Skip to contents

Model and source

  • Citation: Alli A, Paruk F, Roger C, Lipman J, Calleemalay D, Wallis SC, Scribante J, Richards GA, Roberts JA, Ballotta A. Peri-operative pharmacokinetics of cefazolin prophylaxis during valve replacement surgery. PLoS One. 2023;18(9):e0291425. doi:10.1371/journal.pone.0291425
  • Description: Three-compartment joint total/unbound population PK model for intravenous cefazolin surgical prophylaxis in adults undergoing elective cardiac valve replacement with cardiopulmonary bypass (Alli 2023). The dose enters a central compartment whose concentration is the TOTAL plasma cefazolin; it exchanges first-order with a peripheral compartment and with a separate unbound-drug compartment (volume Vt) whose concentration is the UNBOUND plasma cefazolin. The return rate constant from the unbound compartment to the central (total) compartment is the protein-binding term PB = (ktc * Bmax + NS) * (ALB / 40), so higher serum albumin pulls drug back into the total pool. Clearance acts on the central compartment and scales inversely with serum creatinine, CL = TVCL * (90 / SeCr). Estimated with the Pmetrics non-parametric adaptive grid (NPAG); the non-parametric distribution is approximated here by log-normal IIV from the reported CV%.
  • Article (open access): https://doi.org/10.1371/journal.pone.0291425

Alli and colleagues measured total and unbound plasma cefazolin in adults receiving 2 g cefazolin prophylaxis for elective cardiac valve replacement on cardiopulmonary bypass (CPB), fitted a joint total/unbound population PK model in Pmetrics (NPAG), and used it to estimate the probability of target attainment (PTA) of 100% fT>MIC over the first 4 hours of surgery.

The structure (Figure 2 of the paper) has three compartments:

  1. a central (“dosing”) compartment of volume Vc whose concentration is the total plasma cefazolin; clearance acts here and is scaled by serum creatinine, CL = TVCL * (90 / SeCr);
  2. a peripheral compartment exchanging with central through kcp / kpc;
  3. an unbound-drug compartment of volume Vt whose concentration is the unbound plasma cefazolin. Drug moves into it from central at rate kct and returns at the albumin-scaled protein-binding rate constant PB = (ktc * Bmax + NS) * (ALB / 40).

Population

Sixteen adults (8 women, 8 men; age 18-59 years, mean 44; weight 51-97 kg, mean 75.1) undergoing CPB for elective valve replacement at Charlotte Maxeke Johannesburg Academic Hospital, South Africa (Table 1 of the paper). Mean serum creatinine was 76.7 umol/L (range 42-102), mean serum albumin 41.6 g/L (range 33-47) and mean CPB duration 165 min. Each patient received 2 g cefazolin IV 30 minutes before incision, re-dosed with 2 g at 4 hours if surgery was ongoing. The analysis used 195 total and 64 unbound concentrations.

The same information is available programmatically:

str(readModelDb("Alli_2023_cefazolin")()$population)
#> List of 13
#>  $ species       : chr "human"
#>  $ n_subjects    : int 16
#>  $ n_studies     : int 1
#>  $ n_samples     : chr "195 total and 64 unbound plasma cefazolin concentrations"
#>  $ age_range     : chr "18-59 years (mean 44.0, SD 12.0)"
#>  $ weight_range  : chr "51-97 kg (mean 75.1, SD 15.6)"
#>  $ sex_female_pct: num 50
#>  $ disease_state : chr "Adults (18-60 years eligible) undergoing cardiopulmonary bypass (CPB) for elective cardiac valve replacement, r"| __truncated__
#>  $ renal_function: chr "Serum creatinine mean 76.7 umol/L (SD 18.6, range 42-102); Cockcroft-Gault creatinine clearance mean 81.5 mL/mi"| __truncated__
#>  $ albumin       : chr "Serum albumin mean 41.6 g/L (SD 4.57, range 33-47)."
#>  $ dose_range    : chr "2 g cefazolin IV (Figure 2: 1-min bolus infusion) 30 min before skin incision, re-dosed 2 g after 4 h if surgery was ongoing."
#>  $ regions       : chr "South Africa (Charlotte Maxeke Johannesburg Academic Hospital)"
#>  $ notes         : chr "Demographics from Alli 2023 Table 1. Prospective convenience-sampled PK study; plasma sampled pre-dose and 2-30"| __truncated__

