Skip to contents

Model and source

Bukkems 2023 relates the dose, factor VIII (FVIII) activity and bleeding for rFVIII-SingleChain (lonoctocog alfa, Afstyla), a recombinant single-chain FVIII, in persons with severe hemophilia A on prophylaxis. The analysis was sequential: a population PK model was fitted first, and its individual PK parameters then drove FVIII activity in a repeated time-to-event (RTTE) model for bleeds. The two models are packaged as two files that share this vignette:

  • Bukkems_2023_lonoctocogAlfa – the PopPK model: two compartments, body weight on CL and V1, von Willebrand factor antigen (VWF) on CL, an estimated endogenous FVIII level, a decaying residual FVIII level from an unknown pre-study dose, and simultaneous description of chromogenic (CSA) and one-stage (OSA) assay FVIII activity.

  • Bukkems_2023_lonoctocogAlfa_bleed – the bleeding RTTE model: a Weibull hazard declining with time since study start, inhibited by CSA FVIII activity through an Imax function with IC50 8.9 IU/dL, with the PopPK model embedded.

  • Citation: Bukkems LH, Jonsson S, Cnossen MH, Karlsson MO, Mathot RAA. Relationship between factor VIII levels and bleeding for rFVIII-SingleChain in severe hemophilia A: A repeated time-to-event analysis. CPT Pharmacometrics Syst Pharmacol. 2023;12(5):706-718. doi:10.1002/psp4.12938

  • Parameter tables: Table 2 (PopPK) and Table 3 (RTTE). The supplement holds the NONMEM control streams of both final models (Appendices S2 and S3); their unrounded final estimates are the values packaged here.

mod_pk <- rxode2::rxode2(readModelDb("Bukkems_2023_lonoctocogAlfa"))
#> ℹ parameter labels from comments will be replaced by 'label()'
mod_bl <- rxode2::rxode2(readModelDb("Bukkems_2023_lonoctocogAlfa_bleed"))
#> ℹ parameter labels from comments will be replaced by 'label()'

Model equations

PopPK (Table 2 footnote; Appendix S2), with CL and Q in dL/h, V1 and V2 in dL, the dose in IU and FVIII activity in IU/dL:

CL=2.08(WT68)0.702(VWF113)−0.652eηCLV1=34.4(WT68)0.834eηV1CL = 2.08 \left(\frac{WT}{68}\right)^{0.702} \left(\frac{VWF}{113}\right)^{-0.652} e^{\eta_{CL}} \qquad V_1 = 34.4 \left(\frac{WT}{68}\right)^{0.834} e^{\eta_{V1}}

FVIIICSA=BASE+A1V1+EXTRe−βtFVIIIOSA=BASE+0.526A1V1+EXTRe−βtFVIII_{CSA} = BASE + \frac{A_1}{V_1} + EXTR\, e^{-\beta t} \qquad FVIII_{OSA} = BASE + 0.526 \frac{A_1}{V_1} + EXTR\, e^{-\beta t}

where BASEBASE (0.265 IU/dL) is the endogenous FVIII level and the last term, present only for persons whose FVIII before the first study dose exceeded 1 IU/dL, is residual FVIII (1.41 IU/dL) from an unknown pre-study dose decaying at the terminal disposition rate β\beta.

Bleeding hazard (Appendix S3; Equation 1):

hi(t)=λiγ(λit)γ−1(1−FVIIIi(t)FVIIIi(t)+IC50),λi=λeηih_i(t) = \lambda_i \gamma (\lambda_i t)^{\gamma - 1} \left(1 - \frac{FVIII_i(t)}{FVIII_i(t) + IC_{50}}\right), \qquad \lambda_i = \lambda e^{\eta_i}

with λ\lambda = 6.37 per year, γ\gamma = 0.777, IC50IC_{50} = 8.90 IU/dL and FVIIIi(t)FVIII_i(t) the CSA activity including BASEBASE.

Population

The analysis pooled 241 persons with severe hemophilia A (endogenous FVIII < 1 IU/dL) on rFVIII-SingleChain prophylaxis from three trials (Table 1): CSL627_1001 (phase I-III, 12-58 years, n = 74), CSL627_3002 (children < 12 years, n = 81) and the CSL627_3001 extension (all ages, n = 210); the 124 who continued into the extension were analysed as one person each. Median age was 19 years (range 1-58) and median body weight 60 kg (10-120). Prophylaxis was 15-50 IU/kg every second day or two to three times per week. The PK data were 3052 OSA and 2545 CSA FVIII levels; the bleeding data were 2080 bleeds over a median follow-up of 1090 days (65% in joints; 43% spontaneous, 44% trauma related). VWF antigen was missing for 154 persons.

