Skip to contents

Levofloxacin in pregnancy (Sawe 2025)

Source: vignettes/articles/Sawe_2025_levofloxacin.Rmd
Sawe_2025_levofloxacin.Rmd

Model and source 

mod <- rxode2::rxode(readModelDb("Sawe_2025_levofloxacin"))
#> ℹ parameter labels from comments will be replaced by 'label()'
#> Warning: some etas defaulted to non-mu referenced, possible parsing error: etaiov_mtt_1, etaiov_mtt_2, etaiov_mtt_3, etaiov_mtt_4, etaiov_ka_1, etaiov_ka_2, etaiov_ka_3, etaiov_ka_4, etaiov_fdepot_1, etaiov_fdepot_2, etaiov_fdepot_3, etaiov_fdepot_4
#> as a work-around try putting the mu-referenced expression on a simple line
  • Citation: Sawe S, Tsirizani L, Court R, Gausi K, Poswa A, Badat T, Wiesner L, Loveday M, Maartens G, Conradie F, Denti P. The effect of pregnancy on the population pharmacokinetics of levofloxacin in South Africans with rifampicin-resistant tuberculosis. Antimicrob Agents Chemother. 2025 May;69(5):e01626-24. doi:10.1128/aac.01626-24. Values taken from the corrected version posted 17 April 2025, which revised the Table 2 footnote defining the reported %CV.
  • Description: One-compartment population PK model for oral levofloxacin in South African adults treated for rifampicin-resistant tuberculosis (RR-TB), characterising the effect of third-trimester pregnancy (Sawe 2025; n = 47 pooled from two studies, 21 pregnant, 12 with matched antepartum / postpartum profiles). Savic transit-compartment absorption (analytical form, N fixed to 20, MTT = 1.07 h) feeds first-order absorption into a one-compartment disposition model. Clearance and volume are scaled allometrically by fat-free mass (Janmahasatian formula) with fixed exponents 0.75 and 1 and the cohort-median 39.4 kg as reference. Higher serum creatinine lowers clearance via a power function (exponent -0.367) centred on the cohort median 56.2 umol/L, and third-trimester pregnancy raises clearance by a further 38.1%. Between-subject variability is retained on CL only; between-occasion variability is carried on MTT, ka and F, with the BOV on F inflated 2.35-fold on the two occasions whose preceding dose was self-reported rather than observed. Bioavailability is fixed at 1.
  • Article: https://doi.org/10.1128/aac.01626-24
  • Supplement (Tables S1-S3, Fig. S1-S7, NONMEM control stream): https://doi.org/10.1128/aac.01626-24 (file aac.01626-24-s0001.docx, retrieved via Europe PMC PMC12057369)

Sawe 2025 pooled two South African studies in adults treated for rifampicin-resistant tuberculosis (RR-TB) to ask what third-trimester pregnancy does to levofloxacin exposure. The answer is a 38.1% increase in clearance on top of the increases already explained by body size and renal function, which together drive antepartum exposures well below those of non-pregnant participants on the same weight-based dose.

Population

47 participants contributed 320 levofloxacin concentrations: 33 (70%) female, 38 (81%) black, median (range) age 32 (19-51) years, weight 58.0 (37.0-98.0) kg, height 1.60 (1.46-1.88) m, fat-free mass 39.4 (27.3-51.2) kg, serum creatinine 56.2 (25.3-110) umol/L and serum albumin 30.5 (17.0-40.0) g/L (Table 1). 31 (66%) were living with HIV and on antiretroviral therapy. Of the 33 women, 21 were pregnant and sampled in the third trimester; 12 of those contributed matched antepartum and postpartum profiles, giving 19 antepartum, 14 postpartum, 12 non-pregnant-female and 14 male pharmacokinetic profiles.

Levofloxacin was dosed 750 or 1000 mg once daily by WHO weight band (34-50 kg and above 50 kg), with a standard breakfast about 1 h before the observed dose. Sampling was pre-dose and at 2, 4, 6, 8, 10 and 24 h (BEAT Tuberculosis, NCT04062201), pre-dose and 2, 4, 6, 8 and 24 h in its pregnancy sub-study, and pre-dose and 2, 4 and 6 h in the King Dinuzulu Hospital observational cohort. Assay was HPLC-MS/MS with an LLOQ of 0.0781 mg/L; 6 of 320 samples (1.88%) were below it, all pre-dose.

The same information is available programmatically:

str(readModelDb("Sawe_2025_levofloxacin")()$population, max.level = 1)
#> Warning: some etas defaulted to non-mu referenced, possible parsing error: etaiov_mtt_1, etaiov_mtt_2, etaiov_mtt_3, etaiov_mtt_4, etaiov_ka_1, etaiov_ka_2, etaiov_ka_3, etaiov_ka_4, etaiov_fdepot_1, etaiov_fdepot_2, etaiov_fdepot_3, etaiov_fdepot_4
#> as a work-around try putting the mu-referenced expression on a simple line
#> List of 20
#>  $ species       : chr "human"
#>  $ n_subjects    : int 47
#>  $ n_studies     : int 2
#>  $ n_observations: int 320
#>  $ n_profiles    : chr "19 antepartum, 14 postpartum, 12 non-pregnant female, 14 male"
#>  $ age_range     : chr "19-51 years (Table 1: median 32 years)"
#>  $ weight_range  : chr "37.0-98.0 kg (Table 1: median 58.0 kg)"
#>  $ height_range  : chr "1.46-1.88 m (Table 1: median 1.60 m)"
#>  $ ffm_range     : chr "27.3-51.2 kg (Table 1: median 39.4 kg)"
#>  $ creat_range   : chr "25.3-110 umol/L (Table 1: median 56.2 umol/L)"
#>  $ alb_range     : chr "17.0-40.0 g/L (Table 1: median 30.5 g/L)"
#>  $ sex_female_pct: num 70.2
#>  $ race_ethnicity: chr "38 (81%) black, 9 (19%) white (Table 1)"
#>  $ n_hiv_positive: int 31
#>  $ n_pregnant    : int 21
#>  $ disease_state : chr "Rifampicin-resistant tuberculosis (RR-TB) on treatment. 21 of the 33 female participants were pregnant, sampled"| __truncated__
#>  $ dose_range    : chr "Levofloxacin 750 or 1000 mg orally once daily by body-weight band per WHO guidance (34-50 kg and above 50 kg). "| __truncated__
#>  $ regions       : chr "South Africa (two sites for BEAT Tuberculosis; King Dinuzulu Hospital, Durban)"
#>  $ co_medication : chr "Antiretroviral therapy in 31 participants, commonly tenofovir/lamivudine/dolutegravir; RR-TB co-treatment most "| __truncated__
#>  $ notes         : chr "Pooled from BEAT Tuberculosis (ClinicalTrials.gov NCT04062201, a phase 3 RCT with a pregnancy PK sub-study; sam"| __truncated__

Source trace

Every ini() entry carries an in-file comment naming its source location in inst/modeldb/specificDrugs/Sawe_2025_levofloxacin.R. Collected here for review. “Control stream” refers to the NONMEM code printed in the supplementary material, whose $THETA block agrees with Table 2 to every printed digit and supplies additional significant figures.