Source trace

Equation / parameter Value Source location
lcl (CL at SeCr 90 umol/L) log(3.23) L/h Table 2, Cl mean (CV 36.06%)
lvc (Vc, total) log(3.38) L Table 2, Vc mean (CV 21.59%)
lvp2 (Vt, unbound compartment) log(6.08) L Table 2, Vt mean (CV 33.42%)
lk12 (kcp) log(5.92) 1/h Table 2, kcp mean (CV 65.58%)
lk21 (kpc) log(2.24) 1/h Table 2, kpc mean (CV 51.81%)
lk13 (kct) log(144.27) 1/h Table 2, kct mean (CV 10.17%)
lk31 (ktc) log(195.91) 1/h Table 2, ktc mean (CV 12.94%)
lbmax (Bmax) log(1.01) g/L Table 2, Bmax mean (CV 45.23%)
lkns (NS) log(15.14) Table 2, NS mean (CV 24.79%)
IIV (eta*) omega^2 = log(CV^2 + 1) Table 2, CV% column
addSd, propSd (total) 1.0 mg/L, 0.1 Results, assay error C0 and C1 for plasma data
addSd_Cu, propSd_Cu (unbound) 0.1 mg/L, 0.1 Results, assay error C0 and C1 for unbound data
cl = TVCL * (90 / CREAT) n/a Results, ‘The final clearance model’
kpb = (k31 * bmax + kns) * (ALB / 40) n/a Results, ‘The protein binding was described as follows’
Three-compartment ODE structure n/a Results text and Figure 2
Cc = central / vc, Cu = peripheral2 / vp2 n/a Figure 2 volumes; Results (Vc central volume, Vt volume of the unbound compartment)

Virtual cohort

Observed data are not public. The virtual cohort draws serum creatinine and albumin from normal distributions with the Table 1 means and SDs, truncated to the observed ranges. Subjects receive 2 g over 1 minute (Figure 2: ‘1 min bolus intravenous infusion’) at 0 and 4 hours and are sampled at the paper’s schedule (2-300 minutes).

set.seed(20230920)
rxode2::rxSetSeed(20230920)

n_sub <- 200
rtnorm <- function(n, mean, sd, lo, hi) {
  pmin(pmax(rnorm(n, mean, sd), lo), hi)
}
cohort <- tibble(
  id = seq_len(n_sub),
  CREAT = rtnorm(n_sub, 76.7, 18.6, 42, 102),
  ALB = rtnorm(n_sub, 41.6, 4.57, 33, 47)
)

# Paper sampling schedule (minutes after the first dose), converted to hours.
obs_min <- c(0, 2, 5, 10, 30, 60, 90, 120, 150, 180, 210, 240, 270, 300)

# Builds a dosing + observation event table. Doses are 1-minute infusions into
# `central`; observation rows carry dvid = 1 because the model declares two
# endpoints (Cc, Cu) -- both come back as columns on every observation row.
make_events <- function(cov, dose_times, dose_amts, obs_times) {
  doses <- tidyr::expand_grid(id = cov$id, k = seq_along(dose_times)) |>
    mutate(time = dose_times[k], amt = dose_amts[k], rate = amt * 60,
           evid = 1L, cmt = "central", dvid = NA_integer_) |>
    select(-k)
  obs <- tidyr::expand_grid(id = cov$id, time = obs_times) |>
    mutate(amt = 0, rate = 0, evid = 0L, cmt = "central", dvid = 1L)
  bind_rows(doses, obs) |>
    left_join(cov, by = "id") |>
    arrange(id, time, desc(evid))
}

events <- make_events(cohort, c(0, 4), c(2000, 2000), obs_min / 60)
stopifnot(!anyDuplicated(unique(events[, c("id", "time", "evid")])))