The same information is available programmatically through each model’s population metadata.

Source trace

Quantity Value Source
CL at 68 kg, VWF 113 IU/dL 2.07866 dL/h Appendix S2 THETA(3); Table 2 2.08
V1 at 68 kg 34.3587 dL Appendix S2 THETA(1); Table 2 34.4
Q 1.52841 dL/h Appendix S2 THETA(4); Table 2 1.53
V2 2.40027 dL Appendix S2 THETA(2); Table 2 2.40
Endogenous FVIII (BASE) 0.265462 IU/dL Appendix S2 THETA(7); Table 2 0.27
Residual FVIII from a previous dose (EXTR) 1.40884 IU/dL Appendix S2 THETA(8); Table 2 1.41
Weight exponent on V1 / CL 0.833885 / 0.702087 Appendix S2 THETA(5), THETA(6); Table 2
VWF exponent on CL -0.652459 Appendix S2 THETA(10); Table 2
OSA / CSA ratio 0.52613 Appendix S2 THETA(9); Table 2 0.53
IIV V1, CL, BASE (log-scale variance) 0.0238, 0.0611, 1.319 Appendix S2 $OMEGA; Table 2 15.5, 25.1, 165.6 CV%
CL-BASE covariance 0.0427 Appendix S2 $OMEGA BLOCK(2)
IIV EXTR 3.505 Appendix S2 $OMEGA; see Assumptions
IIV on residual error 0.12698 Appendix S2 $OMEGA; Table 2 36.8 CV%
CSA proportional / replicate proportional / additive (variance) 0.00514 / 0.0178 / 0.546 Appendix S2 $SIGMA EPS(1), EPS(5), EPS(3)
OSA power-term scale (variance) and exponent 0.173 and 0.7266 Appendix S2 $SIGMA EPS(2), THETA(11)
OSA replicate proportional (variance) 0.0456 Appendix S2 $SIGMA EPS(7)
PK structure and error equations – Appendix S2 $PK / $ERROR; Table 2 footnote
Hazard equation Weibull x Imax Appendix S3 $DES; Equation 1
Scale λ\lambda 6.36754 per year Appendix S3 THETA(1); Table 3 6.37
Shape γ\gamma 0.776767 Appendix S3 THETA(3); Table 3 0.78
IC50IC_{50} 8.90353 IU/dL Appendix S3 THETA(2); Table 3 8.90
IIV on λ\lambda 2.47759 Appendix S3 $OMEGA; Table 3 330.8 CV%

PopPK validation

Typical troughs behind the Figure 4 simulations

The paper’s simulations used a 70 kg person with typical PK parameters, for whom 1500, 5500, 9500, 19,500 and 39,500 IU every 72 h “resulted in maintaining FVIII trough levels of 1, 3, 5, 10, and 20 IU/dL” (Results, Simulations). The typical-value steady-state trough (CSA, including the endogenous level) is a deterministic check of CL, V1, Q, V2, the weight centring and BASE. The VWF covariate is set to its 113 IU/dL reference.

trough_tab <- data.frame(
  dose = c(1500, 5500, 9500, 19500, 39500),
  trough_pub = c(1, 3, 5, 10, 20)
)

# Twenty doses (0 to 1368 h); 1440 h is the trough before the next dose.
make_pk_events <- function(dose, id, obs_times, wt = 70, vwf = 113, assay_osa = 0) {
  ev <- rxode2::et(amt = dose, ii = 72, addl = 19, cmt = "central") |>
    rxode2::et(obs_times, cmt = "central") |>
    as.data.frame()
  ev$id <- id
  ev$WT <- wt
  ev$VWF <- vwf
  ev$ASSAY_OSA <- assay_osa
  ev
}

t_ss <- 72 * 20
ev_tr <- bind_rows(lapply(seq_len(nrow(trough_tab)), function(i) {
  make_pk_events(trough_tab$dose[i], i, seq(t_ss - 72, t_ss, by = 0.5))
}))
typ_tr <- rxode2::rxSolve(rxode2::zeroRe(mod_pk), ev_tr, returnType = "data.frame") |>
  suppressWarnings() |>
  suppressMessages()

