Busulfan dosing algorithm and sampling strategy in stem cell transplantation patients

Abstract

Aim

The aim of this investigation was to develop a model-based dosing algorithm for busulfan and identify an optimal sampling scheme for use in routine clinical practice.

Methods

Clinical data from an ongoing study (n = 29) in stem cell transplantation patients were used for the purposes our analysis. A one compartment model was selected as basis for sampling optimization and subsequent evaluation of a suitable dosing algorithm. Internal and external model validation procedures were performed prior to the optimization steps using ED-optimality criteria. Using systemic exposure as parameter of interest, dosing algorithms were considered for individual patients with the scope of minimizing the deviation from target range as determined by AUC(0,6 h).

Results

Busulfan exposure after oral administration was best predicted after the inclusion of adjusted ideal body weight and alanine transferase as covariates on clearance. Population parameter estimates were 3.98 h^–1, 48.8 l and 12.3 l h^–1 for the absorption rate constant, volume of distribution and oral clearance, respectively. Inter-occasion variability was used to describe the differences between test dose and treatment. Based on simulation scenarios, a dosing algorithm was identified, which ensures target exposure values are attained after a test dose. Moreover, our findings show that a sparse sampling scheme with five samples per patient is sufficient to characterize the pharmacokinetics of busulfan in individual patients.

Conclusion

The use of the proposed dosing algorithm in conjunction with a sparse sampling scheme may contribute to considerable improvement in the safety and efficacy profile of patients undergoing treatment for stem cell transplantation.

What is Already Known about this Subject

• Intravenous administration of busulfan has been associated with more predictable pharmacokinetics (PK). However, oral busulfan remains a viable, cost-effective therapeutic option for stem cell transplantation in many countries.

• Dosing recommendations for oral and intravenous busulfan are based on body weight (mg kg^–1) or body surface area (mg m^–2). Given the wide intra- and interindividual variability in PK, improved criteria must be identified for individualization of the dose of busulfan.

What this Study Adds

• This study provides the basis for a new dosing algorithm in conjunction with sparse sampling which may lead to significant improvement in conditioning treatment before haematopoietic stem cell transplantation (HSCT), increasing effectiveness and reducing the chance of busulfan-related adverse events.

Introduction

Busulfan is one of the most frequently used chemotherapy agents in high dose conditioning regimens for patients undergoing haematopoietic stem cell transplantation (HSCT). Whilst the intravenous route has been associated with more predictable pharmacokinetics (PK) and, in some studies, has improved the tolerability of the myeloablative effects of busulfan and cyclophosphamide 1–3, oral busulfan remains a viable, cost-effective option in many countries 4,5.

Current dosing recommendations for oral and intravenous busulfan are based on body weight (mg kg^–1) or body surface area (mg m^–2). Given the wide intra- and interindividual variability in pharmacokinetics and narrow therapeutic window at steady-state (600–900 ng ml^–1), many HSCT centres have attempted to personalize the dose of busulfan using therapeutic drug monitoring (TDM) 6–8. However, the initial dose, the dosing regimen, the timing of pharmacokinetic blood samples and the target busulfan exposure are still defined empirically by the physician 9,10.

Among the currently accepted empirical criteria, pharmacokinetic parameters have been used as proxy for the onset and severity of adverse drug reactions. Busulfan exposure assessed either as area under the concentration–time curve (AUC) or as mean steady-state concentration (C_ss), have been related to sinusoidal obstructive syndrome (SOS) (C_ss > 900 ng ml^–1) or increased incidence of graft failure and relapse of disease (C_ss < 600 ng ml^–1) 11–15. SOS is a specific complication of HSCT that can lead to substantial morbidity and treatment-related mortality. Heparin is frequently used as prophylaxis of and defibrotide as therapy for mild to moderate SOS. However, in severe cases of SOS these therapies are often ineffective, and orthotopic liver transplantation may be the only option.