Equation / parameter Value Source location
lcl (CL/F) log(6.05859) L/h Table 2 “Clearance, CL (L/h)” 6.06 (95% CI 5.47-6.53); control stream $THETA 1
lvc (V/F) log(85.8803) L Table 2 “Volume of distribution, V (L)” 85.9 (95% CI 80.6-91.7); control stream $THETA 2
lka log(1.59498) 1/h Table 2 “Absorption rate constant, Ka (1/h)” 1.59 (95% CI 1.11-2.40); control stream $THETA 3
lmtt log(1.07437) h Table 2 “Mean transit time, MTT (h)” 1.07 (95% CI 0.771-1.32); control stream $THETA 7
lnn (N transit) fixed(log(20)) Table 2 “Number of transit compartments, NN” 20 (fixed); control stream $THETA 13 = 20 FIX
lfdepot (F) fixed(log(1)) Table 2 “Bioavailability, F” 1 (fixed); control stream $THETA 4 = 1 FIX
e_ffm_cl fixed(0.75) Methods, “allometric scaling of clearance (with a fixed exponent of 0.75)”; control stream ALLMCL_FFM = (FFM/39.4)**0.75
e_ffm_vc fixed(1) Methods, “volume of distribution (with a fixed exponent of 1)”; control stream ALLMV_FFM = (FFM/39.4)
e_preg_cl 0.380866 Table 2 “Effect of pregnancy on CL (%)” +38.1 (95% CI +23.4 to +57.1); control stream IF (PREGNANT.EQ.1) preg_CL = 1 + THETA(12), $THETA 12
e_creat_cl -0.366504 Table 2 “Effect of serum creatinine on CL (power exponent)” -0.367 (95% CI -0.493 to -0.104); Methods equation (supplementary inline equation m001); control stream $THETA 15
etalcl 0.0489868 Table 2 “Between-subject variability, CL” 22.1% (95% CI 17.1-28.3); control stream $OMEGA 1
etaiov_mtt_1..4 0.211291 Table 2 “Between-occasion variability, MTT” 45.9% (95% CI 30.5-70.3); control stream $OMEGA 21 + 3x SAME
etaiov_ka_1..4 0.732947 Table 2 “Between-occasion variability, Ka” 85.6% (95% CI 61.7-119); control stream $OMEGA 17 + 3x SAME
etaiov_fdepot_2, _4 0.0562092 Table 2 “Between-occasion variability, F” 23.7% (95% CI 19.3-28.1); control stream $OMEGA 13 + 3x SAME
etaiov_fdepot_1, _3 0.310775 2.35136^2 * 0.0562092, from Table 2 “Scaling factor for BOV on F for unobserved doses” 2.35 (95% CI 1.68-3.28) and control stream $THETA 14 with IF(UNOBS.EQ.1) BOVBIO = E_BOVF*BOVBIO
addSd 0.259443 mg/L Table 2 “Additive error (mg/L)” 0.244 ($THETA 6 = 0.243823) plus the control stream’s unconditional ADD = THETA(6) + LLOQ*0.2 with LLOQ 0.0781 mg/L
propSd 0.0733236 Table 2 “Proportional error (%)” 7.33 (95% CI 6.49-8.10); control stream $THETA 5
Transit input rate transit(nn, mtt, fbio) Control stream PIZZA/TRANSIT block with KTR = (NN+1)/MTT, citing Savic 2007; supplementary Fig. S1
Dose enters via transit only no f(depot) line Control stream F1 = 0; under rxUi the bolus is not separately added, so no explicit suppression is needed (see Errata)
CL covariate model (FFM/39.4)^0.75 * (1 + e_preg_cl*PREG) * (CREAT/56.2)^e_creat_cl Control stream TVCL = THETA(1)*ALLMCL_FFM*preg_CL*CR_CL
Residual error add(addSd) + prop(propSd) Control stream W = SQRT(ADD**2 + PROP**2)
Reference FFM / creatinine 39.4 kg / 56.2 umol/L Table 2 footnote b; control stream TVFFM = 39.4, MEDCREAT = 56.2

The variance scale deserves a note of its own, because it is what the publisher’s 17 April 2025 correction changed. Table 2 footnote d defines the reported percentages as

%CV=ω2⋅100\%CV = \sqrt{\omega^2}\cdot 100

so a reported “22.1%” is omega * 100 and is not a log-normal CV. The usual omega^2 = log(CV^2 + 1) conversion must therefore NOT be applied here. The check below confirms the reading against the control stream’s $OMEGA values, all four of which reproduce their printed Table 2 percentage.

omega_cs <- c(
  "BSV CL"  = 0.0489868,
  "BOV MTT" = 0.211291,
  "BOV Ka"  = 0.732947,
  "BOV F"   = 0.0562092
)
printed <- c("BSV CL" = 22.1, "BOV MTT" = 45.9, "BOV Ka" = 85.6, "BOV F" = 23.7)

data.frame(
  Parameter          = names(omega_cs),
  `omega^2 (stream)` = omega_cs,
  `sqrt(omega^2)*100` = round(sqrt(omega_cs) * 100, 1),
  `Table 2 %CV`      = printed,
  `log(CV^2+1) would give` = round(log(printed^2 / 1e4 + 1), 4),
  check.names = FALSE, row.names = NULL
) |>
  knitr::kable(caption = "Table 2 footnote d reads %CV = sqrt(omega^2)*100.")
Table 2 footnote d reads %CV = sqrt(omega^2)*100.
Parameter omega^2 (stream) sqrt(omega^2)*100 Table 2 %CV log(CV^2+1) would give
BSV CL 0.0489868 22.1 22.1 0.0477
BOV MTT 0.2112910 46.0 45.9 0.1912
BOV Ka 0.7329470 85.6 85.6 0.5497
BOV F 0.0562092 23.7 23.7 0.0546 

# sqrt(omega^2)*100 reproduces every printed Table 2 percentage to the
# printed precision (MTT prints 45.9%, a truncation of 45.97%).
stopifnot(all(abs(sqrt(omega_cs) * 100 - printed) < 0.1))

Covariate derivation: fat-free mass

The size descriptor is fat-free mass, not total body weight. The supplementary material gives the Janmahasatian formula

FFM=WHSmax⋅HT2⋅WTWHS50⋅HT2+WTFFM = \frac{WHS_{max}\cdot HT^2 \cdot WT}{WHS_{50}\cdot HT^2 + WT}

with WHSmax 37.99 and WHS50 35.98 for females, 42.92 and 30.93 for males, HT in metres and WT in kilograms. FFM beat total body weight as the allometric descriptor (dOFV 12.9 versus 3.46), which is why no explicit sex term appears in the final model: the sex difference in exposure is mediated through FFM.

ffm_janmahasatian <- function(wt, ht, sexf) {
  whs_max <- ifelse(sexf == 1, 37.99, 42.92)
  whs_50  <- ifelse(sexf == 1, 35.98, 30.93)
  whs_max * ht^2 * wt / (whs_50 * ht^2 + wt)
}

Applying it to the published median weight and height of each participant group should reproduce that group’s published median FFM. Group medians of different quantities need not be exactly self-consistent, so this is checked at 6%.

