Skip to contents

Model and source

  • Citation: Ajavon-Hartmann A, Jarugula P, Back H, Obianom O, Ludwig E, Crevar C, Marchisin D, Wang Z, Chen W, He B, Payne RM, Burns KM, Glatz AC, Male C, Monagle P, Perera V, Murthy B, Merali S. Characterization of Apixaban Pharmacokinetics/Pharmacodynamics and Dose Assessment in Pediatric Patients with Congenital or Acquired Heart Disease. Clin Pharmacol Ther. 2025;118:365-377. doi:10.1002/cpt.3689
  • Description: Population PK model for oral apixaban in adults and pediatric patients, including children with congenital or acquired heart disease (SAXOPHONE) (Ajavon-Hartmann 2025): two-compartment with first-order absorption and elimination, dose-dependent relative bioavailability, body-weight power effects on CL/F and Vc/F, a fixed literature CYP maturation function on CL/F, age on Q/F, a 9-month-to-18-year age-group effect and an ALL/LL patient-type effect on ka, and a congenital-or-acquired-heart-disease effect on CL/F, with a zero-intercept linear mixed-effects PK/PD relationship to anti-factor Xa activity.
  • Article: https://doi.org/10.1002/cpt.3689 (open access, PMC12272319)
  • Supplement: Data S1 of the article (estimation methods, pediatric demographics, dataset composition, exposure summaries by age group, and the model-development schematic)

Ajavon-Hartmann 2025 added the sparse PK data of the phase II SAXOPHONE trial (children aged 28 days to <18 years with congenital or acquired heart disease, CAHD, needing chronic anticoagulation for thromboembolism prevention) to the pooled adult and pediatric apixaban population PK model. It then re-tested the patient-type covariates and used the model to check the trial’s fixed-dose weight-tiered regimen against adult exposure. The companion paper (Jarugula 2025, Jarugula_2025_apixaban in this package) is the same model fitted before the SAXOPHONE data were added. A zero-intercept linear mixed-effects regression relates anti-factor Xa activity (AXA) to apixaban concentration.

Compared with the Jarugula 2025 fit, backward elimination here kept the ALL/LL patient-type effect on ka, dropped it from Vc/F, and added a CAHD effect on CL/F (Figure S1 of the supplement).

Population

The final dataset held 6098 apixaban concentrations from 602 patients in 12 studies: 198 adults from eight adult studies, 56 pediatric patients at risk of venous or arterial thromboembolism from two phase I studies (CV185079, CV185118), 224 PREVAPIX-ALL patients with acute lymphoblastic leukemia or lymphoblastic lymphoma on asparaginase (CV185155), and 124 SAXOPHONE patients with CAHD (CV185362, 447 concentrations; Table S2). Table S1 describes the 404 pediatric patients: age 0.044 to 18 years (median 6.73), baseline weight 3 to 133 kg (median 22.6), 47.3% female, 78.5% White, 7.43% Asian and 6.44% Black. The SAXOPHONE patients were 0.378 to 18 years (median 7.5) and 6.1 to 133 kg (median 23.6); none was younger than 4 months. Adult demographics are not reported. The PK/PD analysis used 390 AXA values from 123 PREVAPIX-ALL and SAXOPHONE patients, measured with a drug-calibrated assay in ng/mL.

The same information is available programmatically via readModelDb("AjavonHartmann_2025_apixaban")()$population.

Source trace

Each ini() entry carries an in-file comment naming its source. The table collects them. The functional forms come from the parameter-covariate equations under Table 1.

Equation / parameter Value Source location
lka (adult ka) log(0.525) 1/h Table 1, theta1
lcl (CL/F, 67.4 kg, mature) log(4.7) L/h Table 1, theta2
lvc (Vc/F, 67.4 kg) log(32.5) L Table 1, theta3
lq (Q/F, 25 years) log(1.43) L/h Table 1, theta4
lvp (Vp/F) log(20.8) L Table 1, theta5
lgamma (F1 dose shape) log(0.89) Table 1, theta6
logitifdepot50 (logit max F1 reduction) -0.457 Table 1, theta7
e_wt_vc 0.882 Table 1, theta8
e_wt_cl 0.68 Table 1, theta9
e_age_q 0.751 Table 1, theta10
e_ped9m_ka 1.23 Table 1, theta11
lhill (ontogeny shape) log(0.83), fixed Table 1, theta12
lpna50 (ontogeny AGE50) log(0.244) years, fixed Table 1, theta13
e_form_capsule_fdepot 1.1, fixed Table 1, theta14
fbirth_cl 0.05, fixed Table 1, theta15
e_dis_all_ka -0.988 Table 1, theta16
e_dis_heart_congenital_acquired_cl -0.185 Table 1, theta17
etalka 0.399 Table 1, ZKA
etalcl, etalvc, covariance 0.149, 0.0719, 0.0838 Table 1, ZCL, ZVC, ZVC:ZCL
etalvp, etalq, covariance 1.39, 0.989, 1.07 Table 1, ZVP, ZQ, ZQ:ZVP
expSd 0.328 Table 1, sigma 1,1 = 0.108 (SD 0.328), additive on log concentration
slope_antiFXa 0.916 Table 3, slope
etaslope_antiFXa 0.0528^2 Table 3, random-effect standard deviation 0.0528
addSd_antiFXa 12.8 ng/mL Table 3, residual 12.8
ka equation theta1 * (1 + PED9M * theta11) * exp(theta16 if PTYPE = ALL) Table 1 equations
CL/F equation incl. ontogeny theta2 * (WT/67.4)^theta9 * MF * exp(theta17 if PTYPE = CAHD) Table 1 equations and MF footnote
Vc/F, Q/F, Vp/F equations see model file Table 1 equations
F1 equation 1 for dose <= 2.5 mg, else 1 - expit(theta7) * ((DOSE - 2.5)/47.5)^theta6; times 1.1 for sprinkle capsule Table 1 equations
Cc = 1000 * central / vc ng/mL dose in mg, volume in L
antiFXa = (slope + eta) * Cc ng/mL Results and Discussion (PK/PD analysis); Table 3
Residual Cc ~ lnorm(expSd) n/a Results; Supplementary Methods (additive error on log scale)