Simulation

mod <- readModelDb("Alli_2023_cefazolin")
sim <- rxode2::rxSolve(mod, events = events, returnType = "data.frame") |>
  filter(!is.na(Cc))
#> ℹ parameter labels from comments will be replaced by 'label()'

# Typical patient at the cohort-mean covariates.
typ_events <- make_events(tibble(id = 1L, CREAT = 76.7, ALB = 41.6),
                          c(0, 4), c(2000, 2000), obs_min / 60)
sim_typ <- rxode2::rxSolve(rxode2::zeroRe(mod), events = typ_events,
                           returnType = "data.frame")
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc', 'etalvp2', 'etalk12', 'etalk21', 'etalk13', 'etalk31', 'etalbmax', 'etalkns'

Replicate published figures

Figure 1 – mean total and unbound concentrations

Figure 1 of the paper plots the observed mean (SD) total and unbound concentrations at each sampling time. The means below were read from that figure by the maintainers (approximately +/- 5 mg/L).

fig1 <- tibble(
  minute = obs_min[-1],
  total_obs = c(317, 260, 210, 159, 112, 80, 72, 66, 61, 56, 50, 169, 138),
  unbound_obs = c(145, 119, 97, 73, 52, 38, 35, 32, 29, 27, 24, 78, 64)
)

typ_cmp <- sim_typ |>
  mutate(minute = round(time * 60)) |>
  filter(minute > 0) |>
  select(minute, total_typ = Cc, unbound_typ = Cu) |>
  left_join(fig1, by = "minute") |>
  mutate(
    total_pct = 100 * (total_typ - total_obs) / total_obs,
    unbound_pct = 100 * (unbound_typ - unbound_obs) / unbound_obs
  )

typ_cmp |>
  mutate(across(where(is.numeric), \(x) round(x, 1))) |>
  rename(
    "Minute" = minute,
    "Total, typical (mg/L)" = total_typ,
    "Total, Fig 1 mean (mg/L)" = total_obs,
    "Total diff (%)" = total_pct,
    "Unbound, typical (mg/L)" = unbound_typ,
    "Unbound, Fig 1 mean (mg/L)" = unbound_obs,
    "Unbound diff (%)" = unbound_pct
  ) |>
  knitr::kable(caption = "Typical-patient prediction (SeCr 76.7 umol/L, ALB 41.6 g/L) against the Figure 1 observed means.")
Typical-patient prediction (SeCr 76.7 umol/L, ALB 41.6 g/L) against the Figure 1 observed means.
Minute Total, typical (mg/L) Unbound, typical (mg/L) Total, Fig 1 mean (mg/L) Unbound, Fig 1 mean (mg/L) Total diff (%) Unbound diff (%)
2 318.6 117.4 317 145 0.5 -19.0
5 263.6 97.0 260 119 1.4 -18.5
10 201.9 74.1 210 97 -3.9 -23.7
30 116.5 42.3 159 73 -26.7 -42.0
60 93.9 34.0 112 52 -16.1 -34.5
90 82.8 30.0 80 38 3.5 -21.1
120 73.3 26.6 72 35 1.8 -24.1
150 64.9 23.5 66 32 -1.7 -26.5
180 57.5 20.8 61 29 -5.8 -28.2
210 50.9 18.4 56 27 -9.1 -31.7
240 45.1 16.3 50 24 -9.8 -31.9
270 156.4 56.8 169 78 -7.4 -27.2
300 129.3 46.9 138 64 -6.3 -26.8
# Deterministic (typical-value) comparison, so a tight bound is safe. The total
# concentration is the primary endpoint (195 samples) and the model reproduces
# its mean profile: a mis-transcribed CL, Vc or binding constant moves the whole
# profile by tens of percent.
stopifnot(abs(median(typ_cmp$total_pct)) < 10)
# Unbound fraction implied by the binding structure at the typical covariates.
# The model's Cu / Cc settles near 0.36; the observed means in Figure 1 sit
# nearer 0.45-0.5 (see Assumptions and deviations).
fu_typ <- with(sim_typ[sim_typ$time > 1, ], median(Cu / Cc))
stopifnot(fu_typ > 0.3, fu_typ < 0.45)
fu_typ
#> [1] 0.3624384
# Replicates Figure 1 of Alli 2023 (observed means) with a VPC overlay.
vpc <- sim |>
  select(id, time, Total = Cc, Unbound = Cu) |>
  pivot_longer(c(Total, Unbound), names_to = "analyte", values_to = "conc") |>
  group_by(time, analyte) |>
  summarise(
    Q05 = quantile(conc, 0.05), Q50 = quantile(conc, 0.50),
    Q95 = quantile(conc, 0.95), mean = mean(conc), .groups = "drop"
  )
obs_long <- fig1 |>
  pivot_longer(c(total_obs, unbound_obs), names_to = "analyte", values_to = "conc") |>
  mutate(analyte = ifelse(analyte == "total_obs", "Total", "Unbound"),
         time = minute / 60)