ffm_ref <- tibble::tribble(
  ~group,                 ~wt,  ~ht,  ~sexf, ~ffm_published, ~source,
  "Antepartum",           61.5, 1.60, 1,     38.7,           "Table S1 study data, pregnant (3rd trimester)",
  "Postpartum",           57.6, 1.57, 1,     37.0,           "Discussion (median postpartum WT 57.6, FFM 37.0)",
  "Non-pregnant female",  52.9, 1.57, 1,     35.7,           "Table S1 study data, non-pregnant females",
  "Male",                 54.4, 1.73, 0,     46.4,           "Table S1 study data, males",
  "Overall cohort",       58.0, 1.60, 1,     39.4,           "Table 1 overall participants (median WT/HT/FFM)"
) |>
  dplyr::mutate(
    ffm_formula = ffm_janmahasatian(wt, ht, sexf),
    pct_diff    = 100 * (ffm_formula - ffm_published) / ffm_published
  )

ffm_ref |>
  dplyr::transmute(
    Group = group, `WT (kg)` = wt, `HT (m)` = ht,
    `FFM published (kg)` = ffm_published,
    `FFM from formula (kg)` = round(ffm_formula, 1),
    `% diff` = round(pct_diff, 1), Source = source
  ) |>
  knitr::kable(caption = "Janmahasatian FFM reproduces the published group median FFM.")
Janmahasatian FFM reproduces the published group median FFM.
Group WT (kg) HT (m) FFM published (kg) FFM from formula (kg) % diff Source
Antepartum 61.5 1.60 38.7 38.9 0.6 Table S1 study data, pregnant (3rd trimester)
Postpartum 57.6 1.57 37.0 36.9 -0.3 Discussion (median postpartum WT 57.6, FFM 37.0)
Non-pregnant female 52.9 1.57 35.7 35.0 -2.0 Table S1 study data, non-pregnant females
Male 54.4 1.73 46.4 47.5 2.5 Table S1 study data, males
Overall cohort 58.0 1.60 39.4 37.6 -4.6 Table 1 overall participants (median WT/HT/FFM) 

stopifnot(max(abs(ffm_ref$pct_diff)) < 6)

Reconstructing the paper’s decomposition of the pregnancy effect

This is the sharpest available check on the covariate model, because the paper reports the same quantity twice, computed two different ways.

The final model splits the pregnancy effect three ways: part is body size (pregnant women are larger), part is renal function (pregnant women have lower serum creatinine), and the remainder is the estimated +38.1% “pregnancy” term. The Discussion then states that the overall clearance increase due to pregnancy is “around 53%”, the value obtained from an alternative model (Table S3) that omits the serum-creatinine effect entirely.

So: evaluating the final model’s three effects at the published antepartum and postpartum median covariates must reconstruct that ~53%. Nothing in this calculation is fitted to the 53% figure, and the two models were estimated separately, so agreement is a genuine cross-check.

cl_typical <- function(ffm, creat, preg) {
  6.05859 * (ffm / 39.4)^0.75 * (1 + 0.380866 * preg) * (creat / 56.2)^-0.366504
}

# Published medians for the matched antepartum / postpartum visits
# (Discussion: FFM 38.7 vs 37.0 kg; serum creatinine 45.3 vs 55.4 umol/L).
cl_ante <- cl_typical(ffm = 38.7, creat = 45.3, preg = 1)
cl_post <- cl_typical(ffm = 37.0, creat = 55.4, preg = 0)

contributions <- c(
  "Pregnancy term (+38.1%)"        = 1 + 0.380866,
  "Fat-free mass (38.7 vs 37.0 kg)" = (38.7 / 37.0)^0.75,
  "Serum creatinine (45.3 vs 55.4 umol/L)" = (45.3 / 55.4)^-0.366504
)

data.frame(
  Component = c(names(contributions), "Total CL ratio (antepartum / postpartum)"),
  `CL multiplier` = round(c(contributions, prod(contributions)), 4),
  check.names = FALSE
) |>
  knitr::kable(caption = "Decomposed pregnancy effect on levofloxacin clearance.")
Decomposed pregnancy effect on levofloxacin clearance.
Component CL multiplier
Pregnancy term (+38.1%) Pregnancy term (+38.1%) 1.3809
Fat-free mass (38.7 vs 37.0 kg) Fat-free mass (38.7 vs 37.0 kg) 1.0343
Serum creatinine (45.3 vs 55.4 umol/L) Serum creatinine (45.3 vs 55.4 umol/L) 1.0766
Total CL ratio (antepartum / postpartum) 1.5375 

total_ratio <- cl_ante / cl_post
cat(sprintf(
  "Reconstructed total CL increase in pregnancy: %.1f%%\nPaper's alternative model (Table S3, no creatinine effect): 53%%\n",
  100 * (total_ratio - 1)
))
#> Reconstructed total CL increase in pregnancy: 53.8%
#> Paper's alternative model (Table S3, no creatinine effect): 53%

# The three decomposed effects reproduce the independently-estimated 53%.
stopifnot(abs(total_ratio / 1.53 - 1) < 0.06)

The efficacy target drawn on Fig. 2 is likewise reproducible in closed form: the caption defines AUC0-24 >= 146 * MIC / fu with the hollow-fibre fAUC0-24/MIC target of 146, the wild-type MIC of 0.5 mg/L and an unbound fraction of 69%.

auc_target <- 146 * 0.5 / 0.69
cat(sprintf("146 * 0.5 / 0.69 = %.1f mg*h/L (paper states 106)\n", auc_target))
#> 146 * 0.5 / 0.69 = 105.8 mg*h/L (paper states 106)
stopifnot(round(auc_target) == 106)

Structural checks against closed-form results

These compare a solved profile against an algebraic result derived from the published equations, with parameter values computed independently in R rather than read back out of the model. Both sides use the same fixed parameter values, so the residual difference is pure numerical / trapezoidal error and the tolerances are tight.

tau       <- 24
n_dose    <- 10                       # 10 daily doses: well past steady state
t_last    <- (n_dose - 1) * tau       # 216 h, the final ("observed") dose
grid_ss   <- seq(t_last, t_last + tau, by = 0.1)

#' Build a plain data.frame event table (never an rxEt object, which silently
#' drops post-hoc covariate column assignment).
#'
#' Occasions follow the paper's design: every dose before the final one is
#' self-administered at home (OCC = 1, the inflated BOV on F), the final dose
#' is the observed clinic dose (OCC = 2), the pre-dose trough belongs to
#' occasion 1 and the subsequent profile to occasion 2.
build_events <- function(cohort, obs_times = grid_ss) {
  doses <- tidyr::expand_grid(
    cohort,
    time = seq(0, t_last, by = tau)
  ) |>
    dplyr::mutate(
      evid = 1L,
      amt  = dose,
      cmt  = "depot",
      OCC  = ifelse(time < t_last, 1L, 2L)
    )
  obs <- tidyr::expand_grid(cohort, time = obs_times) |>
    dplyr::mutate(
      evid = 0L,
      amt  = NA_real_,
      cmt  = "central",
      OCC  = ifelse(time <= t_last, 1L, 2L)
    )
  dplyr::bind_rows(doses, obs) |>
    dplyr::arrange(id, time, dplyr::desc(evid)) |>
    dplyr::select(id, time, amt, evid, cmt, FFM, CREAT, PREG, OCC,
                  dplyr::any_of(c("group", "dose", "arm", "weight_band")))
}
# Four deterministic typical subjects spanning the covariate space.
typ <- tibble::tribble(
  ~id, ~group,                ~FFM,  ~CREAT, ~PREG, ~dose,
  1L,  "Antepartum",          38.7,  45.3,   1,     1000,
  2L,  "Postpartum",          37.0,  55.4,   0,     1000,
  3L,  "Non-pregnant female", 35.7,  58.5,   0,     1000,
  4L,  "Male",                46.4,  64.1,   0,     1000
)