Typical-value checks

The Results quote derived typical values that pin the covariate structure: ka of 0.525, 1.17 and 0.436 1/h for adults, other children aged 9 months to <18 years, and children with ALL/LL (17% and 62.8% lower for ALL/LL), and CL/F of 3.91 L/h for a 67.4-kg CAHD patient against 4.7 L/h for the reference (16.9% lower). Table 1’s footnote gives the IIV correlations 0.809 and 0.912. These follow from the parameters without any randomness, so they are checked tightly. The paper’s 4.7 and 3.91 L/h are for a fully matured patient; the CL/F column of the table below includes the maturation fraction (0.98 at 25 years, 0.96 at 10 years), which the check divides out.

mod <- readModelDb("AjavonHartmann_2025_apixaban")
mod_typ <- rxode2::zeroRe(mod)
#> ℹ parameter labels from comments will be replaced by 'label()'

# zeroRe() removes the random effects on purpose; rxode2 warns on every
# multi-subject solve without omega, so muffle that one warning only.
solve_typical <- function(...) {
  withCallingHandlers(
    rxode2::rxSolve(...),
    warning = function(w) {
      if (grepl("without 'omega'", conditionMessage(w), fixed = TRUE)) {
        invokeRestart("muffleWarning")
      }
    }
  )
}

typ_cov <- tibble::tibble(
  id = 1:4,
  group = c("Adult", "Pediatric other, 9 mo to <18 y", "Pediatric ALL/LL", "Pediatric CAHD"),
  WT = 67.4,
  AGE = c(25, 10, 10, 10),
  DIS_ALL = c(0, 0, 1, 0),
  DIS_HEART_CONGENITAL_ACQUIRED = c(0, 0, 0, 1),
  DOSE_APIXABAN_MG = 2.5,
  FORM_CAPSULE = 0
)
typ_ev <- dplyr::bind_rows(
  typ_cov |> dplyr::mutate(time = 0, amt = 2.5, evid = 1L, cmt = "depot", dvid = NA_integer_),
  typ_cov |> dplyr::mutate(time = 1, amt = 0, evid = 0L, cmt = "central", dvid = 1L)
) |>
  dplyr::arrange(id, time, dplyr::desc(evid))

typ <- solve_typical(mod_typ, typ_ev, keep = "group", returnType = "data.frame") |>
  dplyr::distinct(id, .keep_all = TRUE) |>
  dplyr::select(group, ka, vc, cl, fmat)
#> ℹ omega/sigma items treated as zero: 'etalka', 'etalcl', 'etalvc', 'etalvp', 'etalq', 'etaslope_antiFXa'

typ |>
  dplyr::rename(
    "Group" = group, "ka (1/h)" = ka, "Vc/F (L)" = vc, "CL/F (L/h)" = cl,
    "Maturation fraction" = fmat
  ) |>
  knitr::kable(digits = 3, caption = "Typical values for a 67.4-kg patient.")
Typical values for a 67.4-kg patient.
Group ka (1/h) Vc/F (L) CL/F (L/h) Maturation fraction
Adult 0.525 32.5 4.606 0.980
Pediatric other, 9 mo to <18 y 1.171 32.5 4.504 0.958
Pediatric ALL/LL 0.436 32.5 4.504 0.958
Pediatric CAHD 1.171 32.5 3.743 0.958

ka_adult <- typ$ka[1]
ka_other <- typ$ka[2]
ka_all <- typ$ka[3]
# The paper's 4.7 vs 3.91 L/h comparison is for a fully matured patient; at
# 10 years the maturation fraction is 0.995, so divide it out.
cl_ref <- typ$cl[2] / typ$fmat[2]
cl_cahd <- typ$cl[4] / typ$fmat[4]

omega <- rxode2::rxode2(mod)$omega
#> ℹ parameter labels from comments will be replaced by 'label()'
r_cl_vc <- omega["etalcl", "etalvc"] / sqrt(omega["etalcl", "etalcl"] * omega["etalvc", "etalvc"])
r_vp_q <- omega["etalvp", "etalq"] / sqrt(omega["etalvp", "etalvp"] * omega["etalq", "etalq"])