To mitigate the risk of SOS and other serious adverse events, the conditioning regimen for HSCT is based on an individualized regimen, which includes the administration of a test dose prior to the actual treatment. The goal of the test dose is to ensure the assessment of a patient’s individual clearance estimate. Using non-compartmental methods, the dose is subsequently adjusted to achieve C_ss values between 600–900 ng ml^–1 16–19. Yet, there is evidence that busulfan exposure does not remain within the expected target levels. Such a deviation is partly explained by the fact that this approach does not treat variability in the pharmacokinetics of busulfan as consequence of the effect of covariate factors on drug disposition, but as random noise. Another important practical challenge is the number of blood samples collected for TDM purposes 8,17,18,20–22.

The aim of the present investigation was therefore to develop and evaluate a model-based dosing algorithm for oral busulfan in which factors known to influence the pharmacokinetics of busulfan were taken into account. Furthermore, it is our endeavour to identify a simplified sampling scheme for TDM in stem cell transplantation patients, which would allow a reduction in the overall blood volume required for the characterization of the pharmacokinetic profile in individual patients.

It should be noted that whereas several approaches have been proposed for routine estimation of the individual clearance of busulfan 20,23–31, most methods are still based on linear equations. Such a practice contradicts current evidence on the relationship between covariates and drug disposition. Moreover, no formal validation has been performed for the models describing the population pharmacokinetics of oral busulfan, in which a limited sampling scheme strategy has been used 16,20,23,32,33.

Methods

Study population

Busulfan plasma samples were obtained from 29 patients who received busulfan at the Bone Marrow Transplant Unit of the University Hospital, School of Medicine of Ribeirão Preto, University of São Paulo (HCFMRP-USP). An overview of the demographic data is presented in Table1. The study was approved by the Ethics Committee of HCFMRP-USP. All patients who participated in the trial signed an informed consent form.

Table 1.

Demographic and clinical characteristics of the study population (n = 17 male, 12 female patients)

	Mean	Median
Demographics(n =29)	(SD)	(5th–95th percentiles)
Age(years)	30.3 ± 12.7	30.0 (10.2–50)
Adjusted ideal body weight*(kg)	65.9 ± 16.9	68.0 (46.3–78.9)
Height (m)	1.63 ± 0.16	1.66 (1.4–1.8)
Body weight (kg)	69.7 ± 18.7	64.5 (35.3–93.6)
ALT (µkat l–1)	0.56 ± 0.48	0.43 (0.17–1.54)
Underlying disease (n)
Malignancy	24
BM failure syndrome	1
Sickle-cell disease	4
Concomitant medications (%)
Glucocorticoids	37.9
Anti-emetics	89.6
Antidepressants	6.9
Antimycotics	100.0
Antibacterials	100.0
Antiparasitics	27.6
Phenytoin	100.0
Omeprazole	100.0
Dexchlorpheniramine	24.1
Fludarabine	51.7
Ursodeoxychol acid	27.6
Tramadol	3.4
Furosemide	3.4

Busulfan dose and pharmacokinetic sampling

All patients were treated with a test dose of 0.25mg kg^–1 of oral busulfan (Myleran ®, GlaxoSmithKline) a day before the beginning of the conditioning treatment. Serial blood samples (1 ml/sample) were collected into sodium heparin tubes before and 15, 30, 45, 60, 75, 90, 105, 120, 135, 150, 180, 240, 300 and 360 min after drug administration of busulfan. Blood samples were analyzed using LC-MS/MS. The assay was based on the published method of dos Reis et al. 34. As per study protocol procedures, the AUC was estimated by the trapezoidal method. A first order kinetics model was used for the analysis of the pharmacokinetics of the test dose whereas a non-compartmental method was used for the analysis of the data obtained after the treatment dose (WinNonlin, version 5.2; Pharsight Corporation, Mountain View, CA, USA).

