Abstract
Background and Objectives
Busulfan, used in conditioning regimens prior to haematopoietic stem cell transplantation, has a narrow therapeutic index; therefore, due to pharmacokinetic variability, therapeutic drug monitoring (TDM) is employed to ensure drug exposure levels associated with optimal clinical outcomes. This study aimed to facilitate dose adjustment and potentially reduce the number of samples required for TDM in clinical practice.
Methods
Population pharmacokinetic (popPK) and machine learning (ML) models (Random Forest and XGBoost) were developed using a local, retrospective, extensive-sampling dataset of 67 children (637 observations, Denmark) and externally evaluated using prediction error (PE) on a TDM dataset from another centre with 171 children (1871 observations, Netherlands). Limited sampling schedules were evaluated for both methods.
Results
A one-compartment model was the best data fit, showing time-varying clearance decreases of 10.4% and 13% on Days 2 and 3, respectively, compared to Day 1. The popPK, Random Forest, and XGBoost models performed well on the external dataset for concentration prediction, with median PE ranging from −4.82 to −7.44%, median absolute PE of approximately 15%, and approximately 80% of data predicted within ±30% of the PE. Sampling strategies at 0, 0.5, 1, 4 h or 0, 1, 2, 4 h post-infusion adequately predicted busulfan plasma concentration and exposure using the popPK model.
Conclusions
A new popPK model was developed to support individualised dose adjustments in children using fewer blood samples. Both Random Forest and XGBoost could predict busulfan's plasma concentration with performance like that of the popPK model (population prediction). However, a larger dataset is needed for ML methods to describe a complete pharmacokinetic profile.
Supplementary Information
The online version contains supplementary material available at https://doi.org/10.1007/s40262-026-01702-9.
Key Points
| A novel population pharmacokinetic (popPK) model for busulfan in our cohort of paediatric patients was developed and externally evaluated before being implemented into a Shiny app for clinical use. |
| Limited sampling strategies, using samples taken at 0, 0.5, 1, 4 h or 0, 1, 2, 4 h post-infusion, effectively predicted the exposure of busulfan employing the popPK model in our cohort of paediatric patients. |
| Random Forest and XGBoost achieved similar predictions for busulfan's plasma concentration as a traditional popPK model but required data sampling for the entire timespan for accurate pharmacokinetic profiling. |
Introduction
Busulfan, an antineoplastic alkylating agent, is used in conditioning regimens before haematopoietic stem cell transplantation (HSCT) to treat both malignant and non-malignant diseases [1]. It creates DNA interstrand crosslinks that inhibit DNA replication, ultimately triggering apoptosis. Busulfan has a narrow therapeutic index, and studies have shown that the cumulative area under the concentration-time curve (AUC) over four days of dosing is related to both efficacy and side effects [2–4]. Several studies show that a cumulative AUC of 75–100 mg·h∙L−1 optimises engraftment rates with acceptable busulfan toxicity [5].
To ensure optimal exposure of busulfan, therapeutic drug monitoring (TDM) is recommended. However, dose adjustments are still challenging as the causes of interindividual variability need to be considered. Factors such as total body weight (WT), body size by use of normal fat mass, body surface area, postmenstrual age for children younger than 2.5 years, and hepatic function can influence busulfan clearance (CL), potentially impacting optimal clinical results [3, 6–8]. Busulfan is mainly metabolised by glutathione S-transferase (GST) in the liver. When busulfan exposure is too high, it increases the risk of toxicity, such as veno-occlusive disease (VOD) and transplant-related death [9, 10]. Conversely, when exposure is too low, it leads to a high risk of graft rejection and disease relapse in children and young adults [6]. Therapeutic drug monitoring of busulfan in a busulfan/cyclophosphamide regimen has shown beneficial effects such as increasing engraftment rates in children, lowering hepatotoxicity in adults, and lowering relapse rates in patients with previously untreated chronic myeloid leukaemia [8]. However, the association between exposure and relapse has been less clear in children with acute myeloid leukaemia, probably due to low sample sizes [8].