stopifnot(
  abs(ka_adult - 0.525) < 5e-4,
  abs(ka_other - 1.17) < 5e-3,
  abs(ka_all - 0.436) < 5e-4,
  abs(100 * (1 - ka_all / ka_adult) - 17) < 0.5,
  abs(100 * (1 - ka_all / ka_other) - 62.8) < 0.1,
  abs(cl_ref - 4.7) < 1e-6,
  abs(cl_cahd - 3.91) < 5e-3,
  abs(100 * (1 - cl_cahd / cl_ref) - 16.9) < 0.05,
  abs(r_cl_vc - 0.809) < 0.005,
  abs(r_vp_q - 0.912) < 0.005
)

Table 2 internal consistency: dose, bioavailability and clearance

Table 2 reports, per weight tier, the geometric means of the SAXOPHONE patients’ individual CL/F and steady-state AUC(TAU). The paper’s AUC equation is dose x F1 / (CL/F). Evaluated at the tier geometric-mean CL/F with the dose of each tier and this model’s F1, it reproduces the Table 2 AUC(TAU) column. That pins the twice-daily doses the trial used (1, 1.5, 2, 3, 4 and 5 mg, the regimen proposed in the Results) and the F1 reduction above 2.5 mg, which only this regimen’s 3-, 4- and 5-mg tiers reach. The geometric mean of a ratio is not the ratio of the geometric means, so the agreement is close but not exact.

t2 <- tibble::tibble(
  tier = c("6 to <9 kg", "9 to <12 kg", "12 to <18 kg", "18 to <25 kg", "25 to <35 kg", ">=35 kg"),
  id = 1:6,
  dose = c(1, 1.5, 2, 3, 4, 5),
  cl_t2 = c(0.682, 0.812, 1.25, 1.64, 2.12, 2.99),
  auc_t2 = c(1460, 1840, 1610, 1760, 1840, 1630)
)
# Solve once per tier only to read the model's F1 at that dose.
f_ev <- t2 |>
  dplyr::transmute(
    id, time = 0, amt = dose, evid = 1L, cmt = "depot", dvid = NA_integer_,
    WT = 20, AGE = 8, DIS_ALL = 0, DIS_HEART_CONGENITAL_ACQUIRED = 1,
    DOSE_APIXABAN_MG = dose, FORM_CAPSULE = 0
  ) |>
  dplyr::bind_rows(
    t2 |> dplyr::transmute(
      id, time = 1, amt = 0, evid = 0L, cmt = "central", dvid = 1L,
      WT = 20, AGE = 8, DIS_ALL = 0, DIS_HEART_CONGENITAL_ACQUIRED = 1,
      DOSE_APIXABAN_MG = dose, FORM_CAPSULE = 0
    )
  ) |>
  dplyr::arrange(id, time, dplyr::desc(evid))
f_tab <- solve_typical(mod_typ, f_ev, returnType = "data.frame") |>
  dplyr::distinct(id, .keep_all = TRUE) |>
  dplyr::select(id, frel)
#> ℹ omega/sigma items treated as zero: 'etalka', 'etalcl', 'etalvc', 'etalvp', 'etalq', 'etaslope_antiFXa'

t2_chk <- dplyr::inner_join(t2, f_tab, by = "id") |>
  dplyr::mutate(
    auc_calc = 1000 * dose * frel / cl_t2,
    pct_diff = 100 * (auc_calc / auc_t2 - 1)
  )
t2_chk |>
  dplyr::select(tier, dose, frel, cl_t2, auc_calc, auc_t2, pct_diff) |>
  dplyr::rename(
    "Weight tier" = tier, "Dose (mg BID)" = dose, "F1" = frel,
    "Table 2 CL/F (L/h)" = cl_t2, "Dose x F1 / CL/F" = auc_calc,
    "Table 2 AUC(TAU)" = auc_t2, "% difference" = pct_diff
  ) |>
  knitr::kable(digits = c(0, 1, 3, 3, 0, 0, 1), caption = "Steady-state AUC(TAU) (ng.h/mL) implied by the Table 2 geometric-mean CL/F.")
Steady-state AUC(TAU) (ng.h/mL) implied by the Table 2 geometric-mean CL/F.
Weight tier Dose (mg BID) F1 Table 2 CL/F (L/h) Dose x F1 / CL/F Table 2 AUC(TAU) % difference
6 to <9 kg 1.0 1.000 0.682 1466 1460 0.4
9 to <12 kg 1.5 1.000 0.812 1847 1840 0.4
12 to <18 kg 2.0 1.000 1.250 1600 1610 -0.6
18 to <25 kg 3.0 0.993 1.640 1817 1760 3.2
25 to <35 kg 4.0 0.982 2.120 1853 1840 0.7
>=35 kg 5.0 0.972 2.990 1625 1630 -0.3

# A wrong dose for a tier, or an F1 reduction applied with the wrong sign or
# anchor, moves the affected tier by 20% or more.
stopifnot(all(abs(t2_chk$pct_diff) < 5))

Virtual cohort

The paper simulated about 1000 virtual patients per weight tier (9 tiers), assigning age, sex and weight as in its companion analyses. The cohorts below keep 200 patients per tier: for each tier, age is drawn uniformly over the ages at which the CDC 2000 sex-averaged median weight falls inside the tier (28 days to <18 years), and weight is the CDC median for that age with 12% lognormal scatter; a weight that falls outside the tier is redrawn uniformly within it. All virtual patients have CAHD (DIS_HEART_CONGENITAL_ACQUIRED = 1) and take the tablet or oral solution (FORM_CAPSULE = 0).

