1 Introduction
Hemophilia A and B are caused by a deficiency of either clotting factor VIII (FVIII) or IX (FIX), respectively [
1,
2]. Disease severity is categorized by the patient’s endogenous baseline factor activity level. Patients with severe, moderate, and mild hemophilia have a baseline factor level of < 0.01 IU mL
−1, between 0.01 and 0.05 IU mL
−1, and between 0.05 and 0.40 IU mL
−1, respectively [
3]. These definitions of severity are similar for both hemophilia A and B. As patients with severe hemophilia experience spontaneous and more frequent bleeding with development of joint arthropathy and long-term invalidity if left untreated, patients with severe hemophilia and some patients with moderate hemophilia with a severe bleeding phenotype administer factor concentrates prophylactically [
4].
In general, prophylactic dosing of factor concentrates in patients with severe hemophilia is targeted at a trough level of 0.01 IU mL
−1 [
5]. In contrast, factor levels are augmented to physiological levels to maintain optimal hemostasis during a surgical procedure [
6]. Following the surgical procedure, higher target trough levels than during the non-surgical prophylactic setting are maintained up to 2 weeks depending upon the type and severity of the surgery, according to current treatment guidelines.
An alternative to factor concentrates in patients with non-severe hemophilia A is desmopressin (
d-amino
d-arginine vasopressin or DDAVP) [
7]. Desmopressin releases von Willebrand Factor (VWF) from Weibel Palade bodies in endothelial cells of the vessel wall. As VWF functions as a FVIII carrier protein, protecting it from proteolysis in the circulation, FVIII activity will also rise upon administration of desmopressin. The reason desmopressin can only be applied in patients with non-severe hemophilia A is that synthesis of endogenous FVIII is required [
8].
For FVIII and FIX, various plasma-derived (pd) and recombinant (r) concentrate products are available with a standard terminal half-life (SHL). Recently, FVIII and FIX protein molecules have been designed with an extended terminal half-life (EHL) [
9]. For both FVIII and FIX, factor molecules have been linked to the Fc domain of immunoglobulin G (rFVIIIFc and rFIXFc) [
10,
11]. Other methods to extend the terminal half-life were linking FIX with albumin (rIX-FP) and by PEGylation (N9-GP) [
12,
13]. In comparison to SHL products, EHL products of FVIII and FIX exhibit hemostatic activity for longer time periods.
Although factor concentrates are dosed according to a patient’s body weight (BW), it has been demonstrated that FVIII and FIX concentrates show considerable pharmacokinetic (PK) variability [
14,
15]. Population PK analyses are increasingly performed in hemophilia research, as constructed population PK models are able to quantify and explain the variability of PK parameters. These studies have provided population PK models that allowed the characterization and comparison of the pharmacokinetics of FVIII and FIX concentrates, and models that can be applied to perform dose individualization and evaluations of limited sampling schedules. Moreover, population PK analyses can assist in the identification of patient characteristics that describe and predict pharmacokinetics [
16]. A validated population PK model can be used to perform a maximum a posteriori (MAP) Bayesian estimation to obtain individual PK parameter estimates [
17]. The latter estimates are useful to describe the factor level vs time curve for any given dose and help to design individualized dosing regimens. In addition, population PK models allow optimization of clotting factor dosing by an in silico evaluation. In the latter, Monte Carlo simulations are applied that allow the exploration of the resulting factor levels as the dosing regimen or patient characteristics varied [
18‐
20].
In this review, the population PK analyses that have been conducted to develop population PK models describing factor levels after administration of FVIII or FIX concentrates or desmopressin are presented and discussed. Moreover, methods used to construct these models, key model features, patient characteristics of studied populations, and established covariate relationships are discussed in detail.