mod_typ <- rxode2::zeroRe(mod)
#> Warning: some etas defaulted to non-mu referenced, possible parsing error: etaiov_mtt_1, etaiov_mtt_2, etaiov_mtt_3, etaiov_mtt_4, etaiov_ka_1, etaiov_ka_2, etaiov_ka_3, etaiov_ka_4, etaiov_fdepot_1, etaiov_fdepot_2, etaiov_fdepot_3, etaiov_fdepot_4
#> as a work-around try putting the mu-referenced expression on a simple line
sim_typ <- rxode2::rxSolve(mod_typ, build_events(typ), keep = c("group", "dose")) |>
  as.data.frame()
#> Warning: some etas defaulted to non-mu referenced, possible parsing error: etaiov_mtt_1, etaiov_mtt_2, etaiov_mtt_3, etaiov_mtt_4, etaiov_ka_1, etaiov_ka_2, etaiov_ka_3, etaiov_ka_4, etaiov_fdepot_1, etaiov_fdepot_2, etaiov_fdepot_3, etaiov_fdepot_4
#> as a work-around try putting the mu-referenced expression on a simple line
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etaiov_mtt_1', 'etaiov_mtt_2', 'etaiov_mtt_3', 'etaiov_mtt_4', 'etaiov_ka_1', 'etaiov_ka_2', 'etaiov_ka_3', 'etaiov_ka_4', 'etaiov_fdepot_1', 'etaiov_fdepot_2', 'etaiov_fdepot_3', 'etaiov_fdepot_4'
#> Warning: multi-subject simulation without without 'omega'

Steady-state AUC identity

At steady state the amount eliminated over a dosing interval equals the amount absorbed, so CL * AUC(0-tau) = F * dose exactly, whatever the absorption model does. CL here is computed from the published covariate equations, not taken from the solve.

auc_trapz <- function(t, c) sum(diff(t) * (head(c, -1) + tail(c, -1)) / 2)

id_check <- sim_typ |>
  dplyr::filter(time >= t_last) |>
  dplyr::group_by(id, group, dose) |>
  dplyr::summarise(auc_sim = auc_trapz(time, Cc), .groups = "drop") |>
  dplyr::left_join(typ, by = c("id", "group", "dose")) |>
  dplyr::mutate(
    cl_expected  = cl_typical(FFM, CREAT, PREG),
    auc_expected = dose / cl_expected,          # F = 1 (fixed)
    pct_diff     = 100 * (auc_sim - auc_expected) / auc_expected
  )

id_check |>
  dplyr::transmute(
    Group = group, `CL expected (L/h)` = round(cl_expected, 3),
    `AUC dose/CL (mg*h/L)` = round(auc_expected, 1),
    `AUC from solve (mg*h/L)` = round(auc_sim, 1),
    `% diff` = round(pct_diff, 3)
  ) |>
  knitr::kable(caption = "Steady-state AUC(0-24) equals F*dose/CL.")
Steady-state AUC(0-24) equals F*dose/CL.
Group CL expected (L/h) AUC dose/CL (mg*h/L) AUC from solve (mg*h/L) % diff
Antepartum 8.933 111.9 111.9 -0.014
Postpartum 5.810 172.1 172.1 -0.010
Non-pregnant female 5.545 180.4 180.3 -0.009
Male 6.527 153.2 153.2 -0.009 

stopifnot(max(abs(id_check$pct_diff)) < 0.5)

Terminal slope recovers CL/Vc

By 14 h after the dose the transit chain and the depot are empty (ktr = 21/1.07 = 19.6 /h; depot half-life ln2/1.59 = 0.44 h), so the decline is mono-exponential with rate kel = CL/Vc. Regressing log(Cc) on time over 14-24 h post-dose therefore tests lvc and the volume exponent e_ffm_vc, which the AUC identity above cannot see.

kel_check <- sim_typ |>
  dplyr::filter(time >= t_last + 14) |>
  dplyr::group_by(id, group) |>
  dplyr::summarise(
    kel_sim = -stats::coef(stats::lm(log(Cc) ~ time))[["time"]],
    .groups = "drop"
  ) |>
  dplyr::left_join(typ, by = c("id", "group")) |>
  dplyr::mutate(
    vc_expected  = 85.8803 * (FFM / 39.4)^1,
    kel_expected = cl_typical(FFM, CREAT, PREG) / vc_expected,
    pct_diff     = 100 * (kel_sim - kel_expected) / kel_expected
  )

kel_check |>
  dplyr::transmute(
    Group = group, `Vc expected (L)` = round(vc_expected, 1),
    `kel expected (1/h)` = round(kel_expected, 5),
    `kel from slope (1/h)` = round(kel_sim, 5),
    `t1/2 (h)` = round(log(2) / kel_sim, 2),
    `% diff` = round(pct_diff, 3)
  ) |>
  knitr::kable(caption = "Terminal slope recovers CL/Vc; half-life 6-8 h is the published range.")
Terminal slope recovers CL/Vc; half-life 6-8 h is the published range.
Group Vc expected (L) kel expected (1/h) kel from slope (1/h) t1/2 (h) % diff
Antepartum 84.4 0.10590 0.10590 6.55 0
Postpartum 80.6 0.07204 0.07204 9.62 0
Non-pregnant female 77.8 0.07125 0.07125 9.73 0
Male 101.1 0.06453 0.06453 10.74 0 

stopifnot(max(abs(kel_check$pct_diff)) < 0.5)

The recovered half-lives sit inside the 6-8 h that the Introduction quotes for individuals with normal renal function.

Covariate exponents recovered from exposure ratios

Each ratio isolates one covariate by holding the others fixed, so the simulated exposure ratio must equal the algebraic ratio exactly.

ratio_cohort <- tibble::tribble(
  ~id, ~arm,          ~FFM,  ~CREAT, ~PREG, ~dose,
  1L,  "reference",   39.4,  56.2,   0,     1000,
  2L,  "pregnant",    39.4,  56.2,   1,     1000,
  3L,  "creat x2",    39.4,  112.4,  0,     1000,
  4L,  "ffm x1.25",   49.25, 56.2,   0,     1000
) |>
  dplyr::mutate(group = arm)

sim_ratio <- rxode2::rxSolve(mod_typ, build_events(ratio_cohort),
                             keep = c("arm", "dose")) |>
  as.data.frame() |>
  dplyr::filter(time >= t_last) |>
  dplyr::group_by(arm) |>
  dplyr::summarise(auc = auc_trapz(time, Cc), .groups = "drop")
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etaiov_mtt_1', 'etaiov_mtt_2', 'etaiov_mtt_3', 'etaiov_mtt_4', 'etaiov_ka_1', 'etaiov_ka_2', 'etaiov_ka_3', 'etaiov_ka_4', 'etaiov_fdepot_1', 'etaiov_fdepot_2', 'etaiov_fdepot_3', 'etaiov_fdepot_4'
#> Warning: multi-subject simulation without without 'omega'

auc_of <- function(a) sim_ratio$auc[match(a, sim_ratio$arm)]