# CDC 2000 50th-percentile weight-for-age, sexes averaged. Reference data for
# building the virtual cohort only, not a model parameter.
cdc_age_yr <- c(0, 1, 2, 3, 4, 6, 9, 12, 18, 24) / 12
cdc_wt_kg <- c(3.4, 4.4, 5.5, 6.3, 7.0, 7.9, 9.0, 9.9, 11.3, 12.5)
cdc_age_yr <- c(cdc_age_yr, 3:18)
cdc_wt_kg <- c(
  cdc_wt_kg, 14.3, 16.3, 18.3, 20.7, 23.2, 26.0, 29.2, 32.7,
  36.7, 41.3, 46.0, 50.6, 54.6, 58.0, 60.6, 62.6
)
cdc_weight_for_age <- function(age_yr) {
  stats::approx(cdc_age_yr, cdc_wt_kg, xout = age_yr, rule = 2)$y
}
cdc_age_for_weight <- function(wt_kg) {
  stats::approx(cdc_wt_kg, cdc_age_yr, xout = wt_kg, rule = 2)$y
}

n_per_tier <- 200L

# One cohort per weight tier; ages restricted to [age_min, age_max).
make_tier <- function(tier, wt_lo, wt_hi, dose, age_min, age_max, id_offset) {
  a_lo <- max(age_min, cdc_age_for_weight(wt_lo))
  a_hi <- min(age_max, cdc_age_for_weight(min(wt_hi, 70)))
  if (a_hi <= a_lo + 1 / 12) a_hi <- a_lo + 1 / 12
  age <- stats::runif(n_per_tier, a_lo, a_hi)
  wt <- cdc_weight_for_age(age) * exp(stats::rnorm(n_per_tier, 0, 0.12))
  # Where the CDC median lies outside the tier (the lightest tiers), clamping
  # would pile patients on one edge; redraw those uniformly inside the tier.
  out <- wt < wt_lo | wt >= wt_hi
  wt[out] <- stats::runif(sum(out), wt_lo, wt_hi)
  tibble::tibble(
    id = id_offset + seq_len(n_per_tier),
    tier = tier,
    WT = wt,
    AGE = age,
    DIS_ALL = 0,
    DIS_HEART_CONGENITAL_ACQUIRED = 1,
    FORM_CAPSULE = 0,
    DOSE_APIXABAN_MG = dose
  )
}

# Proposed regimen for ages 28 days to <18 years (Results, PPK model
# simulations); the six tiers from 6 kg are the SAXOPHONE regimen.
tiers <- tibble::tribble(
  ~tier,                  ~wt_lo, ~wt_hi, ~dose,
  "3 to <4 kg, 0.2 mg",      3.0,    4.0,   0.2,
  "4 to <5 kg, 0.3 mg",      4.0,    5.0,   0.3,
  "5 to <6 kg, 0.5 mg",      5.0,    6.0,   0.5,
  "6 to <9 kg, 1 mg",        6.0,    9.0,   1.0,
  "9 to <12 kg, 1.5 mg",     9.0,   12.0,   1.5,
  "12 to <18 kg, 2 mg",     12.0,   18.0,   2.0,
  "18 to <25 kg, 3 mg",     18.0,   25.0,   3.0,
  "25 to <35 kg, 4 mg",     25.0,   35.0,   4.0,
  ">=35 kg, 5 mg",          35.0,  133.0,   5.0
)

set.seed(20250623)
cohort <- dplyr::bind_rows(lapply(seq_len(nrow(tiers)), function(i) {
  r <- tiers[i, ]
  make_tier(r$tier, r$wt_lo, r$wt_hi, r$dose,
    age_min = 28 / 365.25, age_max = 18, id_offset = (i - 1L) * n_per_tier
  )
}))
cohort$tier <- factor(cohort$tier, levels = tiers$tier)
stopifnot(!anyDuplicated(cohort$id))

cohort |>
  dplyr::group_by(tier) |>
  dplyr::summarise(
    n = dplyr::n(),
    `median age (y)` = median(AGE),
    `median WT (kg)` = median(WT),
    .groups = "drop"
  ) |>
  knitr::kable(digits = 2, caption = "Virtual CAHD cohort (ages 28 days to <18 years).")
Virtual CAHD cohort (ages 28 days to <18 years).
tier n median age (y) median WT (kg)
3 to <4 kg, 0.2 mg 200 0.11 3.51
4 to <5 kg, 0.3 mg 200 0.12 4.53
5 to <6 kg, 0.5 mg 200 0.17 5.47
6 to <9 kg, 1 mg 200 0.48 7.49
9 to <12 kg, 1.5 mg 200 1.26 10.46
12 to <18 kg, 2 mg 200 3.34 14.93
18 to <25 kg, 3 mg 200 6.05 20.96
25 to <35 kg, 4 mg 200 9.18 29.59
>=35 kg, 5 mg 200 14.13 53.46