Intravenous (IV) busulfan administered every 6 h or every 24 h over four consecutive days is currently the standard of care. The European Medicines Agency (EMA) suggests individual dosing per day for children based on total body weight (WT) (4 mg/kg for WT ≤ 9 kg, 4.8 mg/kg for WT 9–16 kg, 4.4 mg/kg for WT 16–23 kg, 3.8 mg/kg for WT 23–34 kg, and 3.2 mg/kg for WT ≥ 34 kg) [11]. After the first dose, TDM is commonly implemented in many transplant centres to optimise the drug exposure for each patient.
Several studies show that population pharmacokinetic (popPK) models combined with patient covariates (e.g., body weight, age) may improve the clinical target attainment of busulfan [5, 12]. Previous studies have shown that Bayesian-guided monitoring is faster and more reliable than equation-guided monitoring for various drugs [13]. Nevertheless, further improvement in busulfan PK prediction and optimisation of PK sampling strategies remain important for individualised dosing. Nowadays, machine learning (ML) models are becoming a popular method in personalised medicine, pharmacovigilance, and regulatory decision-making [14–16]. Thus, Woillard et al. have demonstrated that the XGBoost algorithm could sufficiently predict the exposure of tacrolimus [17, 18]. McCune et al. used ML to identify a set of 13 endogenous metabolites for predicting busulfan clearance, highlighting the potential advantages of ML in handling high-dimensional biological data [19]. However, compared to popPK, ML methods are less frequently used for drug PK prediction. The key difference between pharmacometrics and ML methodology is that pharmacometric models are based on biological mechanisms. This can be from compartmental models (i.e., popPK) to fully mechanistic models, such as physiology-based pharmacokinetics (PBPK) and quantitative systems pharmacology (QSP) models. In contrast, ML typically relies on purely data-driven methods. More studies and examples are needed to enable researchers to compare the two fields, bridge the gaps between methodologies, and establish more accurate PK predictions for clinical use, drug discovery, and development.
In this study, we developed a new popPK model of busulfan for our cohort of paediatric patients. Although numerous popPK models have been published in paediatric patients, their applicability across different clinical settings may be limited, as exposure monitoring methods and therapeutic targets are heterogeneous across studies [20]. Furthermore, paediatric populations vary in age distribution, disease severity, organ maturation, comedications, and supportive care practices, all of which can alter drug disposition and reduce the external validity of existing models. To ensure accurate exposure predictions and support model-informed dosing in our clinical context, we developed a new popPK model to characterise PK behaviour specifically within our patient population. Additionally, two classical ML methods, Random Forest and XGBoost, were employed to predict the concentration and AUC of busulfan for comparison, using the same dataset. Another external dataset was used to examine the prediction performance of all models. Additionally, we evaluated a limited sampling strategy (LSS) to reduce the number of samples required for TDM.
Methods
Patient Data Collection
The popPK model and ML model were developed using an extensive sampling dataset comprising 67 paediatric patients (637 observations) who received busulfan-based conditioning before HSCT in Denmark between 2014 and 2024. The TDM monitoring was performed at the Laboratory of Paediatric Oncology (Bonkolab) at Copenhagen University Hospital, Rigshospitalet. Each patient received an infusion of busulfan over 3 h, administered for four consecutive days with a 24-h dosing interval. Depending on diagnosis and chemotherapy combination, the target total AUC over the 4-day treatment with busulfan was within 75–90 mg*h/L [6]. Blood samples were collected from each patient at time 0, 0.5, 1, 2, 4, and 6 h after the infusion was completed (i.e., 3, 3.5, 4, 5, 7, and 9 h after the infusion started). One sample was also collected 5 min before the infusion started for patients who were not on their first infusion day. Over the past decade, Bonkolab has employed a traditional compartmental analysis to individualise busulfan dosing, which calculated AUC using either a one- or two-compartment model and compared it to the target AUC for this single infusion. If the AUC calculation resulted in a change of more than 25% between the prior and next busulfan dose, another TDM measurement was performed after the next dose. The total turnaround time of this procedure for one infusion was 5–6 h from sample receipt to report.