The oral busulfan treatment protocol consisted of 1mg kg^–1 6 h^–1 for 4 days. However, an individualized busulfan dose was recommended for each patient using the apparent clearance calculated from the test dose. This individualized dose was meant to yield plasma concentrations between the target interval of 600–900 ng ml^–1 35. Once steady-state conditions were reached, (i.e. after the fifth dosing event), 15 blood samples were collected according a serial sampling scheme to assess whether busulfan had reached target plasma concentrations.

Population pharmacokinetic modelling

In total, 870 measurements from 29 patients were available for the purpose of this analysis. Considering the limited number of patients available, a meta-analytical approach was considered in which a population pharmacokinetic model previously published by Sandstrom et al. 20 was used to support further evaluation of covariate effects on the disposition of busulfan. In their publication, a one compartmental model with first order absorption and lag time was used to describe the concentration–time profile, with body weight, ALT and phenytoin as covariates on clearance. Given the availability of numerous publications in which a similar parameterization has been described, further model building and evaluation of the base structural model was deemed unnecessary. Focus was therefore given to the re-estimation of parameter distributions and on the identification of new, potential covariates.

In the absence of additional external studies, data splitting and re-sampling techniques were used to ensure proper assessment of the predictive performance of the model 36–40. First, the study population was randomly split into two groups: (1) an index group consisting of 20 patients, who were included into the analysis with the objective of updating parameter estimates and their precision and (2) a validation group consisting of the remaining nine patients. An overview of the modelling procedures is summarized in the diagram in Figure1A. A non-linear mixed effects model was implemented in nonmem ® 7.2.0 (ICON Development Solutions), using a GNU FORTRAN 4.6 compiler (Free Software Foundation, Inc.) and Perl-speaks-nonmem (PsN) as a wrapper for control stream statements and output summaries. Data visualization and manipulation was performed with R 41.

Figure 1.

Overview of the modelling procedures and model validation steps

Covariate effects

Evaluation of the influence of the demographic and clinical factors as potential covariates on pharmacokinetic model parameters included weight, ideal body weight (IBW), adjusted ideal body weight (AIBW), gender, concomitant use of phenytoin and serum aminotransferase (ALT). Covariate selection was based on a forward inclusion-backward deletion procedure. The difference in –2 log likelihood between two hierarchical models is approximately χ² distributed, with degrees of freedom equal to the difference in the number of parameters between them. Factors leading to a decrease in the objective function value (OFV) of ≥ 3.84 were considered significant (P < 0.05) and included as covariates. During the backward deletion process stricter criteria were applied and only the covariates which resulted in a difference of at least 7.88 (P < 0.005) were kept in the final model. Likewise, interindividual variability (IIV) and inter-occasion variability (IOV) were assumed to follow a log-normal distribution and were tested on all parameters, but only included into the model when a significant drop in the OFV was observed.

Model evaluation

The re-estimation of pharmacokinetic parameters was guided by graphical inspection of standard goodness-of-fit, such as population prediction (PRED) vs. observed concentrations (DV), individual prediction (IPRED) vs. DV, distributions of conditional weighted residuals (CWRES) vs. time and CWRES vs. PRED 42. Bootstrapping was subsequently performed to assess the precision of the final parameter estimates 43,44. From the original data set, 1000 bootstrap data sets were generated by re-sampling individuals with replacement. The median of the parameter estimates of all 1000 runs along with the non-parametric 90% confidence intervals were compared with the estimates obtained with the original data set.

Visual predictive checks (VPC)

A visual predictive check was carried out by simulating 500 replicates of the original study design. The median and the 5^th and 95^th percentiles of the simulated data were plotted against time and stratified by body weight. In addition to the graphical analysis, a posterior predictive check (PPC) was also performed using AUC as a metric for the predictive performance of the model. One thousand simulations were performed for each observation from the original data set. AUCs were then calculated and the simulated distributions obtained for test dose and treatment were compared with the actual values observed in the clinical study. The distribution of the predicted AUCs was summarized as medians and compared with the median of AUCs of real patients.