Simulation

Twice-daily dosing is simulated directly at steady state (ss = 1, ii = 12) with observations every 0.25 h over one dosing interval.

tau <- 12
obs_grid <- seq(0, tau, by = 0.25)

events <- dplyr::bind_rows(
  cohort |>
    dplyr::mutate(
      time = 0, amt = DOSE_APIXABAN_MG, evid = 1L, cmt = "depot",
      dvid = NA_integer_, ss = 1L, ii = tau
    ),
  cohort |>
    tidyr::crossing(time = obs_grid) |>
    dplyr::mutate(amt = 0, evid = 0L, cmt = "central", dvid = 1L, ss = 0L, ii = 0)
) |>
  dplyr::arrange(id, time, dplyr::desc(evid))

rxode2::rxSetSeed(20250623)
sim <- rxode2::rxSolve(
  mod, events,
  keep = c("tier", "DOSE_APIXABAN_MG"),
  maxsteps = 500000L,
  returnType = "data.frame"
)
#> ℹ parameter labels from comments will be replaced by 'label()'
# A solver failure would drop subjects silently through the PKNCA filter below.
stopifnot(!anyNA(sim$Cc), dplyr::n_distinct(sim$id) == nrow(cohort))

Replicate published figures

Figure 3: steady-state AUC(TAU) by weight tier

sim_nca <- sim |>
  dplyr::filter(!is.na(Cc)) |>
  dplyr::select(id, time, Cc, antiFXa, tier)

dose_df <- events |>
  dplyr::filter(evid == 1L) |>
  dplyr::select(id, time, amt, tier)

intervals <- data.frame(
  start = 0, end = tau,
  auclast = TRUE, cmax = TRUE, cmin = TRUE, tmax = TRUE
)

conc_obj <- PKNCA::PKNCAconc(sim_nca, Cc ~ time | tier + id)
dose_obj <- PKNCA::PKNCAdose(dose_df, amt ~ time | tier + id)
nca_res <- PKNCA::pk.nca(PKNCA::PKNCAdata(conc_obj, dose_obj, intervals = intervals))

nca_wide <- as.data.frame(nca_res) |>
  dplyr::select(id, tier, PPTESTCD, PPORRES) |>
  tidyr::pivot_wider(names_from = PPTESTCD, values_from = PPORRES)

# Steady-state AUC(TAU) equals F1 * dose / CL for the same drawn parameters;
# the only difference is trapezoidal error on the 0.25-h grid.
cl_by_id <- sim |>
  dplyr::distinct(id, .keep_all = TRUE) |>
  dplyr::transmute(id, auc_closed = 1000 * DOSE_APIXABAN_MG * frel * fform / cl)
auc_chk <- dplyr::inner_join(nca_wide, cl_by_id, by = "id") |>
  dplyr::mutate(pct_diff = 100 * (auclast / auc_closed - 1))
stopifnot(
  abs(median(auc_chk$pct_diff)) < 1,
  quantile(abs(auc_chk$pct_diff), 0.9) < 3
)

The Methods describe the Figure 3 simulation as virtual pediatric patients with CAHD. Simulated that way (DIS_HEART_CONGENITAL_ACQUIRED = 1), the tier medians come out about 20% above the published Figure 3 medians in every tier from 6 kg upward, and they agree instead with the Table 2 AUC(TAU) of the real SAXOPHONE patients. A 20% shift in every tier is exactly the CAHD effect on CL/F, exp(0.185) = 1.20; no difference in the age or weight mix of a cohort can produce it in the narrow middle tiers, where maturation is complete and weight is pinned by the tier. The published Figure 3 medians are reproduced when the same virtual patients are simulated with the reference patient type (DIS_HEART_CONGENITAL_ACQUIRED = 0), so the figure appears to have been generated without the CAHD clearance effect. Both scenarios are shown below; the Figure 3 gate uses the reference-type scenario and the Table 2 gate the CAHD scenario.