# At steady state AUC = F*dose/CL, so the exposure ratio between two arms is
# the inverse ratio of their clearances: AUC(ref)/AUC(arm) = CL(arm)/CL(ref).
# Doubling creatinine LOWERS clearance (the exponent is negative), so that
# ratio is below 1; raising FFM RAISES clearance, so that ratio is above 1.
ratios <- tibble::tribble(
  ~Effect,                              ~observed,                                   ~expected,
  "Pregnancy: AUC(ref)/AUC(pregnant)",  auc_of("reference") / auc_of("pregnant"),    1 + 0.380866,
  "Creatinine x2: AUC(ref)/AUC(2x)",    auc_of("reference") / auc_of("creat x2"),    2^-0.366504,
  "FFM x1.25: AUC(ref)/AUC(1.25x)",     auc_of("reference") / auc_of("ffm x1.25"),   1.25^0.75
) |>
  dplyr::mutate(pct_diff = 100 * (observed - expected) / expected)

ratios |>
  dplyr::transmute(
    Effect, `Expected ratio` = round(expected, 5),
    `Simulated ratio` = round(observed, 5), `% diff` = round(pct_diff, 4)
  ) |>
  knitr::kable(caption = "Exposure ratios recover the published covariate exponents.")
Exposure ratios recover the published covariate exponents.
Effect Expected ratio Simulated ratio % diff
Pregnancy: AUC(ref)/AUC(pregnant) 1.38087 1.38092 0.0036
Creatinine x2: AUC(ref)/AUC(2x) 0.77566 0.77565 -0.0019
FFM x1.25: AUC(ref)/AUC(1.25x) 1.18218 1.18217 -0.0005 

stopifnot(max(abs(ratios$pct_diff)) < 0.2)

Virtual cohort

The cohort mirrors the four participant groups in the numbers in which they contributed profiles (19 antepartum, 14 postpartum, 12 non-pregnant female, 14 male), scaled 10-fold and capped at 200 per arm. Weight, height and serum creatinine are drawn log-normally around each group’s published median (Table S1 study-data columns; postpartum medians from the Discussion) and truncated to the published range; FFM is then derived per subject with the Janmahasatian formula, and dose is assigned by WHO weight band. The dispersion of the drawn covariates is an assumption – Sawe 2025 reports medians and ranges, not distributions – and is listed under Assumptions below.

group_spec <- tibble::tribble(
  ~group,                ~n,   ~preg, ~sexf, ~wt_med, ~wt_lo, ~wt_hi, ~ht_med, ~creat_med, ~creat_lo, ~creat_hi,
  "Antepartum",          190L, 1,     1,     61.5,    43.0,   98.0,   1.60,    45.3,       37.0,      66.4,
  "Postpartum",          140L, 0,     1,     57.6,    45.0,   83.5,   1.57,    55.4,       25.3,      63.0,
  "Non-pregnant female", 120L, 0,     1,     52.9,    37.0,   67.9,   1.57,    58.5,       37.0,      110,
  "Male",                140L, 0,     0,     54.4,    41.9,   64.0,   1.73,    64.1,       52.3,      102
)

rxode2::rxSetSeed(20250401)   # seeded per stochastic block, not once globally
set.seed(20250401)

draw_group <- function(spec) {
  n <- spec$n
  tibble::tibble(
    group = spec$group,
    PREG  = spec$preg,
    SEXF  = spec$sexf,
    WT    = pmin(pmax(spec$wt_med    * exp(stats::rnorm(n, 0, 0.20)), spec$wt_lo),    spec$wt_hi),
    HT    = spec$ht_med * exp(stats::rnorm(n, 0, 0.045)),
    CREAT = pmin(pmax(spec$creat_med * exp(stats::rnorm(n, 0, 0.25)), spec$creat_lo), spec$creat_hi)
  )
}

cohort <- dplyr::bind_rows(
  lapply(seq_len(nrow(group_spec)), function(i) draw_group(group_spec[i, ]))
) |>
  dplyr::mutate(
    FFM         = ffm_janmahasatian(WT, HT, SEXF),
    weight_band = ifelse(WT > 50, "above 50 kg", "34-50 kg"),
    dose        = ifelse(WT > 50, 1000, 750),
    id          = dplyr::row_number()
  )

cohort |>
  dplyr::group_by(group) |>
  dplyr::summarise(
    n = dplyr::n(),
    `WT median` = round(stats::median(WT), 1),
    `FFM median` = round(stats::median(FFM), 1),
    `CREAT median` = round(stats::median(CREAT), 1),
    `% on 1000 mg` = round(100 * mean(dose == 1000)),
    .groups = "drop"
  ) |>
  knitr::kable(caption = "Simulated cohort versus the published group medians.")
Simulated cohort versus the published group medians.
group n WT median FFM median CREAT median % on 1000 mg
Antepartum 190 61.8 39.3 44.5 83
Male 140 54.0 47.7 64.1 71
Non-pregnant female 120 53.3 35.1 56.0 61
Postpartum 140 57.1 36.8 57.3 73 
rxode2::rxSetSeed(20250402)
sim <- rxode2::rxSolve(
  mod,
  build_events(cohort, obs_times = seq(t_last, t_last + tau, by = 0.25)),
  keep = c("group", "dose", "weight_band")
) |>
  as.data.frame()
#> Warning: some etas defaulted to non-mu referenced, possible parsing error: etaiov_mtt_1, etaiov_mtt_2, etaiov_mtt_3, etaiov_mtt_4, etaiov_ka_1, etaiov_ka_2, etaiov_ka_3, etaiov_ka_4, etaiov_fdepot_1, etaiov_fdepot_2, etaiov_fdepot_3, etaiov_fdepot_4
#> as a work-around try putting the mu-referenced expression on a simple line

Replicating Figure 1: VPC stratified by pregnancy status

Figure 1 of Sawe 2025 is a visual predictive check of concentration versus time after dose, stratified into a pregnant and a non-pregnant stratum, where the non-pregnant stratum pools the postpartum, never-pregnant-female and male records. Its message is that median exposure is lower in the pregnant stratum.

vpc_dat <- sim |>
  dplyr::mutate(
    tad     = time - t_last,
    stratum = ifelse(PREG == 1, "Pregnant", "Non-pregnant")
  ) |>
  dplyr::filter(!is.na(Cc))

vpc_pct <- vpc_dat |>
  dplyr::group_by(stratum, tad) |>
  dplyr::summarise(
    p05 = stats::quantile(Cc, 0.05),
    p50 = stats::median(Cc),
    p95 = stats::quantile(Cc, 0.95),
    .groups = "drop"
  )

ggplot2::ggplot(vpc_pct, ggplot2::aes(tad)) +
  ggplot2::geom_ribbon(ggplot2::aes(ymin = p05, ymax = p95), alpha = 0.25,
                       fill = "steelblue") +
  ggplot2::geom_line(ggplot2::aes(y = p50), linewidth = 0.9) +
  ggplot2::geom_line(ggplot2::aes(y = p05), linetype = "dashed") +
  ggplot2::geom_line(ggplot2::aes(y = p95), linetype = "dashed") +
  ggplot2::facet_wrap(~stratum) +
  ggplot2::labs(x = "Time after dose (h)", y = "Levofloxacin concentration (mg/L)") +
  ggplot2::theme_bw()
Replicates Figure 1 of Sawe 2025: simulated 5th, 50th and 95th percentiles of levofloxacin concentration versus time after dose, stratified by pregnancy status.

