results and 6 in Table 1. None of these enters the model.
  • Convention-check note. checkModelConventions() warns that e_q_ecmo_cl “looks like a reversed-order covariate effect” because its first token q is also a PK parameter name. The covariate is Q_ECMO, so the name is in the canonical e_<cov>_<param> order, matching e_q_ecmo_cl_1ohm in Kim_2026_midazolam_ecmo.