trough_tab$trough_sim <- typ_tr |>
  dplyr::filter(time == t_ss) |>
  arrange(id) |>
  pull(Cc)

trough_tab |>
  mutate(trough_sim = signif(trough_sim, 3)) |>
  dplyr::rename(
    "Dose every 72 h (IU)" = dose,
    "Stated trough (IU/dL)" = trough_pub,
    "Model typical trough (IU/dL)" = trough_sim
  ) |>
  knitr::kable(caption = "Typical-value steady-state CSA trough, 70 kg, VWF 113 IU/dL.")
Typical-value steady-state CSA trough, 70 kg, VWF 113 IU/dL.
Dose every 72 h (IU) Stated trough (IU/dL) Model typical trough (IU/dL)
1500 1 0.973
5500 3 2.860
9500 5 4.750
19500 10 9.460
39500 20 18.900

stopifnot(
  # Deterministic typical-value solve: a mis-transcribed CL, V1 or weight
  # centring moves the trough by far more than 10%.
  all(abs(trough_tab$trough_sim / trough_tab$trough_pub - 1) < 0.10)
)

Virtual cohort and PKNCA

Two virtual arms of 100 persons each mirror the two trials that contributed most of the rich PK sampling (Table 1): adolescents and adults like CSL627_1001 (median 27 years, 69 kg, VWF 118 IU/dL) and children like CSL627_3002 (median 6 years, 25 kg, VWF 84 IU/dL). Body weight and VWF are drawn log-normally around the trial medians and kept inside the trial ranges. Each person receives a single 50 IU/kg dose with FVIII measured by both assays.

n_arm <- 100
draw_lnorm <- function(n, median, cv, lo, hi) {
  out <- numeric(0)
  while (length(out) < n) {
    x <- median * exp(rnorm(n, 0, cv))
    out <- c(out, x[x >= lo & x <= hi])
  }
  out[seq_len(n)]
}
arms <- data.frame(
  arm = c("Adolescents/adults", "Children < 12 years"),
  wt_med = c(69, 25), wt_lo = c(37, 10), wt_hi = c(112, 88),
  vwf_med = c(118, 84), vwf_lo = c(58, 43), vwf_hi = c(296, 154)
)
obs_pk <- c(0, 0.25, 0.5, 1, 2, 4, 6, 8, 12, 24, 30, 48, 72)

ev_coh <- bind_rows(lapply(seq_len(nrow(arms)), function(a) {
  wt <- draw_lnorm(n_arm, arms$wt_med[a], 0.25, arms$wt_lo[a], arms$wt_hi[a])
  vwf <- draw_lnorm(n_arm, arms$vwf_med[a], 0.3, arms$vwf_lo[a], arms$vwf_hi[a])
  bind_rows(lapply(seq_len(n_arm), function(j) {
    id <- (a - 1) * n_arm + j
    dose <- rxode2::et(amt = 50 * wt[j], cmt = "central") |> as.data.frame()
    obs <- rxode2::et(obs_pk, cmt = "central") |> as.data.frame()
    ev <- bind_rows(
      dose |> mutate(ASSAY_OSA = 0),
      obs |> mutate(ASSAY_OSA = 0),
      obs |> mutate(ASSAY_OSA = 1)
    ) |>
      arrange(time, desc(evid))
    ev$id <- id
    ev$WT <- wt[j]
    ev$VWF <- vwf[j]
    ev$arm <- arms$arm[a]
    ev
  }))
}))

sim_pk <- rxode2::rxSolve(mod_pk, ev_coh, keep = c("arm", "ASSAY_OSA"), returnType = "data.frame") |>
  suppressMessages() |>
  mutate(assay = ifelse(ASSAY_OSA == 1, "OSA", "CSA"))
sim_pk |>
  group_by(arm, assay, time) |>
  summarise(
    p05 = quantile(Cc, 0.05), p50 = median(Cc), p95 = quantile(Cc, 0.95),
    .groups = "drop"
  ) |>
  ggplot(aes(time, p50, colour = assay, fill = assay)) +
  geom_ribbon(aes(ymin = p05, ymax = p95), alpha = 0.2, colour = NA) +
  geom_line() +
  scale_y_log10() +
  facet_wrap(~arm) +
  labs(
    x = "Time after a 50 IU/kg dose (h)", y = "FVIII activity (IU/dL)",
    colour = "Assay", fill = "Assay"
  ) +
  theme_bw()

Median and 90% interval of FVIII activity by assay after a single 50 IU/kg dose. The one-stage assay reads about half the chromogenic activity of the drug, as in Figure S2 of the paper.