Replicates Figure 1 of Sawe 2025: simulated 5th, 50th and 95th percentiles of levofloxacin concentration versus time after dose, stratified by pregnancy status. 

med_by_stratum <- vpc_pct |>
  dplyr::group_by(stratum) |>
  dplyr::summarise(auc_med = auc_trapz(tad, p50), .groups = "drop")

cat(sprintf(
  "Median-profile AUC(0-24): pregnant %.0f, non-pregnant %.0f mg*h/L (ratio %.2f)\n",
  med_by_stratum$auc_med[med_by_stratum$stratum == "Pregnant"],
  med_by_stratum$auc_med[med_by_stratum$stratum == "Non-pregnant"],
  med_by_stratum$auc_med[med_by_stratum$stratum == "Pregnant"] /
    med_by_stratum$auc_med[med_by_stratum$stratum == "Non-pregnant"]
))
#> Median-profile AUC(0-24): pregnant 108, non-pregnant 153 mg*h/L (ratio 0.70)

# The paper's central qualitative finding: median exposure is lower in the
# pregnant stratum. Asserted on the ratio of median profiles, which is robust
# to which subjects land in the tails.
stopifnot(
  med_by_stratum$auc_med[med_by_stratum$stratum == "Pregnant"] <
    med_by_stratum$auc_med[med_by_stratum$stratum == "Non-pregnant"]
)

PKNCA validation 

conc_df <- sim |>
  dplyr::filter(!is.na(Cc)) |>
  dplyr::select(id, group, weight_band, dose, time, Cc)

dose_df <- cohort |>
  dplyr::transmute(id, group, weight_band, dose, time = t_last)

conc_obj <- PKNCA::PKNCAconc(
  data = conc_df, formula = Cc ~ time | group + id,
  concu = "mg/L", timeu = "h"
)
dose_obj <- PKNCA::PKNCAdose(
  data = dose_df, formula = dose ~ time | group + id,
  doseu = "mg"
)

intervals <- data.frame(
  start = t_last, end = t_last + tau,
  cmax = TRUE, tmax = TRUE, cmin = TRUE, auclast = TRUE, cav = TRUE
)

nca <- PKNCA::pk.nca(PKNCA::PKNCAdata(conc_obj, dose_obj, intervals = intervals))
nca_res <- as.data.frame(nca$result)
nca_wide <- nca_res |>
  dplyr::filter(PPTESTCD %in% c("cmax", "tmax", "auclast", "cav", "cmin")) |>
  dplyr::group_by(group, PPTESTCD) |>
  dplyr::summarise(
    median = stats::median(PPORRES),
    q25    = stats::quantile(PPORRES, 0.25),
    q75    = stats::quantile(PPORRES, 0.75),
    .groups = "drop"
  )

nca_wide |>
  dplyr::mutate(
    Parameter = dplyr::recode(PPTESTCD,
      cmax = "Cmax (mg/L)", tmax = "Tmax (h)", auclast = "AUC0-24 (mg*h/L)",
      cav = "Cavg (mg/L)", cmin = "Ctrough (mg/L)"
    ),
    `Median (IQR)` = sprintf("%.2f (%.2f-%.2f)", median, q25, q75)
  ) |>
  dplyr::select(Group = group, Parameter, `Median (IQR)`) |>
  tidyr::pivot_wider(names_from = Group, values_from = `Median (IQR)`) |>
  knitr::kable(caption = "Simulated steady-state NCA by participant group.")
Simulated steady-state NCA by participant group.
Parameter Antepartum Male Non-pregnant female Postpartum
AUC0-24 (mg*h/L) 106.33 (89.97-126.93) 143.40 (115.04-165.11) 162.34 (132.33-185.12) 159.94 (134.62-190.47)
Cavg (mg/L) 4.43 (3.75-5.29) 5.98 (4.79-6.88) 6.76 (5.51-7.71) 6.66 (5.61-7.94)
Cmax (mg/L) 10.07 (9.02-11.13) 10.00 (8.88-11.10) 12.47 (10.72-13.64) 12.18 (10.60-13.72)
Ctrough (mg/L) 1.16 (0.70-1.66) 2.67 (1.81-3.54) 2.69 (1.85-3.57) 2.68 (1.88-3.56)
Tmax (h) 1.75 (1.25-2.75) 1.75 (1.25-3.00) 1.88 (1.25-2.75) 1.75 (1.25-2.75)

The paper reports exposures for the pooled cohort: median (IQR) AUC0-24 131 (108-170) mg*h/L and Cmax 11.3 (9.68-13.7) mg/L (Results; Fig. S4).

pooled <- nca_res |>
  dplyr::filter(PPTESTCD %in% c("cmax", "tmax", "auclast"))

reference <- data.frame(
  cmax    = 11.3,   # Results / Fig. S4, pooled observed median
  auclast = 131,
  tmax    = 2       # first post-dose sampling time; see Assumptions
)

cmp <- nlmixr2lib::ncaComparisonTable(
  simulated = pooled |> dplyr::select(PPTESTCD, PPORRES),
  reference = reference,
  units = c(cmax = "mg/L", auclast = "mg*h/L", tmax = "h")
)
knitr::kable(cmp, caption = "Simulated versus published pooled observed NCA.")
Simulated versus published pooled observed NCA.
NCA parameter Reference Simulated % diff
Cmax (mg/L) 11.3 10.8 -4.5%
Tmax (h) 2 1.75 -12.5%
AUClast (mg*h/L) 131 137 +4.6% 
attr(cmp, "footnote")
#> NULL
pooled_med <- pooled |>
  dplyr::group_by(PPTESTCD) |>
  dplyr::summarise(median = stats::median(PPORRES), .groups = "drop")

auc_med  <- pooled_med$median[pooled_med$PPTESTCD == "auclast"]
cmax_med <- pooled_med$median[pooled_med$PPTESTCD == "cmax"]

cat(sprintf("Pooled simulated median AUC0-24 %.0f mg*h/L (observed IQR 108-170)\n", auc_med))
#> Pooled simulated median AUC0-24 137 mg*h/L (observed IQR 108-170)
cat(sprintf("Pooled simulated median Cmax    %.1f mg/L   (observed IQR 9.68-13.7)\n", cmax_med))
#> Pooled simulated median Cmax    10.8 mg/L   (observed IQR 9.68-13.7)

# Assert against the published interquartile range rather than the point
# median: the simulated cohort's covariate dispersion is an assumption, so the
# defensible claim is that the centre of the simulated distribution falls
# inside the observed IQR.
stopifnot(
  auc_med  > 108,  auc_med  < 170,
  cmax_med > 9.68, cmax_med < 13.7
)

# Table S2 range of median AUC0-24 (98.8-145) and Cmax (7.40-14.86) reported
# for adults with drug-resistant TB on 750-1000 mg; the paper's Fig. 2 ribbon.
lit_auc  <- c(98.8, 145)
lit_cmax <- c(7.40, 14.86)
stopifnot(cmax_med > lit_cmax[1], cmax_med < lit_cmax[2])

# Antepartum exposure must sit below all three non-pregnant groups.
group_auc <- nca_res |>
  dplyr::filter(PPTESTCD == "auclast") |>
  dplyr::group_by(group) |>
  dplyr::summarise(median = stats::median(PPORRES), .groups = "drop")
knitr::kable(
  group_auc |> dplyr::transmute(Group = group, `Median AUC0-24 (mg*h/L)` = round(median)),
  caption = "Antepartum exposure is the lowest of the four groups."
)
Antepartum exposure is the lowest of the four groups.
Group Median AUC0-24 (mg*h/L)
Antepartum 106
Male 143
Non-pregnant female 162
Postpartum 160 
stopifnot(
  group_auc$median[group_auc$group == "Antepartum"] ==
    min(group_auc$median)
)