An external TDM dataset from the University Medical Centre Utrecht, the Netherlands, was used for external evaluation of the developed popPK and ML models. This dataset contained 1871 observations from 171 children collected from 2014 to 2021. Busulfan was given once a day (around 3 h of infusion) for four consecutive days, with a starting dose according to a previous publication [6]. The samples were collected at 5 min, 1, 2, and 3 h after the end of a 3-h infusion. If patients need a dose adjustment of more than 25%, samples would be taken at the next dose round. For all patients, samples were collected on infusion Day 4 for quality purposes.
Bioanalysis
Busulfan was quantified in plasma samples from Bonkolab (Denmark, hereafter called the Danish dataset) with positive-mode electrospray ionisation LC-MS/MS, using either Waters Acquity UPLC (Waters, Milford, MA) coupled to a Waters Xevo TQS mass spectrometer (Waters, Milford, MA), or ExionLC AD UHPLC (AB Sciex, MA, USA) coupled to a Sciex QTRAP 6500+ mass spectrometer (AB Sciex, MA, USA). Samples were analysed according to a published method [21] with adjustments. In brief, 100 µL heparin plasma was added to 40 µL internal standard solution (1 µg/mL busulfan-d8). To each sample, 1 mL ethyl acetate was added and thoroughly vortex mixed, followed by centrifugation for 5 min at 15,000×g. Then, 350 µL of the organic phase was transferred to new tubes and evaporated with nitrogen gas at 30 °C. Samples were resuspended in 50 µL 1.4 mM ammonium formate, and 0.5 µL were injected into the UPLC-MS/MS system (Waters UPLC BEH C18 column (50 × 2.1 mm), isocratic elution with 3.5 mM ammonium formate with 0.05% formic acid, 0.3 mL/min flow rate). Samples taken at time points 0, 0.5 h, and 1 h after infusion were additionally diluted 1:1 with foetal bovine serum before analysis. The lower limit of quantification (LLOQ) was 0.156 mg/L.
The busulfan samples from the University Medical Centre Utrecht (the Netherlands, hereafter called the Dutch dataset) were analysed using LC-MS/MS as previously published [2, 22].
Equations
The prediction error (PE), absolute prediction error (APE), and root mean squared error (RMSE) were determined as Eqs. (1), (2), and (3), respectively:
| 1 |
| 2 |
| 3 |
where yPre and yObs are the predictions and corresponding observations, respectively, and N is the number of observations considered.
The R2 was used when developing the ML model (and the Lasso model) to evaluate model performance on the training and validation datasets. The R2 between observations and predictions was calculated using Eq. (4):
| 4 |
where yPre and yObs are the individual predictions through the ML model and corresponding observations, respectively. The SSres is the sum of squared residuals, and the SStot is the total sum of squares. A higher R2 value indicates a better model fit so that model can explain a greater proportion of the observed variation.
Population PK Model
Development
The Danish dataset was used for the popPK model development. The first-order conditional estimation with interaction (FOCE-I) method was employed for parameter estimation. A low proportion of observations (5.2%) were below the limit of quantification (BLOQ) and were replaced with half LLOQ (M5 method), as it has been shown in other studies that this method is acceptable with such low number of data points BLOQ [23, 24]. The exponential model was used to describe the IIV. For residual unexplained variability (RUV), the additive, proportional, and combined additive and proportional error models were tested. For hierarchical models, the likelihood ratio test (LRT, p < 0.05) was used for model selection, whereas the Akaike information criterion (AIC) was used for non-hierarchical models.