# Same virtual patients with the reference patient type. Steady-state AUC(TAU)
# is F1 x dose / CL (the paper's AUC equation), so one observation row per
# patient is enough to read the individual parameters.
events_ref <- dplyr::bind_rows(
  cohort |> dplyr::mutate(time = 0, amt = DOSE_APIXABAN_MG, evid = 1L, cmt = "depot", dvid = NA_integer_),
  cohort |> dplyr::mutate(time = 1, amt = 0, evid = 0L, cmt = "central", dvid = 1L)
) |>
  dplyr::mutate(DIS_HEART_CONGENITAL_ACQUIRED = 0) |>
  dplyr::arrange(id, time, dplyr::desc(evid))

rxode2::rxSetSeed(20250625)
sim_ref <- rxode2::rxSolve(mod, events_ref, keep = c("tier", "DOSE_APIXABAN_MG"), returnType = "data.frame") |>
  dplyr::mutate(auc_ss = 1000 * DOSE_APIXABAN_MG * frel * fform / cl)
stopifnot(dplyr::n_distinct(sim_ref$id) == nrow(cohort), !anyNA(sim_ref$auc_ss))

fig3_sim <- dplyr::bind_rows(
  nca_wide |> dplyr::transmute(tier, auc = auclast, scenario = "CAHD (indicator = 1)"),
  sim_ref |> dplyr::transmute(tier, auc = auc_ss, scenario = "Reference patient type (indicator = 0)")
)
# Replicates Figure 3 of Ajavon-Hartmann 2025 (pediatric tiers): steady-state
# AUC(TAU) against the adult VTE-treatment reference (median 1240, 90% CI
# 675-2300 ng.h/mL).
fig3_published <- tibble::tibble(
  tier = factor(tiers$tier, levels = tiers$tier),
  published = c(937, 1090, 1290, 1440, 1390, 1340, 1480, 1520, 1280)
)

q5_95 <- function(y) {
  data.frame(ymin = quantile(y, 0.05), ymax = quantile(y, 0.95))
}
ggplot(fig3_sim, aes(tier, auc, fill = scenario)) +
  annotate("rect", xmin = -Inf, xmax = Inf, ymin = 675, ymax = 2300, alpha = 0.15, fill = "red") +
  geom_hline(yintercept = 1240, linetype = "dashed", colour = "red") +
  geom_boxplot(outlier.shape = NA, coef = 0, position = position_dodge(width = 0.8), width = 0.7) +
  stat_summary(fun.data = q5_95, geom = "errorbar", width = 0.2, position = position_dodge(width = 0.8)) +
  geom_point(
    data = fig3_published, aes(tier, published), inherit.aes = FALSE,
    shape = 4, size = 3, colour = "blue"
  ) +
  scale_fill_manual(values = c("grey85", "white")) +
  coord_cartesian(ylim = c(0, 5000)) +
  labs(
    x = "Weight tier and twice-daily dose",
    y = "Steady-state AUC(TAU) (ng.h/mL)",
    fill = NULL,
    caption = paste(
      "Replicates Figure 3 of Ajavon-Hartmann 2025. Boxes: 25th-75th percentiles;",
      "whiskers: 5th-95th. Blue x: published tier medians."
    )
  ) +
  theme_bw() +
  theme(axis.text.x = element_text(angle = 30, hjust = 1), legend.position = "top")

auc_med <- fig3_sim |>
  dplyr::group_by(tier, scenario) |>
  dplyr::summarise(sim = median(auc), .groups = "drop") |>
  tidyr::pivot_wider(names_from = scenario, values_from = sim) |>
  dplyr::inner_join(fig3_published, by = "tier") |>
  dplyr::mutate(
    pct_cahd = 100 * (`CAHD (indicator = 1)` / published - 1),
    pct_ref = 100 * (`Reference patient type (indicator = 0)` / published - 1)
  )
auc_med |>
  dplyr::rename(
    "Weight tier" = tier, "Figure 3 median" = published,
    "% difference, CAHD" = pct_cahd, "% difference, reference type" = pct_ref
  ) |>
  knitr::kable(digits = 1, caption = "Steady-state AUC(TAU) medians (ng.h/mL) by simulated patient type.")
Steady-state AUC(TAU) medians (ng.h/mL) by simulated patient type.
Weight tier CAHD (indicator = 1) Reference patient type (indicator = 0) Figure 3 median % difference, CAHD % difference, reference type
3 to <4 kg, 0.2 mg 1005.0 842.2 937 7.3 -10.1
4 to <5 kg, 0.3 mg 1264.3 1054.1 1090 16.0 -3.3
5 to <6 kg, 0.5 mg 1528.2 1297.3 1290 18.5 0.6
6 to <9 kg, 1 mg 1745.2 1508.3 1440 21.2 4.7
9 to <12 kg, 1.5 mg 1703.0 1490.8 1390 22.5 7.2
12 to <18 kg, 2 mg 1601.8 1291.6 1340 19.5 -3.6
18 to <25 kg, 3 mg 1790.0 1458.7 1480 20.9 -1.4
25 to <35 kg, 4 mg 1818.7 1568.2 1520 19.7 3.2
>=35 kg, 5 mg 1400.3 1325.8 1280 9.4 3.6

# A wrong clearance, weight exponent, maturation function, F1 or unit shifts
# every tier by tens of percent; the tier-to-tier spread reflects only how
# closely the CDC-based cohort reproduces the paper's age and weight sampling.
stopifnot(
  abs(median(auc_med$pct_ref)) < 10,
  all(abs(auc_med$pct_ref) < 25),
  # The CAHD scenario sits one CAHD effect, about 20%, above the figure.
  abs(median(auc_med$pct_cahd) - 100 * (exp(0.185) - 1)) < 10
)

Comparison against published values (Table 2)