Replicating Figure 2: the +250 mg dose recommendation

Figure 2 plots simulated steady-state AUC0-24 and Cmax against weight band with serum creatinine split into tertiles, adds a dashed box for the proposed extra 250 mg once daily in pregnancy, ribbons the literature range of reported medians, and draws the 106 mg*h/L efficacy line. The paper concludes that a 250 mg increase brings pregnant exposures up to those of non-pregnant participants across both weight bands.

rxode2::rxSetSeed(20250403)

preg_extra <- cohort |>
  dplyr::filter(group == "Antepartum") |>
  dplyr::mutate(
    group = "Antepartum +250 mg",
    dose  = dose + 250,
    id    = id + 10000L
  )

cohort2 <- dplyr::bind_rows(cohort, preg_extra)

sim2 <- rxode2::rxSolve(
  mod,
  build_events(cohort2, obs_times = seq(t_last, t_last + tau, by = 0.5)),
  keep = c("group", "dose", "weight_band")
) |>
  as.data.frame()

nca2 <- PKNCA::pk.nca(PKNCA::PKNCAdata(
  PKNCA::PKNCAconc(
    data = sim2 |> dplyr::filter(!is.na(Cc)) |>
      dplyr::select(id, group, weight_band, time, Cc),
    formula = Cc ~ time | group + weight_band + id, concu = "mg/L", timeu = "h"
  ),
  PKNCA::PKNCAdose(
    data = cohort2 |> dplyr::transmute(id, group, weight_band, dose, time = t_last),
    formula = dose ~ time | group + weight_band + id, doseu = "mg"
  ),
  intervals = data.frame(start = t_last, end = t_last + tau,
                         cmax = TRUE, auclast = TRUE)
))

exposure <- as.data.frame(nca2$result) |>
  dplyr::filter(PPTESTCD %in% c("cmax", "auclast")) |>
  dplyr::left_join(
    cohort2 |> dplyr::select(id, CREAT), by = "id"
  ) |>
  dplyr::group_by(group, weight_band) |>
  # Sawe 2025 describes the split as tertiles "within each group of
  # participants, based on the weight, sex, and pregnancy status", i.e. a
  # data-driven division into three equal parts. dplyr::ntile() does exactly
  # that and, unlike cut() on sample quantiles, tolerates the ties produced by
  # truncating the drawn creatinine values to the published range.
  dplyr::mutate(
    creat_tertile = factor(
      dplyr::ntile(CREAT, 3), levels = 1:3, labels = c("low", "medium", "high")
    )
  ) |>
  dplyr::ungroup()
ribbons <- data.frame(
  PPTESTCD = c("auclast", "cmax"),
  ymin     = c(98.8, 7.40),
  ymax     = c(145, 14.86)
)
targets <- data.frame(PPTESTCD = "auclast", yint = auc_target)

exposure |>
  dplyr::mutate(
    panel = dplyr::recode(PPTESTCD, auclast = "AUC0-24 (mg*h/L)", cmax = "Cmax (mg/L)")
  ) |>
  ggplot2::ggplot(ggplot2::aes(creat_tertile, PPORRES, fill = group)) +
  ggplot2::geom_rect(
    data = ribbons |> dplyr::mutate(
      panel = dplyr::recode(PPTESTCD, auclast = "AUC0-24 (mg*h/L)", cmax = "Cmax (mg/L)")
    ),
    ggplot2::aes(ymin = ymin, ymax = ymax, xmin = -Inf, xmax = Inf),
    inherit.aes = FALSE, fill = "grey70", alpha = 0.35
  ) +
  ggplot2::geom_hline(
    data = targets |> dplyr::mutate(panel = "AUC0-24 (mg*h/L)"),
    ggplot2::aes(yintercept = yint), inherit.aes = FALSE,
    colour = "red", linetype = "dashed"
  ) +
  ggplot2::geom_boxplot(outlier.size = 0.4, position = ggplot2::position_dodge(0.8)) +
  ggplot2::facet_grid(panel ~ weight_band, scales = "free_y", switch = "y") +
  ggplot2::labs(x = "Serum creatinine tertile", y = NULL, fill = NULL) +
  ggplot2::theme_bw() +
  ggplot2::theme(legend.position = "bottom", strip.placement = "outside")
Replicates Figure 2 of Sawe 2025: simulated steady-state AUC0-24 (top) and Cmax (bottom) by weight band and serum-creatinine tertile. Grey ribbon is the Table S2 literature range of reported adult medians; red dashed line is the 106 mg*h/L efficacy target.

Replicates Figure 2 of Sawe 2025: simulated steady-state AUC0-24 (top) and Cmax (bottom) by weight band and serum-creatinine tertile. Grey ribbon is the Table S2 literature range of reported adult medians; red dashed line is the 106 mg*h/L efficacy target. 

auc_by_group <- exposure |>
  dplyr::filter(PPTESTCD == "auclast") |>
  dplyr::group_by(group, weight_band) |>
  dplyr::summarise(median = stats::median(PPORRES), .groups = "drop") |>
  tidyr::pivot_wider(names_from = group, values_from = median)

auc_by_group |>
  dplyr::mutate(dplyr::across(-weight_band, ~round(.x))) |>
  knitr::kable(caption = "Median simulated steady-state AUC0-24 by weight band.")
Median simulated steady-state AUC0-24 by weight band.
weight_band Antepartum Antepartum +250 mg Male Non-pregnant female Postpartum
34-50 kg 100 129 123 152 146
above 50 kg 107 135 148 172 169 

recovery <- auc_by_group |>
  dplyr::transmute(
    weight_band,
    base_gap     = 100 * (`Antepartum` / `Non-pregnant female` - 1),
    with_250     = 100 * (`Antepartum +250 mg` / `Non-pregnant female` - 1)
  )
knitr::kable(
  recovery |> dplyr::transmute(
    `Weight band` = weight_band,
    `Antepartum vs non-pregnant female, base dose (%)` = round(base_gap),
    `... with +250 mg (%)` = round(with_250)
  ),
  caption = "The proposed 250 mg increase closes most of the exposure gap."
)
The proposed 250 mg increase closes most of the exposure gap.
Weight band Antepartum vs non-pregnant female, base dose (%) … with +250 mg (%)
34-50 kg -34 -15
above 50 kg -38 -21 

# Antepartum exposure is below non-pregnant at the base dose, and the +250 mg
# arm closes most of the gap in both weight bands. The residual shortfall is
# expected: fully offsetting a 38.1% clearance increase needs a 1.381-fold
# dose, i.e. +286 mg on 750 mg and +381 mg on 1000 mg, and the paper chose
# 250 mg because it is the available tablet increment.
stopifnot(
  all(recovery$base_gap < -10),
  all(recovery$with_250 > recovery$base_gap),
  all(abs(recovery$with_250) < abs(recovery$base_gap))
)

for (i in seq_len(nrow(auc_by_group))) {
  cat(sprintf(
    "%-12s | antepartum %.0f | +250 mg %.0f | postpartum %.0f | non-preg F %.0f | male %.0f  (target %.0f)\n",
    auc_by_group$weight_band[i], auc_by_group$`Antepartum`[i],
    auc_by_group$`Antepartum +250 mg`[i], auc_by_group$`Postpartum`[i],
    auc_by_group$`Non-pregnant female`[i], auc_by_group$`Male`[i], auc_target
  ))
}
#> 34-50 kg     | antepartum 100 | +250 mg 129 | postpartum 146 | non-preg F 152 | male 123  (target 106)
#> above 50 kg  | antepartum 107 | +250 mg 135 | postpartum 169 | non-preg F 172 | male 148  (target 106)