Predictive performance

Given the objective of our analysis, i.e. the development of a dosing algorithm for dose adjustment in prospective patients, additional diagnostic measures were used to assess the predictive performance of the model in subsequent simulations. First, normalized prediction distribution errors (NPDEs) were calculated using the NPDE package in R (version 2.15.1). Graphical summaries, including NPDEs vs. time, QQ plots and a histogram for NPDE distribution indicated no evident bias in model predictions. Finally, mirror plots were generated using the Xpose library available in R to assess issues in the variance-covariance structure, as determined by the degree of similarity, including scattering patterns, between original and simulated data.

Optimized sampling scheme

Optimization of the sampling scheme was considered after completion of the evaluation of model predictive performance. Using PopED (Population Experimental Design, University of Uppsala, Sweden), ED-optimality principles were used to evaluate different sampling scenarios for therapeutic drug monitoring purposes. The steps of this phase of our investigation are outlined in Figure1B. The precision of busulfan clearance was used as the primary optimization criterion, in conjunction with eventual changes in OFV, as observed during minimization procedures. The scenarios that showed lower OFV and reduced coefficient of variation for clearance were further evaluated as described below. Eighty hypothetical patients were added one by one to the original data set. For each patient, new estimates of individual and population pharmacokinetic parameters were obtained taking into account two sampling schemes, a typical one, in which serial samples werere collected throughout the dosing interval and a sparse sampling scheme, as determined by PopED. The difference between individual clearance values estimated for each sampling scheme was then used to evaluate the impact of information loss (i.e. due to the limited number of samples per individual) on model stability and parameter precision, including AUC values.

Simulations

Model-based simulations using the final estimates of the fixed and random effect parameters were carried out in a cohort of virtual patients to describe busulfan exposure after a test dose and subsequently during treatment (steady-state). In contrast to current practice, which relies on a test dose based on body weight, the proposed simulation scenarios included patients receiving a test dose corresponding to one quarter of the dose predicted by the pharmacokinetic model. As indicated in previous paragraphs, the final model includes the demographic covariates AIBW [AIBW= IBW + 25% (actual body weight – IBW), IBW = 25 (height)²] and ALT concentrations. Busulfan concentration vs. time profiles were simulated and AUCs were subsequently calculated using the trapezoidal rule. Based on the individual clearance predicted by the analysis of the concentration data collected after administration of the test dose, an optimized dose for the prospective treatment with busulfan was selected for each hypothetical patient, with the objective of attaining plasma concentrations between 600 and 900 ng ml^–1. The sparse sampling scheme that best described drug exposure, as compared to the estimates obtained with 15 blood samples was selected for evaluation in a prospective clinical trial.

Clinical relevance

To evaluate the clinical implications of this study, four different busulfan dosage and TDM protocols were simulated (using the same 20 patients) and compared. The first simulated scenario was based on the clinical protocol used in HSCT centres that do not use TDM. This implies that the busulfan dose is based solely on the actual body weight of the patient, for those individuals who weighed less than IBW or on AIBW if the actual body weight was greater than IBW. The second scenario includes a test dose, based on actual body weight or IBW or AIBW, which is subsequently used to determine the treatment dose. Finally, two scenarios are proposed using a model-based approach, in which a test dose is derived from the estimates from a one compartment population model and typical and optimized sampling schemes are considered with 15 and five blood samples, respectively.

Results

An overview of the demographic data is presented in Table1.

Pharmacokinetics

A total of 870 plasma busulfan concentrations from an ongoing study were used to re-estimate population pharmacokinetic parameters according to the model published by Sandstrom et al. 20. The analysis was performed taking into account the potential contribution of additional covariate factors, which had not been described previously.