PKNCA computes Cmax, AUC0-72 and the 72 h concentration for each arm and assay. The paper reports no NCA, so the check is against the typical-value prediction: the median Cmax should sit near the typical-person value 50⋅WT/V1(WT)⋅r+BASE50 \cdot WT / V_1(WT) \cdot r + BASE, where rr is 1 for CSA and 0.526 for OSA, at the arm’s median weight.

conc_df <- sim_pk |>
  dplyr::filter(!is.na(Cc)) |>
  select(id, arm, assay, time, Cc)
dose_df <- ev_coh |>
  dplyr::filter(evid == 1) |>
  select(id, arm, time, amt) |>
  tidyr::crossing(assay = c("CSA", "OSA"))

o_conc <- PKNCA::PKNCAconc(conc_df, Cc ~ time | arm + assay + id)
o_dose <- PKNCA::PKNCAdose(dose_df, amt ~ time | arm + assay + id)
intervals <- data.frame(start = 0, end = 72, cmax = TRUE, auclast = TRUE, clast.obs = TRUE)
nca <- PKNCA::pk.nca(PKNCA::PKNCAdata(o_conc, o_dose, intervals = intervals))

nca_sum <- as.data.frame(nca$result) |>
  group_by(arm, assay, PPTESTCD) |>
  summarise(median = median(PPORRES), .groups = "drop") |>
  tidyr::pivot_wider(names_from = PPTESTCD, values_from = median)

ini_pk <- setNames(mod_pk$iniDf$est, mod_pk$iniDf$name)
cmax_typ <- function(wt, assay) {
  vc <- exp(ini_pk[["lvc"]]) * (wt / 68)^ini_pk[["e_wt_vc"]]
  r <- ifelse(assay == "OSA", ini_pk[["e_assay_osa_cc"]], 1)
  50 * wt / vc * r + exp(ini_pk[["lrbase"]])
}
nca_sum <- nca_sum |>
  left_join(select(arms, arm, wt_med), by = "arm") |>
  mutate(cmax_typ = cmax_typ(wt_med, assay))

nca_sum |>
  transmute(
    arm, assay,
    cmax = signif(cmax, 3), cmax_typ = signif(cmax_typ, 3),
    auclast = signif(auclast, 3), clast.obs = signif(clast.obs, 3)
  ) |>
  dplyr::rename(
    "Arm" = arm,
    "Assay" = assay,
    "Cmax, PKNCA median (IU/dL)" = cmax,
    "Cmax, typical person (IU/dL)" = cmax_typ,
    "AUC0-72, PKNCA median (IU*h/dL)" = auclast,
    "C72h, PKNCA median (IU/dL)" = clast.obs
  ) |>
  knitr::kable(caption = "PKNCA medians after a single 50 IU/kg dose.")
PKNCA medians after a single 50 IU/kg dose.
Arm Assay Cmax, PKNCA median (IU/dL) Cmax, typical person (IU/dL) AUC0-72, PKNCA median (IU*h/dL) C72h, PKNCA median (IU/dL)
Adolescents/adults CSA 101.0 99.5 1670 2.120
Adolescents/adults OSA 53.4 52.5 885 1.390
Children < 12 years CSA 83.1 84.1 1010 0.782
Children < 12 years OSA 44.1 44.4 554 0.621

stopifnot(
  # Cohort medians of 100 persons per arm: V1 IIV is 15.5% CV and weight
  # varies about the arm median, so 15% is a robust envelope.
  all(abs(nca_sum$cmax / nca_sum$cmax_typ - 1) < 0.15)
)

Residual FVIII from a pre-study dose

For persons whose FVIII before the first study dose exceeded 1 IU/dL, the model adds residual FVIII from an unknown earlier dose, 1.41 IU/dL at the first record (typical value; IIV variance 3.505), decaying at the person’s terminal disposition rate. In the packaged model this is a unit dose into the prior_dose state at the first record, scaled by its bioavailability. Persons without such a pre-dose level receive no prior_dose record.