# The three non-pregnant arms clear the 106 mg*h/L efficacy target in both
# weight bands, and so does the pregnant arm once the proposed 250 mg is
# added. The pregnant arm on the CURRENT dose is deliberately NOT asserted to
# clear it: that is the paper's point. Its median lands just above the target
# in the above-50 kg band and below it in the 34-50 kg band, which is why the
# paper says AUC is larger than 106 in "most" subjects rather than all, and
# why it recommends the increase.
stopifnot(
  all(auc_by_group$`Postpartum`          > auc_target),
  all(auc_by_group$`Non-pregnant female` > auc_target),
  all(auc_by_group$`Male`                > auc_target),
  all(auc_by_group$`Antepartum +250 mg`  > auc_target)
)

# Antepartum on the current dose is the only arm at risk of missing the target,
# and the lighter weight band is where it misses.
stopifnot(
  auc_by_group$`Antepartum`[auc_by_group$weight_band == "34-50 kg"] < auc_target
)

Both weight bands behave as the paper describes: antepartum medians sit materially below the non-pregnant-female medians on the same weight-based dose, and adding 250 mg once daily recovers most of the shortfall without pushing exposures above the range reported in the comparator literature.

Assumptions and deviations

Where the paper did not state something the simulation needed, the choice is recorded here.

  • Covariate dispersion in the virtual cohort. Sawe 2025 reports medians and ranges (Table 1, Table S1), not distributions. Weight and serum creatinine are drawn log-normally at 20% and 25% CV respectively and height at 4.5% CV, each truncated to the published range. Group medians are matched (see the cohort table); the spread is an assumption and the interquartile ranges of the simulated NCA parameters should not be read as replications of the paper’s.
  • Postpartum median covariates come from the Discussion (weight 57.6 kg, FFM 37.0 kg, serum creatinine 55.4 umol/L) rather than Table 1, which splits postpartum across the two studies. Postpartum height is taken as 1.57 m, the midpoint of the two studies’ medians (1.58 and 1.56 m).
  • Occasion assignment. The model’s OCC covariate is set to 1 for every self-administered dose before the final one and 2 for the observed clinic dose, with the pre-dose trough on occasion 1 and the subsequent profile on occasion 2. This mirrors the design described in the supplementary material (“occasion 1 is linked to the dose taken on the day prior to sampling and its associated concentration (trough sample)”), and it is what makes the inflated BOV on bioavailability apply to the trough only. A user simulating a prospective regimen in which every dose is observed should pass an even occasion number throughout.
  • Between-occasion variability degenerates within an arm. Because each simulated subject is given a single occasion value per dose class rather than a fresh occasion per dose, the BOV on MTT, ka and F is drawn once per subject per class and behaves like between-subject variability within that class. That is the correct reading of the paper’s design, where each participant contributed one observed profile, but it means the simulation does not show dose-to-dose absorption variability within a subject.
  • Tmax has no published counterpart. The paper reports no NCA Tmax. The reference value of 2 h in the comparison table is the first post-dose sampling time in both studies, so the observed Tmax could not have been earlier than that; the row is informational and is not asserted on.
  • Simulated versus observed exposure. The pooled simulated median AUC0-24 and Cmax are asserted to fall inside the published observed interquartile ranges rather than to match the published medians, because the observed medians come from 59 profiles with the real joint covariate distribution and real dose assignment, while the simulation uses drawn covariates and a weight-band dose rule.

Errata and omissions relative to the source

  • Table 2 footnote correction. The article was published on 1 April 2025 with an error in a Table 2 footnote and corrected on 17 April 2025. The corrected footnote d defines the reported variability percentages as %CV = sqrt(omega^2) * 100. This model file uses the corrected reading, verified three ways: against the footnote as typeset in the corrected PDF, against the supplementary equation image, and arithmetically against the control stream $OMEGA values (see the variance-scale table above). Applying the more common omega^2 = log(CV^2 + 1) conversion instead would understate the BOV on ka by 26% on the variance scale.
  • Additive residual error is 0.259443 mg/L, not the 0.244 mg/L of Table 2. The control stream’s $ERROR block sets ADD = THETA(6) + (LLOQ*0.2) unconditionally on every record, with LLOQ 0.0781 mg/L, so the additive standard deviation actually applied is 0.243823 + 0.01562. The same 20%-of-LLOQ floor appears in the sibling paediatric model Denti_2018_levofloxacin.
  • BLQ-handling error inflations are omitted. The control stream adds a further 50% of the LLOQ to the additive error for the imputed BLQ records (Beal’s M6) and substitutes a huge fixed additive term to drop trailing BLQ records from the fit. Both are conditional on the CENS data column, apply to no simulated observation, and are deliberately not encoded.
  • Peripheral-compartment parameters are absent. The control stream retains $THETA and $OMEGA slots for V3, Q, V4 and Q2, all fixed to zero: the residue of testing a second and third compartment. Adding a second compartment did not improve the fit (dOFV 0.674, 2 df), so the packaged model is one-compartment, as Table 2 reports.
  • Between-subject variability is on clearance only. Every other $OMEGA in the control stream is BLOCK(1) FIX 0, and between-visit variability on clearance was tested and not retained. No IIV has been invented for volume or the absorption parameters.
  • No f(depot) <- 0 line, despite the control stream’s F1 = 0. The NONMEM stream sets F1 = 0 so the dose enters only through the analytic transit input rate rather than as a bolus into the absorption compartment. Transcribing that literally as f(depot) <- 0 is the natural move, and several older transit models in this package do exactly that, but under rxUi with rxode2 5.1.7 it silently zeroes the transit() input as well and the model absorbs nothing – every simulated concentration comes back 0, with no error or warning. Omitting the line gives the intended behaviour, not a double-counted dose: transit() is already the sole input pathway under rxUi. The steady-state AUC identity above is the standing regression gate, and a single-dose check gives AUC(0-inf) * CL = 999.98 mg against a 1000 mg dose with F fixed at 1.
  • The transit input accounts only for the most recent dose. The published analytic Savic form tracks a single dose time (TDOS/PNXD in the control stream), so contributions from earlier doses are not superposed. This is faithful to the source and numerically immaterial here: with ktr = 19.6 /h the chain delivers essentially the whole dose within a few hours of a 24-hour interval.
  • Screened-but-unretained covariates (total body weight as an alternative size descriptor, sex, age, albumin, HIV status, gestational age, study and study site) are recorded in the model file’s covariatesDataExcluded for provenance. None is referenced in model().
  • Gestational age. Only third-trimester data were collected, and gestational age was tested within the pregnant participants and found not significant, so the model must not be used to interpolate a pregnancy effect at earlier gestation.
  • Serum creatinine extrapolation. Values were measured within two weeks of the PK visit rather than on the day, and no validated renal-function equation exists for pregnancy; the authors caution against extrapolating outside the observed 25.3-110 umol/L range.