The pharmacokinetics of busulfan could be accurately described by a one compartment model with first order absorption and elimination. However, individual pharmacokinetic profiles were characterized by high inter and intra-individual variability. IIV was identified on clearance, volume of distribution and absorption rate constant. When exploring the impact of covariates, adjusted ideal body weight (AIBW) was found to be non-linearly correlated with clearance and linearly correlated with volume of distribution. On the other hand, the expected non-linear effect of alanine transferase (ALT) on clearance was not identifiable in this population, most probably due to the limited number of patients. A fixed exponent was used instead to ensure this factor was accounted for in the estimation of clearance. The individual clearance of each patient can be described by the following equation: CL/F = 12.3*[(AIBW/69.9)^1.35]*(ALT/0.4)^-0.01. From this equation, it can be seen that the effect of AIBW on clearance is far larger than the effect of ALT alone. In fact, a change of 20% in body weight corresponds to approximately 26% change in the clearance of busulfan. On the other hand, increases in ALT of, for example, as high as 50% relative to median values, have a much smaller effect on clearance.

Whilst data on ALT suggest that changes in liver function may not have clinical relevance for a considerable number of patients, the effect of phenytoin on busulfan clearance was not detected in the current study setting, i.e. there were no statistically significant differences between patients who received phenytoin and those who did not. Based on the aforementioned, we decided not to include this covariate into the final model. The variability in clearance was therefore further investigated for the contribution of random effects. Our data suggested that differences between test dose and treatment can be accurately described by an additional random component, i.e. IOV.

The individual plots presented in Figure2A show that the model fitted the individual data well, after both test dose and treatment. Goodness-of-fit was further demonstrated by the results obtained for the visual predictive checks (VPC, Figure2B). As it can be seen from the median of the simulated distribution, the predictions adequately reflected the trend of the observed data. In addition, the 95% confidence interval appeared to describe accurately the variability of the observed data. On the other hand, a small, but acceptable, over prediction was observed in the estimated variability of the data. This pattern was evident in patients with body weight above the median value.

Figure 2.

(A) Individual plots. The dots represent observed concentrations, the dashed line represents the population prediction (PRED) and the solid line represents the individual predictions (IPRE). (B) Visual predicted check (VPC). The right panel shows the VPC of the population. The dots represent observed concentrations and the dotted lines represent the 5^th and 95^th percentiles of the simulated values. The solid line represents the median of the simulated profiles. Left: panels show the VPC for patients with weight lower (top) and higher (bottom) than 64.5 kg (median) during the test dose. Right: panels show the VPC for patients with weight lower (top) and higher (bottom) than 64.5 kg (median) during treatment. DV , PRED , IPRE

Further evaluation of the predictive performance was performed using diagnostic metrics which provide evidence of model stability and demonstrate the suitability of the stochastic model for simulation purposes. As shown in Table2, confidence intervals were successfully estimated by bootstrapping. In addition, the estimates of the parameters of interest (clearance and volume of distribution) were very similar to the values obtained by the initial fitting of the data. Furthermore, mirror plots reveal that the variance-covariance structure was well characterized, as shown by the similar scattering pattern observed by re-fitting of simulated data sets (see supplemental material - Figure 1Sa). Finally, no particular trends were observed in the distribution of NPDEs for the index dataset (n = 20, see supplemental material, Figure 1Sb).

Table 2.

Summary of pharmacokinetic parameter estimates for busulfan (final model)

		Bootstrap
Parameter	Final model estimates	Median	CV(%)
Fixed effects
CL/F (l h–1)	12.3	12.27	5.02
V/F (l)	48.8	48.99	5.17
Ka (h–1)	3.98	3.93	20.5
tlag (h)	0.20	0.20	5.2
Effect of AIBW on CL	1.35	1.48	0.01
Effect of AIBW on V	2.33	2.4	21.96
Effect of ALT on CL	–0.01(FIX)	–0.01	-
Random effects
IIV CL/F (CV%)	10.0	10.0	57.97
IIV V/F (CV%)	3.16 (FIX)	3.16	-
IIV Ka (CV%)	72.38	69.28	55.3
IOV CL/F (CV%)	7.98	21.6	41.6
IOV V/F (CV%)	31.3	30.0	31.8
IOV Ka (CV%)	84.2	83.6	39.9
Residual error
Proportional error	0.06	0.06	24.5
proportional	14