Table 2 reports geometric means of the individual (Bayesian) estimates for the 124 SAXOPHONE patients by weight tier. ncaComparisonTable() compares the simulated medians of the six SAXOPHONE tiers with them.

published <- tibble::tribble(
  ~tier,                  ~auclast, ~cmax, ~cmin, ~tmax,
  "6 to <9 kg, 1 mg",         1460,   185,  57.9,  2.14,
  "9 to <12 kg, 1.5 mg",      1840,   218,  82.7,  2.46,
  "12 to <18 kg, 2 mg",       1610,   222,  64.3,  1.72,
  "18 to <25 kg, 3 mg",       1760,   244,  67.4,  1.65,
  "25 to <35 kg, 4 mg",       1840,   249,  73.1,  1.69,
  ">=35 kg, 5 mg",            1630,   203,  72.7,  1.89
)

cmp <- nlmixr2lib::ncaComparisonTable(
  simulated = nca_res,
  reference = published,
  by = "tier",
  units = c(auclast = "ng.h/mL", cmax = "ng/mL", cmin = "ng/mL", tmax = "h"),
  tolerance_pct = 20
)
knitr::kable(
  cmp,
  caption = paste(
    "Simulated (median, virtual CAHD cohort) vs published (Table 2 geometric",
    "means of SAXOPHONE individual estimates). * differs from reference by >20%."
  )
)
Simulated (median, virtual CAHD cohort) vs published (Table 2 geometric means of SAXOPHONE individual estimates). * differs from reference by >20%.
NCA parameter tier Reference Simulated % diff
Cmax (ng/mL) 6 to <9 kg, 1 mg 185 194 +5.1%
Cmax (ng/mL) 9 to <12 kg, 1.5 mg 218 227 +4.2%
Cmax (ng/mL) 12 to <18 kg, 2 mg 222 212 -4.3%
Cmax (ng/mL) 18 to <25 kg, 3 mg 244 243 -0.6%
Cmax (ng/mL) 25 to <35 kg, 4 mg 249 238 -4.5%
Cmax (ng/mL) >=35 kg, 5 mg 203 178 -12.1%
Cmin (ng/mL) 6 to <9 kg, 1 mg 57.9 86.4 +49.1%*
Cmin (ng/mL) 9 to <12 kg, 1.5 mg 82.7 69.3 -16.2%
Cmin (ng/mL) 12 to <18 kg, 2 mg 64.3 64.9 +1.0%
Cmin (ng/mL) 18 to <25 kg, 3 mg 67.4 77.8 +15.4%
Cmin (ng/mL) 25 to <35 kg, 4 mg 73.1 83.6 +14.3%
Cmin (ng/mL) >=35 kg, 5 mg 72.7 60.7 -16.5%
Tmax (h) 6 to <9 kg, 1 mg 2.14 2.75 +28.5%*
Tmax (h) 9 to <12 kg, 1.5 mg 2.46 1.75 -28.9%*
Tmax (h) 12 to <18 kg, 2 mg 1.72 1.75 +1.7%
Tmax (h) 18 to <25 kg, 3 mg 1.65 1.75 +6.1%
Tmax (h) 25 to <35 kg, 4 mg 1.69 1.75 +3.6%
Tmax (h) >=35 kg, 5 mg 1.89 1.75 -7.4%
AUClast (ng.h/mL) 6 to <9 kg, 1 mg 1460 1750 +19.5%
AUClast (ng.h/mL) 9 to <12 kg, 1.5 mg 1840 1700 -7.4%
AUClast (ng.h/mL) 12 to <18 kg, 2 mg 1610 1600 -0.5%
AUClast (ng.h/mL) 18 to <25 kg, 3 mg 1760 1790 +1.7%
AUClast (ng.h/mL) 25 to <35 kg, 4 mg 1840 1820 -1.2%
AUClast (ng.h/mL) >=35 kg, 5 mg 1630 1400 -14.1%

The two sides differ in kind: Table 2 summarises the empirical Bayes estimates of the actual SAXOPHONE patients, who had few samples each (shrinkage on the etas of ka, Q/F and Vp/F was 36 to 38%), while the simulation draws a full virtual cohort with the population variances. Only six patients sit in the 6 to <9 kg tier and two in the 9 to <12 kg tier, so those reference values are themselves imprecise. Cmax, Cmin and Tmax are therefore descriptive. AUC(TAU) is checked at the centre only: across the six tiers the median difference between the simulated CAHD medians and the Table 2 geometric means must be small.

t2_auc <- nca_wide |>
  dplyr::group_by(tier) |>
  dplyr::summarise(sim = median(auclast), .groups = "drop") |>
  dplyr::inner_join(published |> dplyr::select(tier, auclast), by = "tier") |>
  dplyr::mutate(pct_diff = 100 * (sim / auclast - 1))
stopifnot(nrow(t2_auc) == 6, abs(median(t2_auc$pct_diff)) < 10)

Figure 4: anti-factor Xa activity

The PK/PD model is a zero-intercept line with a normally distributed slope (mean 0.916, SD 0.0528) and an additive residual SD of 12.8 ng/mL. The simulated AXA below is computed from the model-predicted concentration, so the scatter about the line combines the slope variability and the residual error. Figure 4 plots observed AXA against observed apixaban concentration.