ev_prior <- bind_rows(
  rxode2::et(amt = 1, cmt = "prior_dose") |> as.data.frame(),
  rxode2::et(amt = 3500, time = 0, cmt = "central") |> as.data.frame(),
  rxode2::et(seq(0, 96, by = 1), cmt = "central") |> as.data.frame()
) |>
  arrange(time, desc(evid)) |>
  mutate(id = 1L, WT = 70, VWF = 113, ASSAY_OSA = 0)
ev_none <- ev_prior |>
  dplyr::filter(is.na(cmt) | cmt != "prior_dose") |>
  mutate(id = 2L)

prior_sim <- rxode2::rxSolve(
  rxode2::zeroRe(mod_pk), bind_rows(ev_prior, ev_none),
  returnType = "data.frame"
) |>
  suppressWarnings() |>
  suppressMessages() |>
  mutate(person = ifelse(id == 1, "Pre-dose FVIII > 1 IU/dL", "No residual pre-study FVIII"))

beta_typ <- prior_sim$beta_term[1]
stopifnot(
  # The residual term starts at EXTR and decays exactly as exp(-beta * t).
  abs(prior_sim$prior_dose[prior_sim$id == 1 & prior_sim$time == 48] -
    exp(ini_pk[["lfprior_dose"]]) * exp(-beta_typ * 48)) < 1e-6,
  all(prior_sim$prior_dose[prior_sim$id == 2] == 0)
)

ggplot(prior_sim, aes(time, Cc, colour = person)) +
  geom_line() +
  scale_y_log10() +
  labs(x = "Time (h)", y = "CSA FVIII activity (IU/dL)", colour = NULL) +
  theme_bw() +
  theme(legend.position = "bottom")

Bleeding RTTE validation

Hazard at constant FVIII (Figure 1)

Figure 1 shows the typical hazard for a constant FVIII level: 3.0 bleeds per year after one year at 1 IU/dL (panel a), and 3.3, 3.0, 1.65 and about 1.0 bleeds per year after one year at 0, 1, 8.9 (the IC50) and 20 IU/dL (panel b and Results). A constant FVIII level is obtained from the packaged model by giving no dose and setting the endogenous level to the target.

const_levels <- c(0, 1, 8.90353, 20)
grid_t <- c(seq(0, 24 * 10, by = 2), seq(24 * 11, 24 * 375, by = 24))
const_sim <- bind_rows(lapply(seq_along(const_levels), function(i) {
  ev <- rxode2::et(grid_t, cmt = "central") |>
    as.data.frame() |>
    mutate(id = i, WT = 70, VWF = 113)
  rxode2::rxSolve(
    rxode2::zeroRe(mod_bl), ev,
    params = c(lrbase = log(max(const_levels[i], 1e-12))),
    returnType = "data.frame"
  ) |>
    suppressWarnings() |>
    suppressMessages() |>
    mutate(fviii = const_levels[i])
}))

fig1 <- const_sim |>
  dplyr::filter(time == 24 * 365) |>
  transmute(fviii, haz_sim = hazard * 24 * 365) |>
  mutate(haz_pub = c(3.3, 3.0, 3.3 / 2, 1.0))

fig1 |>
  mutate(fviii = signif(fviii, 3), haz_sim = signif(haz_sim, 3)) |>
  dplyr::rename(
    "Constant FVIII (IU/dL)" = fviii,
    "Hazard at 1 year, model (per year)" = haz_sim,
    "Figure 1 / Results (per year)" = haz_pub
  ) |>
  knitr::kable(caption = "Typical bleeding hazard after one year at a constant FVIII level.")
Typical bleeding hazard after one year at a constant FVIII level.
Constant FVIII (IU/dL) Hazard at 1 year, model (per year) Figure 1 / Results (per year)
0.0 3.27 3.30
1.0 2.94 3.00
8.9 1.64 1.65
20.0 1.01 1.00

stopifnot(
  # Deterministic: lambda, gamma and IC50 all enter. The published values
  # are rounded to two significant figures.
  all(abs(fig1$haz_sim / fig1$haz_pub - 1) < 0.04)
)
const_sim |>
  dplyr::filter(fviii == 1, time > 0) |>
  ggplot(aes(time / 24, hazard * 24 * 365)) +
  geom_line(colour = "steelblue") +
  coord_cartesian(ylim = c(0, 23)) +
  labs(x = "Time (days)", y = "Bleeding hazard (per year)") +
  theme_bw()

Replicates Figure 1a of Bukkems 2023: typical bleeding hazard at a constant FVIII level of 1 IU/dL (point estimate only; the paper’s band is the parameter uncertainty).