ggplot(vpc, aes(time * 60, Q50, colour = analyte, fill = analyte)) +
  geom_ribbon(aes(ymin = Q05, ymax = Q95), alpha = 0.2, colour = NA) +
  geom_line() +
  geom_point(data = obs_long, aes(y = conc), shape = 21, size = 2) +
  labs(x = "Time after first dose (min)", y = "Cefazolin (mg/L)",
       colour = NULL, fill = NULL,
       caption = paste("Lines/ribbons: simulated median and 90% interval (n = 200).",
                       "Points: observed means read from Figure 1 of Alli 2023."))

Figure 3 – probability of target attainment

The paper’s PTA simulations require the unbound concentration to exceed the MIC for 100% of the first 4 hours of surgery, and (Methods) compute the unbound concentration as a fixed 26% of the total – the ‘worst case’ unbound fraction the authors measured at 120 minutes – rather than from the model’s unbound compartment. The replication below does the same: the target is 0.26 * Cc >= MIC at every time point from the end of the first infusion to 4 hours. Doses inside the 4-hour window are those of each regimen (a 4-hourly regimen contributes only its first dose).

pta_grid <- sort(unique(c(seq(1 / 60, 4, by = 2 / 60), 2 - 1e-3, 4)))
mic_grid <- 2^(-4:6)

pta_scenarios <- tribble(
  ~panel, ~label, ~creat, ~alb, ~t2, ~d1, ~d2,
  "a", "1g4h", 50, 35, NA, 1000, NA,
  "a", "2g4h", 50, 35, NA, 2000, NA,
  "a", "3g4h", 50, 35, NA, 3000, NA,
  "a", "1g2h", 50, 35, 2, 1000, 1000,
  "a", "2g2h", 50, 35, 2, 2000, 2000,
  "a", "2g/1g 2h", 50, 35, 2, 2000, 1000,
  "a", "3g/1.5g 2h", 50, 35, 2, 3000, 1500,
  "b", "Cr50", 50, 35, NA, 2000, NA,
  "b", "Cr75", 75, 35, NA, 2000, NA,
  "b", "Cr100", 100, 35, NA, 2000, NA,
  "c", "Alb30", 75, 30, NA, 3000, NA,
  "c", "Alb35", 75, 35, NA, 3000, NA,
  "c", "Alb40", 75, 40, NA, 3000, NA
) |>
  mutate(scenario = row_number())

pta_events <- bind_rows(lapply(seq_len(nrow(pta_scenarios)), function(i) {
  s <- pta_scenarios[i, ]
  times <- if (is.na(s$t2)) 0 else c(0, s$t2)
  amts <- if (is.na(s$t2)) s$d1 else c(s$d1, s$d2)
  cov <- tibble(id = (i - 1L) * n_sub + seq_len(n_sub),
                CREAT = s$creat, ALB = s$alb, scenario = i)
  make_events(cov, times, amts, pta_grid)
}))
stopifnot(!anyDuplicated(unique(pta_events[, c("id", "time", "evid")])))