Taking into account the prospective clinical application(s) of the model, the prediction of systemic exposure, as determined by AUC estimates, was deemed the most important metric of performance. Based on simulations using the same demographic characteristics of the population included in the index data set, comparable parameter distributions were observed for the predicted individual exposure after both the test dose and treatment (Figure3A). Assessment of model performance was complemented by stratification of the VPC plots for treatment occasion and body weight. These plots were subsequently used in conjunction with the reference data set (n = 9), i.e. a subset of randomly selected patients who had not contributed to the initial parameter estimation and model fitting procedures. These data were found to be accurately described by the VPCs (Figure3B).

Figure 3.

(A) Predicted AUC distribution based on model parameter estimates (1000 replicates) obtained from data fitting of 20 patients in both investigated occasions, test dose (left panel) and treatment (right panel). The line represents the true point estimate of AUC in the population. (B) VPC results for the external validation data set (n = 9). The dots represent observed concentrations and the dashed lines represent the 5^th and 95^th percentiles of the simulated values. The solid line represents the median of the simulated profiles. Left: panels show the VPC for patients with weight lower (top) and higher (bottom) than 64.5 kg (median) during the test dose. Right: panels show the VPC for patients with weight lower (top) and higher (bottom) than 64.5 kg (median) during treatment.

Sparse sampling

Taking into account feasibility factors and ED-optimality criteria, five samples were found to be the minimum number required to assess the pharmacokinetics of busulfan, which warrants estimates of AUC(0,∞) and AUC(0,6 h) with acceptable precision. The sampling times associated with the optimized parameter estimation were 0.5, 2.25, 3, 4 and 6 h after drug administration.

To ensure that not only the precision, but also the accuracy of parameter estimates were appropriate for dose adjustment in prospective patients, individual clearance estimates using five samples were compared with those obtained after serial sampling (n = 15). Based on the currently accepted therapeutic window of busulfan and assuming that the clearance estimates obtained by serial sampling were unbiased, a deviation of ± 20% from these values was considered acceptable for estimates obtained by sparse sampling. As shown in Figure4A, all estimates were found to scatter within the proposed range. Finally, our analysis shows that the AUC distributions derived from the rich and the sparse sampling scheme overlap considerably (Figure4B), confirming the possibility of reducing the number from 15 to five blood samples for the purpose of TDM.

Figure 4.

(A) Clearance deviation (%) when estimated with 15 and five samples. Lines represent the acceptable limit (±20%) of individual clearance estimate deviation when estimated with five samples. (B) AUC distribution for both sampling scenarios. The white bars represent AUC estimate with 15 blood samples. Black bars represent AUC estimate with five blood samples. The grey bars are the overlapping of black and white bars

In brief, the dosing algorithm entailed the analysis of busulfan exposure achieved after administration of the test dose based on the patient’s AIBW and liver function as determined by ALT levels. The estimates from individual clearance were then used to define the prospective treatment dose with the objective of attaining plasma concentrations between 600 and 900 ng ml^–1. Our results indicate so far that use of this model-based algorithm for the selection of the treatment dose in prospective patients (Figure5) does allow for improved exposure to busulfan.

Figure 5.