2 Methods
2.1 Search Strategy
To identify the available literature on population PK analyses of FVIII, FIX, and desmopressin in patients with hemophilia A and B, the following PubMed search query was applied: (haemophilia* OR hemophilia*) AND (“VIII”[Tiab] OR “IX”[Tiab] OR “desmopressin”[Tiab] OR “DDAVP”[Tiab]) AND (“population pharmacokinetics”[Tiab] OR “pharmacokinetic model*”[Tiab] OR “Two-Stage*”[Tiab] OR “population pharmacokinetic analysis”[Tiab] OR “population PK”[Tiab]) NOT (“monkey*”[Tiab] OR “mice*”[Tiab] OR “dog*”[Tiab] OR “rabbit*”[Tiab] OR “rat*”[Tiab]) AND (“1960/01/01”[Date—Publication]: “2020/04/30”[Date—Publication]). The last date of publication inclusion was 30 April, 2020.
The retrieved publications were evaluated for eligibility on the basis of title and abstract. From the selected publications, backward citation screening was conducted to identify additional studies from the reference listings. Publications were only selected if they presented a population PK model derived from real-world patient data (i.e., data that are not obtained using Monte Carlo simulations), and were established using either a standard two-stage (STS) analysis or by non-linear mixed-effect modeling (NLMEM). Publications describing non-compartmental PK analyses or population PK models derived solely using simulation techniques, i.e., without real-world patient data, were not included.
2.2 Data Collection
From selected publications, the following data were retrieved: demographics of study population, type of product, laboratory assays used to obtain FVIII or FIX activity levels, modeling software applied, sampling design, other relevant study characteristics related to treatment using FVIII or FIX concentrates or desmopressin, and all presented model parameters.
2.3 Methodology of Population Pharmacokinetic Model Construction
In general, multiple methods can be applied to construct population PK models [
21]. However, two methods have specifically been applied in hemophilia treatment. In hemophilia, the earliest population PK models were established using a STS analysis, which is the “traditional” approach [
22]. However, currently the most widely used method is NLMEM, which is a one-stage method. The STS and the NLMEM approach both have their advantages and limitations. The complexities and constraints of both methods will, therefore, first be elaborated upon below.
2.3.1 Standard Two-Stage Approach
The application of the STS approach is less time consuming and less complex than the NLMEM approach. The first stage of the STS approach consists of obtaining individual PK parameter estimates for each individual in the studied population using a compartmental model [
23,
24]. In the second stage, distributions of these individual PK parameter estimates are described with summary statistics, i.e., the mean and standard deviation of clearance (CL) or volume of distribution (Vd). The values of the means from the parameter distributions represent the population PK parameters, whereas standard deviations represent the inter-individual variability (IIV) for the corresponding parameters.
In this method, it is assumed that each individual from the studied population contributes equally to the estimation of the population PK parameters, although the amount of data may differ between patients. Moreover, the IIV of the population PK parameters is generally overestimated, as the residual error (difference between the predicted and measured factor levels) is included in the IIV for estimating the individual PK parameters [
25].
Another drawback of this method is the need for rich data, i.e., ten or more factor levels are required for each individual to adequately estimate individual PK parameters. However, sparse sampling strategies have also been investigated for this method [
26]. Essentially, this renders the STS method less suitable for analyzing sparse data, which is often the case in clinical studies.
2.3.2 Non-Linear Mixed-Effect Modeling
Non-linear mixed-effect modeling is a one-stage approach, referred to as the “population” approach [
27]. In this method, some or all fixed and random effects enter the model non-linearly, hence the term “non-linear”. The term ‘mixed-effect’ refers to combining fixed and random effects in a single model [
28]. Fixed effects comprise typical (median) parameters and parameters describing covariate effects, whereas random effects refer to the parameters describing the IIV or the residual unknown variability. In addition, random effects can also be estimated within an individual (intra-individual variability). For instance, an individual PK parameter may change between occasions using inter-occasion variability (IOV) [
29]. An occasion can be defined, for example, as a dosing event or a surgical procedure.
After estimating the IIV and IOV associated with the model parameters, these may be explained using covariate relationships such as the patients’ BW or age. In the covariate analysis, various covariate models may be evaluated for their ability to explain parts of the IIV or IOV. Moreover, the covariate relationships may also be applied to explain a part of the residual unknown variability [
28].