pta_sim <- rxode2::rxSolve(mod, events = pta_events, keep = "scenario",
                           returnType = "data.frame")

trough_free <- pta_sim |>
  filter(time >= 1 / 60) |>
  group_by(scenario, id) |>
  summarise(min_free = min(0.26 * Cc), .groups = "drop")

pta <- tidyr::expand_grid(scenario = pta_scenarios$scenario, mic = mic_grid) |>
  left_join(trough_free, by = "scenario", relationship = "many-to-many") |>
  group_by(scenario, mic) |>
  summarise(pta = mean(min_free >= mic), .groups = "drop") |>
  left_join(pta_scenarios |> select(scenario, panel, label), by = "scenario")
# Replicates Figure 3a-c of Alli 2023.
panel_titles <- c(a = "(a) regimens, SeCr 50, ALB 35",
                  b = "(b) 2 g 4-hourly, ALB 35",
                  c = "(c) 3 g 4-hourly, SeCr 75")
ggplot(pta, aes(mic, pta, colour = label)) +
  geom_line() +
  geom_point(size = 1) +
  scale_x_log10(breaks = mic_grid, labels = format(mic_grid, drop0trailing = TRUE)) +
  facet_wrap(~panel, ncol = 2, labeller = labeller(panel = panel_titles)) +
  labs(x = "MIC (mg/L)", y = "PTA (100% fT>MIC, 0-4 h)", colour = NULL,
       caption = "Simulated with 0.26 x total concentration as unbound; n = 200 per curve.")

The legible curves of Figure 3 were read by the maintainers at the plotted MIC values (approximately +/- 0.03). The comparison uses the MIC at which PTA crosses 50% (interpolated on the log2 MIC scale), which is insensitive to which subjects land in the tails.

pub_pta <- tribble(
  ~label, ~mic, ~pta_pub,
  "1g4h", 1, 0.95, "1g4h", 2, 0.80, "1g4h", 4, 0.38, "1g4h", 8, 0.04,
  "2g4h", 2, 0.96, "2g4h", 4, 0.80, "2g4h", 8, 0.39, "2g4h", 16, 0.03,
  "3g4h", 4, 0.92, "3g4h", 8, 0.64, "3g4h", 16, 0.20, "3g4h", 32, 0,
  "Cr75", 4, 0.97, "Cr75", 8, 0.79, "Cr75", 16, 0.16, "Cr75", 32, 0,
  "Cr100", 8, 0.95, "Cr100", 16, 0.37, "Cr100", 32, 0, "Cr100", 64, 0,
  "Alb30", 4, 0.92, "Alb30", 8, 0.65, "Alb30", 16, 0.20, "Alb30", 32, 0,
  "Alb35", 4, 0.99, "Alb35", 8, 0.95, "Alb35", 16, 0.55, "Alb35", 32, 0.03,
  "Alb40", 4, 0.99, "Alb40", 8, 0.95, "Alb40", 16, 0.55, "Alb40", 32, 0.03
)

# log2 MIC at which a decreasing PTA curve crosses 0.5.
mic50 <- function(mic, p) {
  o <- order(mic)
  mic <- mic[o]
  p <- p[o]
  i <- which(p >= 0.5 & c(p[-1], 0) < 0.5)[1]
  if (i == length(p)) return(log2(mic[i]))
  log2(mic[i]) + (p[i] - 0.5) / (p[i] - p[i + 1]) * (log2(mic[i + 1]) - log2(mic[i]))
}

sim_mic50 <- pta |>
  filter(label %in% unique(pub_pta$label)) |>
  group_by(label) |>
  summarise(sim = mic50(mic, pta), .groups = "drop")
pub_mic50 <- pub_pta |>
  group_by(label) |>
  summarise(pub = mic50(mic, pta_pub), .groups = "drop")
mic50_cmp <- left_join(pub_mic50, sim_mic50, by = "label") |>
  mutate(diff_dilutions = sim - pub,
         label = factor(label, levels = unique(pub_pta$label))) |>
  arrange(label)