lam_y <- exp(mod_bl$iniDf$est[mod_bl$iniDf$name == "llam_haz"]) * 24 * 365
gam <- exp(mod_bl$iniDf$est[mod_bl$iniDf$name == "lgam_haz"])
ic50 <- exp(mod_bl$iniDf$est[mod_bl$iniDf$name == "lic50"])
data.frame(fviii = seq(0, 50, by = 0.5)) |>
  mutate(hazard = lam_y * gam * lam_y^(gam - 1) * (1 - fviii / (fviii + ic50))) |>
  ggplot(aes(fviii, hazard)) +
  geom_line(colour = "steelblue") +
  geom_point(data = fig1, aes(fviii, haz_sim)) +
  labs(x = "Constant FVIII level (IU/dL)", y = "Bleeding hazard after 1 year (per year)") +
  theme_bw()

Replicates Figure 1b of Bukkems 2023 (the closed-form curve; the points are the packaged model’s solves from the previous chunk).

Annual bleeding rate at maintained trough levels (Figure 4)

Figure 4 simulates 1000 persons of 70 kg with typical PK parameters dosed every 72 h to maintain troughs of 1, 3, 5, 10 and 20 IU/dL, and reports the annual bleeding rate and the percentage with zero bleeds (20%, 31%, 37%, 48% and 59%, printed above the boxes). The Results and Discussion add the median ABR with its 90% prediction interval: 2 (0-17), 1 (0-11) and 1 (0-8) at troughs of 1, 3 and 5 IU/dL.

Successive bleeds are independent, so the number of bleeds in the first year is Poisson with mean equal to the cumulative hazard HiH_i. With typical PK, Hi=H0eγηiH_i = H_0 e^{\gamma \eta_i}, where H0H_0 is the typical cumulative hazard, because the random effect sits on λ\lambda and enters the hazard as λγ\lambda^\gamma. The ABR distribution is then a Poisson mixture over ηi\eta_i, computed here exactly by integration on a fine grid – no random draws are needed.

ev_f4 <- bind_rows(lapply(seq_len(nrow(trough_tab)), function(i) {
  rxode2::et(amt = trough_tab$dose[i], ii = 72, addl = 122, cmt = "central") |>
    rxode2::et(c(0, 24 * 365), cmt = "central") |>
    as.data.frame() |>
    mutate(id = i, WT = 70, VWF = 113)
}))
H0 <- rxode2::rxSolve(rxode2::zeroRe(mod_bl), ev_f4, returnType = "data.frame") |>
  suppressWarnings() |>
  suppressMessages() |>
  dplyr::filter(time == 24 * 365) |>
  arrange(id) |>
  pull(cumhaz)

omega_haz <- mod_bl$omega["etallam_haz", "etallam_haz"]
z <- seq(-8, 8, by = 0.005)
wz <- dnorm(z) * 0.005
abr_mixture <- function(H, mult) {
  k <- 0:400
  pk <- vapply(k, function(kk) sum(dpois(kk, H * mult) * wz), numeric(1))
  cdf <- cumsum(pk)
  c(
    p0 = 100 * pk[1],
    median = k[which(cdf >= 0.5)[1]],
    p05 = k[which(cdf >= 0.05)[1]],
    p95 = k[which(cdf >= 0.95)[1]]
  )
}
# Random effect on lambda inside the Weibull (the control stream).
f4 <- t(vapply(H0, abr_mixture, numeric(4), mult = exp(gam * sqrt(omega_haz) * z)))
# Random effect multiplying the whole hazard (Equation 1 as printed).
f4_eq1 <- t(vapply(H0, abr_mixture, numeric(4), mult = exp(sqrt(omega_haz) * z)))

fig4 <- data.frame(
  trough = trough_tab$trough_pub,
  H0 = H0,
  p0 = f4[, "p0"], p0_eq1 = f4_eq1[, "p0"],
  p0_pub = c(20, 31, 37, 48, 59),
  median = f4[, "median"],
  pi90 = paste0(f4[, "p05"], "-", f4[, "p95"]),
  pub = c("2 (0-17)", "1 (0-11)", "1 (0-8)", "1", "not stated")
)