In population PK modeling, population parameters are often scaled using the BW of the patients, especially in populations containing both children and adults [
30]. By scaling or normalization, a part of the IIV is explained. Although several methods are available to scale model parameters [
31], the following equation is generally used for allometric scaling using BW:
$$\theta_{{{\text{TV}}}} = \theta_{{{\text{Pop}}}} \times \left( {\frac{{{\text{BW}}_{i} }}{{{\text{BW}}_{{{\text{med}}}} }}} \right)^{{{\text{EXP}}_{{{\text{BW}}}} }} ,$$
(1)
in which
θTV is the estimated typical value of the population PK parameter scaled to a patients’ BW (BW
i),
θPop is the estimated population PK parameter value, BW
med is the median of the BW from the studied population (or a value used to scale the population PK parameters such as 70 kg), and EXP
BW is the value for the estimated allometric exponent. When fixed allometric exponents are used, a value for EXP
BW of 0.75 (or ¾) is used for all clearance parameters, whereas a value of 1 is used for all volume of distribution parameters [
32]. These fixed values for the allometric exponents have been derived from biological principles and observations from diverse areas in biology [
33]. The allometric exponents may also be estimated empirically, on the basis of the collected data. In that case, the value of exponents may differ from 0.75 and 1 and may differ between the PK parameters.
The NLMEM approach is the most frequently applied method for analyzing vast amounts of population data. However, the application of this method is more complex and time consuming than the STS approach and requires considerable expertise. This method allows simultaneous estimation of both fixed and random effects. Moreover, this method is suitable for analyzing both rich and sparse data, which may be heterogeneously collected between the individuals from the studied population [
24,
34]. Similar to the STS method, sparse data impose constraints on the model’s complexity. Therefore, clinical trial design is important to optimally collect data before the modeling process is initiated [
35].
4 Discussion
In this review, all available population PK models describing FVIII and FIX activity or desmopressin after dose administration were summarized. In total, 33 population PK models were retrieved from the literature. In 29 population PK analyses, the NLMEM approach was applied, whereas in four population PK analyses, the STS approach was performed. The latter four analyses described administration of SHL-FVIII only. Moreover, with the advent of NONMEM, this is currently the preferred software to established population PK models. Only one population PK model was developed to describe FVIII levels after administration of desmopressin. From all available population PK analyses, 27 studies also included data from pediatric patients. In total, 18 population PK models were established on the basis of data derived from a single product, whereas the remaining models were established using data from multiple products.
In a minority of the population PK models, IOV of the population PK parameters CL or V1 was described. In three of the four population models describing the pharmacokinetics of SHL-FIX concentrates (Table S7 of the ESM), IOV ranged considerably from 15 to 48.8% and 12–47.2% for CL and V1, respectively. As a consequence of IOV, CL and V1 differ with every dose administration. In a MAP Bayesian analysis, individual PK parameter estimates can still be obtained for population PK models having IOV associated with a population PK parameter [
70]. Using the individual PK parameters, individualized doses of FVIII or FIX concentrates can be calculated. Dose calculations based on individual PK parameters obtained with a MAP Bayesian analysis will compensate for the IIV only. If a MAP Bayesian analysis is iteratively applied with the most recent measured factor level, the resulting individual PK parameter estimates will gradually represent the average for the parameter value of an individual. Therefore, despite administration of individualized doses obtained using (iterative) a MAP Bayesian analysis, the resulting factor levels may vary still depending on the extend of IOV. Therefore, one should be cautious when the IOV considerably exceeds the extent of IIV associated with one or more PK parameters from a population PK model used to perform dose individualization.
Several studies have indicated that CL of FVIII immediately after a surgical procedure may be increased because of consumption of FVIII [
71‐
73]. To address this change in FVIII CL during and after surgery, population PK models were constructed allowing the CL to change after the surgical procedure (Table
1) [
37,
38]. In the study from Longo et al., the time-dependent elimination model allowed a better description of the measured FVIII levels than the constant elimination model in ten out of 20 patients [
37]. These population PK models were constructed using the STS approach only. Although Hazendonk et al. evaluated different relationships allowing the prediction of a time-dependent CL in their perioperative population PK model constructed using NLMEM, no time dependence of CL was observed [
57]. In a recent study, Preijers et al. did not observe changes over time in the PK parameters during perioperative dosing of FIX concentrates (Table S9 of the ESM) [
69]. Whether population PK models constructed using NLMEM allow the description of time-dependent PK parameters in the surgical setting remains to be further studied in prospective studies having well-timed samples before, during, and after surgery [
74].