Optimized dosing algorithm target busulfan systemic exposure. Lines represent the therapeutic interval. CP clinical protocol (dose based on AIBW, 1mg kg^–1); TDM busulfan dose based on observed clearance from test dose (non-compartmental analysis); MB (15 samples) test dose based on model and treatment dose based on individual clearance estimated with nonmem using 15 samples; MB (five samples) test dose based on model and treatment dose based on individual clearance estimated with nonmem using only five samples

Discussion

Busulfan TDM has been performed all over the world for more than a decade. Yet, optimized sampling and dose selection remain a research topic, without clear consensus on the requirements to improve treatment outcome and minimize the risks associated with adverse events 7,10,15,32. These issues remain a clinical concern even when intravenous dosing regimens are used.

Our investigation sheds light on some of the factors driving variability in the systemic exposure to busulfan after oral administration. Whilst formulations and bioavailability have been considered an important driver of such variation 10,45,46, less attention has been paid to the contribution of drug–drug interactions associated with the use of the co-medications that are required before and during transplantation procedures. In fact, none of the published population models has fully addressed the issue of variability in pharmacokinetics following oral administration. Instead, preference has been given to intravenous regimens. From a clinical perspective, this represents an important limitation, as oral busulfan represents a cheap therapeutic option in countries such as Brazil 47.

The availability of a model-based algorithm may therefore provide an opportunity for improved therapeutic use of busulfan in stem cell transplantation. As shown by our results, it is possible to ensure optimization of the initial test and treatment doses, without the need for extensive, serial pharmacokinetic sampling. Given the clinical implications of dose optimization, the burden relative to the proposed experimental procedures for patients and physicians can be considered minor.

As previously reported by Sandstrom et al. 20, the pharmacokinetics of busulfan administered orally could be described by a one compartment model with first order absorption and elimination. Population estimates for clearance (12.3 l h^–1) and volume of distribution (48.8 l) were in accordance with those found in the literature 20,48. Our findings showed that actual estimates of AUC were well within the 95% CI of the predicted exposure after a test dose. Whereas model evaluation has been limited to the confirmation of the effect size of known covariates and further refinement of population parameter estimation, our analysis has also identified the contribution of inter-occasion variability. In contrast to previous models 20,29, which include body surface area and body weight as covariates, ideal body weight was identified as a better descriptor of the effect of the individual differences in clearance and volume of distribution. Even though evidence about the magnitude of the effect of ALT on clearance has been inconsistent across numerous publications 29,32,33, we decided to fix the parameter-covariate correlation to ensure its effect was captured by model predictions. In this respect, we need to point out that the observed effect of ALT was marginal, as compared with the 12% stated in a previous publication 20. Similarly, it was not possible to establish any clear correlation between co-medication and busulfan pharmacokinetics. Published data on the effects of phenytoin on busulfan metabolism show variable results. In fact, the mechanism underlying metabolic induction is questionable since phenytoin is not an inducer of glutathione-S-transferase (GST), the major enzyme involved in busulfan metabolism 49,50. Likewise, no correlation was observed between busulfan pharmacokinetics and GST polymorphism.

A possible explanation for this finding may be linked to the procedures for seizure prophylaxis treatment. Dosing with phenytoin or other anti-epileptic agents generally starts on the same day or 1 day before the beginning of the conditioning regimen with busulfan, which is then administered over 4 days. Considering that the patients included in this study were evaluated on the second treatment day, we believe that the absence of any significant effect may simply be due to the timing of busulfan administration. In other words, the interval between phenytoin and busulfan administration may have been too short to allow metabolic induction.

Model evaluation

Model evaluation and assessment of its predictive performance are extremely important steps for the use of a model-based approach in clinical pharmacology and therapeutics. Yet, formal evaluation procedures are rarely performed 51,52. In contrast to previous studies 18,20,21,27,29,51, we have demonstrated model performance for prospective use in patients undergoing stem cell transplantation. The observed predictive performance provides the basis for a dosing algorithm and reduced blood sampling scheme. Noteworthy is that the proposed sampling scheme warrants the characterization of the erratic absorption process. Different sampling times have been suggested in previous investigations 18,20,21,53, but information regarding model evaluation and predictive performance was not provided in the publications.