fig4 |>
  mutate(across(c(H0, p0, p0_eq1), \(x) signif(x, 3))) |>
  dplyr::rename(
    "Trough (IU/dL)" = trough,
    "Typical 1-year cumulative hazard" = H0,
    "Zero bleeds, model (%)" = p0,
    "Zero bleeds, eta on whole hazard (%)" = p0_eq1,
    "Zero bleeds, Figure 4 (%)" = p0_pub,
    "Median ABR, model" = median,
    "90% PI, model" = pi90,
    "Median ABR (90% PI), paper" = pub
  ) |>
  knitr::kable(caption = "First-year bleeds at maintained trough levels, 70 kg, typical PK.")
First-year bleeds at maintained trough levels, 70 kg, typical PK.
Trough (IU/dL) Typical 1-year cumulative hazard Zero bleeds, model (%) Zero bleeds, eta on whole hazard (%) Zero bleeds, Figure 4 (%) Median ABR, model 90% PI, model Median ABR (90% PI), paper
1 2.450 20.1 23.5 20 2 0-19 2 (0-17)
3 1.460 30.3 32.6 31 1 0-12 1 (0-11)
5 1.080 37.0 38.4 37 1 0-9 1 (0-8)
10 0.677 48.1 47.9 48 1 0-6 1
20 0.393 60.6 58.7 59 0 0-4 not stated

stopifnot(
  # Exact integration, no random draws: the percentages with zero bleeds
  # match the Figure 4 annotations (which carry the paper's own 1000-person
  # sampling noise, about +/- 1.5 points).
  all(abs(fig4$p0 - fig4$p0_pub) < 2.5),
  all(fig4$median[1:4] == c(2, 1, 1, 1))
)

The first-year zero-bleed percentages and median ABRs reproduce Figure 4 and the Results. The model’s 90% prediction intervals are slightly wider at their upper end (19, 12 and 9 bleeds against 17, 11 and 8); the 95th percentile of 1000 simulated persons from this heavy-tailed mixture varies by a few bleeds between simulations. Placing the random effect on the whole hazard, as Equation 1 is printed, moves the zero-bleed percentages by up to 3.4 points and fits the low-trough values less well (23.5% against the published 20% at 1 IU/dL); the packaged model follows the control stream (see Assumptions).

abr_dist <- bind_rows(lapply(seq_along(H0), function(i) {
  k <- 0:30
  data.frame(
    trough = factor(trough_tab$trough_pub[i]),
    abr = k,
    prob = vapply(k, function(kk) {
      sum(dpois(kk, H0[i] * exp(gam * sqrt(omega_haz) * z)) * wz)
    }, numeric(1))
  )
}))
ggplot(abr_dist, aes(abr, prob)) +
  geom_col(fill = "steelblue") +
  facet_wrap(~trough, nrow = 1, labeller = label_both) +
  coord_cartesian(xlim = c(0, 15)) +
  labs(x = "Bleeds in the first year", y = "Probability") +
  theme_bw()

Distribution of the first-year bleed count underlying Figure 4 of Bukkems 2023 (shown as probabilities rather than boxplots).

FVIII and hazard for the two illustrative persons (Figure 2)

Figure 2 shows two persons from the dataset: A (38 kg, 13 years, 1000 IU twice weekly) and B (60 kg, 19 years, 2000 IU twice weekly). Their individual parameters are not published; the typical-value profiles below use their weight and a VWF predicted from their age with the paper’s regression (Appendix S2: 79.367 + 1.6672 x age).

persons <- data.frame(
  person = c("A: 38 kg, 1000 IU twice weekly", "B: 60 kg, 2000 IU twice weekly"),
  wt = c(38, 60), age = c(13, 19), dose = c(1000, 2000)
)
fig2 <- bind_rows(lapply(seq_len(nrow(persons)), function(i) {
  ev <- rxode2::et(amt = persons$dose[i], ii = 84, addl = 120, cmt = "central") |>
    rxode2::et(seq(24 * 330, 24 * 358, by = 1), cmt = "central") |>
    as.data.frame() |>
    mutate(id = i, WT = persons$wt[i], VWF = 79.367 + 1.6672 * persons$age[i])
  rxode2::rxSolve(rxode2::zeroRe(mod_bl), ev, returnType = "data.frame") |>
    suppressWarnings() |>
    suppressMessages() |>
    mutate(person = persons$person[i])
}))

fig2 |>
  transmute(
    person, day = time / 24,
    `FVIII activity (IU/dL)` = Cc,
    `Bleeding hazard (per year)` = hazard * 24 * 365
  ) |>
  tidyr::pivot_longer(-c(person, day)) |>
  ggplot(aes(day, value)) +
  geom_line(colour = "steelblue") +
  facet_grid(name ~ person, scales = "free_y") +
  labs(x = "Time since study start (days)", y = NULL) +
  theme_bw()

Typical-value analogue of the top and middle panels of Figure 2 of Bukkems 2023: the hazard rises as FVIII activity falls between doses.