In the majority of the population PK models, the population PK parameters were allometrically scaled using the actual BW. Only in two cases was LBW used for scaling (Table S3 of the ESM). The allometric exponents can be fixed to a value that is set before the relationship is evaluated (“3/4 rule”) or can be estimated [
32,
75]. However, if allometric exponents are estimated, the allometric scaling is not based on biologic principles but rather based on statistical heuristics driven by an empirical relationship present in the collected data used to construct the model. As a result, obtaining values for the estimated exponents may be influenced by confounding factors. When the allometric exponents are not accurately estimated, potential issues may arise when the model is used for extrapolation to other populations, such as from adults to children [
76]. Moreover, allometric exponents may not be accurate for obese patients if they were not present in the data used to construct the model, as FVIII and FIX are not likely to distribute to the fatty tissue. Nevertheless, the population PK models may be applied to describe the FVIII or FIX levels after administration of a factor concentrate for their corresponding study populations. When applying these models in Therapeutic Drug Monitoring, they should have been validated using independent datasets.
In the majority of the population PK models summarized in this review, the endogenous baseline factor level of the patients with hemophilia from the studied populations was taken into account (Tables
1 and
2). Different methods were applied to correct the predicted factor levels using the endogenous baseline level. Moreover, if the patients used prophylactic doses previous to the loading dose from which, subsequently, the FVIII or FIX levels were measured, a residual pre-dose factor level may be present next to the endogenous baseline level. A correction for the endogenous baseline level may also be performed prior to the initiation of the modeling process. It is important to correct the predicted factor levels using the residual or endogenous baseline levels, as doing not so will affect the estimation of the model parameters [
77]. Not correcting for a residual pre-dose or baseline factor level may result in underestimation of Vd and/or overprediction of CL [
78]. Therefore, the endogenous baseline factor level and the residual pre-dose factor level should be taken into account when a population PK model is constructed at least for patients with non-severe hemophilia.
When performing patient-tailored dosing using a MAP Bayesian analysis, the availability of a population PK model is a prerequisite. However, if a model contains a covariate relationship, a covariate value must be supplied to calculate the typical value for the corresponding population PK parameter. In a MAP Bayesian analysis, empirical Bayesian estimates are obtained for the model parameters containing IIV or IOV. Using the typical values and empirical Bayesian estimates, the individual PK parameters can be calculated. In the majority of the population PK models summarized in this review, the model parameters were allometrically scaled. In one population PK analysis, the levels of VWF and hematocrit were also associated with covariate relationships to population PK parameters (Table S4 of the ESM). In that case, these VWF and hematocrit levels have to be measured or, if measurements are unavailable, imputed before the individual PK parameters can be obtained accurately. The requirement of extra laboratory measurements poses a limitation for obtaining patient-tailored doses using a MAP Bayesian analysis. Nevertheless, taking covariates into account could improve the selection of the right dose for the right patient and, therefore, all covariate relationships that are likely to have a predictive ability should be evaluated in the construction of a population PK model.
For the majority of the established population PK models, parts of the IIV were explained using covariate relationships. Still, there is considerable unexplained IIV present in most of the population PK models with ranges of 17–50% and 8–54.2% for CL and V1, respectively. Further research may explain the extent of explained IIV. Only a minority of the established population PK models has been applied successfully to perform individualized dosing of factor concentrates [
79]. Although the established population PK models are validated internally using the same data set used to construct the model, only a limited number of external validations have been conducted comprising an independent data set [
80]. As the latter is considered to be the most stringent approach for model evaluation, such studies must first be conducted before the population models can be applied standardly into clinical practice. Therefore, to allow the application of the established population PK models in clinical practice to obtain individualized doses, further studies have to be performed to validate these models. Currently, various ongoing studies are investigating individualized dosing using population PK models in patients with non-severe hemophilia A (DAVID study) [
81], surgical patients with hemophilia A (OPTI-CLOT) [
74], and in patients using extended terminal half-life FVIII and FIX products (Target Study).