mic50_cmp |>
  mutate(across(where(is.numeric), \(x) round(2^x, 2), .names = "{.col}_mgL"),
         diff_dilutions = round(diff_dilutions, 2)) |>
  select(label, pub_mgL, sim_mgL, diff_dilutions) |>
  rename(
    "Curve" = label,
    "MIC50, Figure 3 (mg/L)" = pub_mgL,
    "MIC50, simulated (mg/L)" = sim_mgL,
    "Difference (doubling dilutions)" = diff_dilutions
  ) |>
  knitr::kable(caption = "MIC at which PTA crosses 50%: Figure 3 versus simulation.")
MIC at which PTA crosses 50%: Figure 3 versus simulation.
Curve MIC50, Figure 3 (mg/L) MIC50, simulated (mg/L) Difference (doubling dilutions)
1g4h 3.28 3.06 -0.10
2g4h 6.64 6.01 -0.14
3g4h 9.97 9.80 -0.03
Cr75 11.01 10.00 -0.14
Cr100 13.70 12.30 -0.15
Alb30 10.08 16.00 0.67
Alb35 17.10 14.19 -0.27
Alb40 17.10 14.88 -0.20
# Regimen, renal-function and albumin 35 / 40 g/L curves must land within one
# doubling dilution of the published curve. The albumin 30 g/L curve of
# Figure 3c is a known deviation and is not gated -- see Assumptions and
# deviations.
gated <- mic50_cmp |> filter(label != "Alb30")
stopifnot(
  abs(median(gated$diff_dilutions)) < 0.5,
  all(abs(gated$diff_dilutions) < 1)
)

PKNCA validation

The paper reports no NCA table. NCA over the first dosing interval (0-4 h) is computed per analyte on the paper’s sampling schedule, so the simulated Cmax is the 2-minute concentration – the same quantity as the first plotted point of Figure 1, which serves as the reference.

nca_conc <- sim |>
  filter(time <= 4) |>
  select(id, time, Total = Cc, Unbound = Cu) |>
  pivot_longer(c(Total, Unbound), names_to = "analyte", values_to = "Cc") |>
  filter(!is.na(Cc))
# Pre-dose rows are present at time 0 (Cc = 0); guard anyway.
nca_conc <- bind_rows(
  nca_conc,
  nca_conc |> distinct(id, analyte) |> mutate(time = 0, Cc = 0)
) |>
  distinct(id, analyte, time, .keep_all = TRUE) |>
  arrange(analyte, id, time)

nca_dose <- events |>
  filter(evid == 1, time == 0) |>
  select(id, time, amt) |>
  tidyr::expand_grid(analyte = c("Total", "Unbound"))

conc_obj <- PKNCA::PKNCAconc(nca_conc, Cc ~ time | analyte + id)
dose_obj <- PKNCA::PKNCAdose(nca_dose, amt ~ time | analyte + id)
intervals <- data.frame(start = 0, end = 4, cmax = TRUE, tmax = TRUE,
                        auclast = TRUE)
nca_res <- PKNCA::pk.nca(PKNCA::PKNCAdata(conc_obj, dose_obj, intervals = intervals))
summary(nca_res)
#>  start end analyte   N    auclast       cmax                    tmax
#>      0   4   Total 200 301 [38.9] 304 [29.0] 0.0333 [0.0333, 0.0333]
#>      0   4 Unbound 200 108 [74.4] 111 [50.1] 0.0333 [0.0333, 0.0333]
#> 
#> Caption: auclast, cmax: geometric mean and geometric coefficient of variation; tmax: median and range; N: number of subjects

Comparison against published values

published <- tibble::tribble(
  ~analyte, ~cmax,
  "Total", 317,
  "Unbound", 145
)
cmp <- nlmixr2lib::ncaComparisonTable(
  simulated = nca_res,
  reference = published,
  by = "analyte",
  params = "cmax",
  units = c(cmax = "mg/L"),
  tolerance_pct = 20
)
knitr::kable(cmp, caption = "Simulated median Cmax (2 min) against the Figure 1 observed mean at 2 min. * differs by >20%.")
Simulated median Cmax (2 min) against the Figure 1 observed mean at 2 min. * differs by >20%.
NCA parameter analyte Reference Simulated % diff
Cmax (mg/L) Total 317 310 -2.2%
Cmax (mg/L) Unbound 145 111 -23.5%*

The total Cmax agrees closely. The unbound Cmax is lower than observed, for the reason given below.