Assumptions and deviations

  • Parameter values. The unrounded final estimates of the two NONMEM control streams (Appendices S2 and S3) are used; they round to every Table 2 and Table 3 value, with one exception: the IIV of the residual previous-dose level, 3.505 in the control stream, converts to 568 CV%, whereas Table 2 prints 596.6 CV%. The control-stream variance is used.
  • Where the hazard random effect enters. Equation 1 multiplies the whole hazard by eηie^{\eta_i}, but the control stream places the random effect on the Weibull scale, λi=λeηi\lambda_i = \lambda e^{\eta_i}, so it enters the hazard as eγηie^{\gamma \eta_i}. The control stream is the model that produced the estimates (the omega of 2.478 belongs to this placement), and it reproduces the Figure 4 zero-bleed percentages slightly better than the printed one.
  • Residual error model. The control stream’s error components are independent normal terms, so their variances add; the packaged model computes the total SD per row and scales it by the individual residual-error random effect. Two features cannot be carried: (1) the CSA and OSA proportional terms are correlated (covariance 0.0298, a correlation of about 1) within a sample measured by both assays, which matters only when simulating paired assays of the same sample; and (2) the separate error terms for repeated measurements of the same sample have equal variances, so they reduce to one term per assay. Table 2 prints the two “multiple measurements” rows with OSA and CSA interchanged relative to the control stream (13.3% belongs to the CSA terms EPS(5)/EPS(6), 21.4% to the OSA terms EPS(7)-EPS(9)); the control stream is followed. Table 2’s ‘Additive error CSA’ of 0.55 IU/dL is the variance (0.546); the SD is 0.739 IU/dL. The OSA additive term was held at zero and is omitted.
  • Weight centring. The Table 2 footnote says typical values refer to a 70 kg person, but its equations and the control stream centre on 68 kg; 68 kg is used.
  • VWF imputation. For the 154 persons without a VWF measurement, the source imputed VWF inside the model from age (79.367 + 1.6672 x age, with a random effect of variance 1931 held constant). This missing-data device is not part of the packaged models; supply VWF for every simulated person. Figure S1 prints the regression as 79.58 + 1.66 x age; the control-stream coefficients are used where needed here.
  • Residual pre-study FVIII. The source switches the decaying residual term on with a data flag for persons whose FVIII before the first study dose exceeded 1 IU/dL. The packaged model encodes it as a unit dose into the prior_dose state at the first record, which reproduces EXTRe−βtEXTR\, e^{-\beta t} exactly; give that record only to such persons. The RTTE model does not include this term.
  • RTTE output and likelihood. The source fits the RTTE likelihood with right censoring at the end of follow-up and interval censoring over the day for bleeds with an unknown time. The packaged RTTE file exposes the hazard, the cumulative hazard and the survivor function sur for simulation, and attaches a placeholder additive error (SD 0.001) to sur that is not from the source. sur is the probability of no bleed since time zero; the source’s survival since the last bleed is exp(-(cumhaz - cumhaz at the last bleed)).
  • Hazard time origin. The Weibull time is time since the person’s first record (study start) and is not reset after a bleed. The source’s offset of 10−610^{-6} inside (λt+10−6)γ−1(\lambda t + 10^{-6})^{\gamma - 1} is kept.
  • Sequential estimation. In the source the RTTE model was driven by each person’s empirical Bayes PK estimates. The packaged RTTE file embeds the PK model with its IIV, so a simulation draws new PK parameters for each virtual person; fix them with params = to reproduce a known person.
  • Covariates on the hazard. Body weight, VWF, region, hemophilic arthropathy and the pre-study annual bleeding rate were screened in a full fixed-effects covariate model; only the pre-study rate had a clinically important effect, and none is in the final RTTE model (Table 3; Appendix S3). Their effect sizes are reported only graphically (Figure 3), so they are documented in covariatesDataExcluded rather than encoded.
  • Virtual cohorts. Weight and VWF are drawn log-normally around the trial medians of Table 1; the PKNCA cohort is not meant to reproduce a published table, since the paper reports no NCA.