ini_df <- rxode2::rxode2(mod)$iniDf
#> ℹ parameter labels from comments will be replaced by 'label()'
stopifnot(
  abs(ini_df$est[ini_df$name == "slope_antiFXa"] - 0.916) < 1e-9,
  abs(sqrt(ini_df$est[ini_df$name == "etaslope_antiFXa"]) - 0.0528) < 1e-6
)

# Add the residual error to the PK/PD output: rxSolve returns predictions
# without residual error for each endpoint.
set.seed(20250624)
axa_obs <- sim_nca |>
  dplyr::slice_sample(n = 3000) |>
  dplyr::mutate(antiFXa_obs = antiFXa + stats::rnorm(dplyr::n(), 0, 12.8))

ggplot(axa_obs, aes(Cc, antiFXa_obs)) +
  geom_point(alpha = 0.15, size = 0.6) +
  geom_abline(slope = 0.916, intercept = 0) +
  labs(
    x = "Apixaban concentration (ng/mL)", y = "Anti-FXa activity (ng/mL)",
    caption = "Model relationship underlying Figure 4 of Ajavon-Hartmann 2025 (median line)."
  ) +
  theme_bw()


# The ratio antiFXa / Cc is each patient's slope, so its distribution must
# reproduce the Table 3 mean and SD.
slopes <- sim |>
  dplyr::filter(Cc > 1) |>
  dplyr::distinct(id, .keep_all = TRUE) |>
  dplyr::mutate(slope = antiFXa / Cc)
stopifnot(
  abs(mean(slopes$slope) - 0.916) < 0.01,
  abs(stats::sd(slopes$slope) - 0.0528) < 0.01
)

Assumptions and deviations

  • Figure 3 patient type. The Methods describe the Figure 3 virtual patients as children with CAHD, but the published tier medians are reproduced only when the CAHD effect on CL/F is switched off, and sit about 20% (exp(0.185)) below the CAHD simulation and below the Table 2 AUC(TAU) of the real SAXOPHONE patients. The model keeps the published CAHD effect; the vignette shows both simulated patient types and gates Figure 3 against the reference-type scenario.
  • Patient-type covariates. The paper’s single patient-type variable (PTYPE) is carried as two binary indicators. DIS_ALL = 1 for the PREVAPIX-ALL patients (acute lymphoblastic leukemia or lymphoblastic lymphoma on asparaginase, pooled as in the paper’s footnote) and DIS_HEART_CONGENITAL_ACQUIRED = 1 for the SAXOPHONE patients; both 0 selects healthy adults and the pediatric phase I patients at risk of thromboembolism. The ALL/LL indicator acts on ka only, the CAHD indicator on CL/F only. The authors suggest the lower CL/F in CAHD reflects the Fontan circulation of most SAXOPHONE patients; the indicator is also confounded with study.
  • PED9M is derived inside the model from AGE (1 for 9 months <= age <18 years), matching the Table 1 footnote definition.
  • Formulation. The sprinkle-capsule relative bioavailability (1.1, fixed) is carried by the general capsule indicator FORM_CAPSULE (1 = sprinkle capsule); the reference is tablet or oral solution. The paper does not say which SAXOPHONE patients took the sprinkle capsule, so all virtual patients here take the reference formulation.
  • Dose-dependent bioavailability reads the dose from the DOSE_APIXABAN_MG column, which must equal the event-table amt on the dose records. The SAXOPHONE doses per weight tier are not tabulated; the doses used are the regimen proposed in the Results, which the Table 2 consistency check above confirms.
  • PK/PD random effect. Table 3 reports a random-effect standard deviation of 0.0528 for the linear mixed-effects regression with ‘IIV in the slope parameter’ (Discussion). It is encoded as a normally distributed, additive random effect on the slope (slope_antiFXa + etaslope_antiFXa), the usual form of a linear mixed-effects random slope. The PK/PD model was fitted to observed apixaban concentrations; here it is driven by the model-predicted concentration Cc.
  • PK/PD dataset size. The Methods give 541 concentrations from 253 patients for the PK/PD dataset; the Results give 390 AXA values from 123 patients after excluding values below the 35 ng/mL quantification limit and values without a time-matched concentration. The model metadata quotes the Results.
  • Virtual cohort. Weight-for-age uses the CDC 2000 sex-averaged median with 12% lognormal scatter instead of sampling the full sex-specific distribution, and ages are drawn within each weight tier. This mainly affects the age and weight mix inside a tier, and so the tier medians, by a few percent. The three lightest tiers (3 to <6 kg) are below the weight of any SAXOPHONE patient and are model extrapolations, as in the paper.
  • Cohort size. 200 virtual patients per tier, against the paper’s about 1000 per tier.
  • Adult reference. The adult VTE-treatment median of 1240 ng.h/mL (90% CI 675-2300) is taken from the paper’s text; adults are not simulated here.
  • No erratum or correction notice for this article was found as of 2026-11-10.