Assumptions and deviations

  • Typical values are the Table 2 means. Pmetrics NPAG returns a discrete non-parametric population distribution, summarised in Table 2 by mean, SD, CV%, variance and median. The mean is used as the typical value (it is the figure the Results text quotes), and each CV% becomes a log-normal IIV variance, omega^2 = log(CV^2 + 1), with no covariances. This is a parametric approximation of the non-parametric distribution.
  • Residual-error multiplier. Pmetrics multiplies the assay-error polynomial SD = C0 + C1 * obs by an estimated noise factor gamma; the paper prints C0 and C1 but not gamma. The polynomial is used unscaled (gamma = 1) as a combined1 additive-plus-proportional error.
  • Reading of the binding term. The Results define PB = (ktc * Bmax + NS) * (ALB / 40) as ‘the rate constant for protein binding from unbound to total drug’, and Figure 2 draws Bmax and NS acting on the unbound-to-central arrow. PB is therefore the rate constant of that arrow, replacing ktc there; kct is the central-to-unbound rate constant. Bmax is tabulated in g/L and NS without a unit; both are used as dimensionless multipliers exactly as the equation is printed. This reading reproduces the Figure 1 total profile (median difference -5.8%). The alternative (PB as the central-to-unbound constant) was checked by the maintainers and fits the total peak markedly worse (about 250 versus 317 mg/L at 2 minutes).
  • Unbound concentrations are under-predicted. At the cohort-mean covariates the model’s unbound-to-total ratio settles at about 0.36, whereas the Figure 1 means imply about 0.45-0.5, so the typical unbound profile runs roughly 20-40% below the observed means. The paper reports no individual or population predictions for unbound concentrations beyond the S1 goodness-of-fit scatter, so the gap cannot be traced further; it is not tuned away. Note that the paper itself did not use the model’s unbound output for its PTA work but a fixed 26% unbound fraction.
  • Covariate references. The clearance equation normalises SeCr to 90 umol/L and the binding term normalises albumin to 40 g/L, both described as ‘the population mean value’, whereas Table 1 reports means of 76.7 umol/L and 41.6 g/L. The printed equation values are used.
  • The 30 g/L albumin curve of Figure 3c is not reproduced. In the model, albumin barely moves the total-concentration trough, so with the paper’s fixed 26% unbound fraction the three albumin curves nearly coincide. The 35 and 40 g/L curves match Figure 3c, but the published 30 g/L curve sits about two-thirds of a doubling dilution to the left. That curve is numerically identical to the 3 g 4-hourly curve of Figure 3a (SeCr 50 umol/L), which the simulation does reproduce, so the panel may carry a labelling issue. It is shown but not gated.
  • Figure 3d (surgery duration) is not replicated. Its 5- and 6-hour curves fall below the 4-hour curve under a 4-hourly regimen, which implies the re-dose was not given at 4 hours in those simulations; the timing used is not stated.
  • Window of the target. PTA is evaluated from the end of the first 1-minute infusion to 4 hours after the first dose; the paper says ‘the first 4 hours of surgery’ without stating whether the 30-minute pre-incision interval is included.
  • Virtual cohort. Creatinine and albumin are drawn from truncated normals with the Table 1 means, SDs and ranges; the paper’s covariate distribution is otherwise unreported. Fixed covariates (SeCr, ALB) are used for the PTA scenarios, as in the paper.
  • Digitised values. The Figure 1 means and Figure 3 PTA values were read by the maintainers from the published figures.
  • No correction notice for this article was found on Europe PMC (checked 2026-10-27).