Ideally, all the patients undergoing treatment with busulfan should achieve plasma concentrations within the target range, but this goal is currently not met. More efficient methods of achieving target busulfan exposure are therefore desirable. In this context, non-compartmental methods in TDM cannot be considered as acceptable practice any longer. Improvement of TDM requires the use of model-based techniques, which are statistically more robust and enable the distinction between covariate factors and residual or random variability in pharmacokinetics. Due to high variability, the number of patients analyzed and of available demographic covariates, it was not possible to disentangle the contribution of different factors. The proposed busulfan dose based on administration of a test dose (a quarter of busulfan dose calculated by the population model) followed by five blood samples for determination of the apparent clearance (for defining the optimal dose to achieve the therapeutic range of busulfan) was found to perform better than the currently used protocols (Figure5). Based on simulation scenarios, it is possible to demonstrate that a considerable fraction of patients were exposed to AUC(0,6 h) values below or above the therapeutic window (3600–5400 ng ml^–1 h) when the busulfan dose was administered based only on AIBW. The use of TDM in conjunction with a test dose (0.25 mg kg^–1) appeared to result in some improvement, but many patients remained outside the target range. By contrast, the use of a model-based approach to define both the test and treatment results in the most precise distribution around the target range. Furthermore, we do not anticipate that the reduction in the number of samples will have any effect on graft failure. The reduction in the number of samples proposed here may only have implications for the parameter precision, not its accuracy.

Limitations

One of the potential limitations in our investigation was the relatively small number of patients (n = 29) available for the analysis. We believe however that use of a structural model from the published literature represents an effective and unbiased approach for subsequent characterization of individual pharmacokinetic parameters as well as further evaluation of covariate effects in this small group. The range of covariate values observed for ALT and differences related to co-medications were clearly limited in this population. Hence, further evaluation of these putative covariates in a wider population is desirable. It is also clear that lack of intravenous data in the same patients prevents a more comprehensive evaluation of the true changes in metabolic clearance, which would enable us to discriminate the role of pharmaceutical factors (e.g. formulation) from other intrinsic and extrinsic causes of variability, such as drug–drug interaction and altered liver function.

Nevertheless, we do not anticipate that these limitations would alter the overall results and conclusions regarding the clinical implications of a model-based dosing algorithm, as compared with current practice. It should be noted that model performance was robust and unbiased, as demonstrated by the different validation procedures presented here.

In summary, the results from our study contribute to the introduction of a model-based dosing algorithm in conjunction with sparse sampling which may lead to significant improvement in conditioning treatment before HSCT, increasing effectiveness and reducing the chance of busulfan-related adverse events.

Competing Interests

All authors have completed the Unified Competing Interest form at www.icmje.org/coi_disclosure.pdf (available on request from the corresponding author) and declare FAC had support from FAPESP (Fundação de Amparo à Pesquisa do Estado de São Paulo, Brazil) for the submitted work. There are no financial relationships with any organizations that might have an interest in the submitted work in the previous 3 years and no other relationships or activities that could appear to have influenced the submitted work.

The authors thank the São Paulo Research Foundation (FAPESP) for financial support for the completion of this project.

Supporting Information

Figure S1 (a) Individual mirror plots. Mirror plots describe the goodness-of-fit and scattering pattern between observed busulfan plasma concentrations and individual predicted busulfan plasma concentrations based on simulated data using the final population pharmacokinetic model. Left plot shows the original clinical data and the corresponding model predictions. The 3 plots on the right were selected randomly from the simulated data. (b) Normalized Prediction Distribution Errors (NPDE). Left upper plot shows quantile-quantile versus expected normal standard distribution (line). Right upper panel shows the NPDE histogram along with the probability density of the standard normal distribution. Left lower panel shows NPDE versus time (X) after the test dose and during treatment. The right lower panel shows NPDE versus predicted concentrations.