Basic demographic characteristics (weight, age, sex) and number of infusions were evaluated as candidate covariates. The stepwise covariate modelling was used for the covariate model development with forward selection (p < 0.05) and backwards elimination (p < 0.001). The linear model was tested for categorical covariates, while linear, piece-wise linear, power, and exponential models were tested for the continuous covariates. The 90% bootstrap confidence intervals (CI) for parameters were calculated by Bootstrap (samples = 2000). The goodness-of-fit plot and prediction corrected visual predictive check (pcVPC) plot were also conducted to assess the predictive performance of the final model.
Limited Sampling Strategy
A popPK-informed LSS was tested on the Danish dataset and used to estimate busulfan AUC. The Bayesian maximum a posteriori (MAP) estimates of predicted CL were calculated by re-estimating from a limited number of samples from each patient following a single infusion. Then, this predicted CL was used to calculate the predicted AUC. The reference-true AUC is calculated using the MAP estimate of CL, based on all sample information for that patient. Multiple combinations of points, including three and four samples from 0, 0.5, 1, 2, and 4 h post-infusion, were evaluated. The PE, APE, and RMSE were calculated through Eqs. (1), (2), and (3), respectively. The yPre and yObs were the AUCs with a limited number of samples and with all sample information, respectively. For patients with more than one interval AUC observation, PE, APE, and squared prediction errors from the different infusion intervals were first averaged at the patient level before calculating the overall performance metrics. The bias was evaluated by the median of PE (MDPE), whereas the precision was evaluated through the RMSE and the median of APE (MDAPE).
External Evaluation
Prediction performance evaluation: The final popPK model was externally evaluated through the Dutch sample dataset. The PE, APE, and RMSE were determined as Eqs. (1), (2), and (3), respectively. In this part, the yPre and yObs were the population predictions (from NONMEM) and the corresponding observed busulfan concentration at each time point, respectively. Based on previous studies, the criteria for an acceptable model were set to: MDPE ≤ ±15%; MDAPE ≤ 30%; the fraction of PE within ±20% (F20) ≥ 35%; the fraction of PE within ±30% (F30) ≥ 50% [25–28]. The fraction of PE within ±15% (F15) was also calculated.
Bayesian Forecasting Evaluation: The Bayesian forecasting performance was evaluated on three different occasions: (1) Scenario 1: The plasma concentrations at 0 and 2 h after end of infusion during the first day were used as a priori information for each patient to make the MAP prediction of the concentrations for the rest of the infusion days (170 children). (2) Scenario 2: All plasma concentration information during the first day of infusion was used to predict the concentration on the remaining days (170 children). (3) Scenario 3: All plasma concentration information during the first 2 days of infusion was used to predict the concentration on the remaining days (99 children). The PE, APE, and RMSE of the individual MAP predictions were calculated as previously described in “Prediction performance evaluation” but using individual prediction (from NONMEM) as yPre, instead of population prediction.
Machine Learning Models
The ML models (Random Forest and XGBoost) were evaluated in this study for predicting the concentration and exposure (AUC) of busulfan.
For ML models predicting the busulfan concentration, features included in the training dataset were time after dose (TAD), dose, infusion day, infusion duration, body weight, age, and sex. The Danish sample dataset was used for model training and was split in 80% training and 20% validation for fivefold cross-validation. The overall predictive performance was compared through the average of the five validation datasets. Hyperparameter tuning was performed using a grid search to optimise the final ML model. Subsequently, the Dutch dataset, which was not included in the model development process, was used as an external testing dataset.
For ML models predicting the busulfan exposure (AUC), features included in the training dataset were concentrations at different time points (0, 0.5, 1, 2, 4, and 6 h post-infusion), dose, infusion day, infusion duration, body weight, age, and sex. The model development was similar to that previously described, but instead of predicting the concentration, the AUC calculated from the popPK model was predicted. Thus, for this specific analysis, only the Danish dataset was employed. The model performance was assessed by averaging the scores (testing dataset) across fivefold cross-validation.
Feature importance and their effects on the model were assessed using Shapley additive explanations (SHAP) summary plots, a game-theoretic approach that quantifies the contribution of each variable to the prediction using Shapley values [29]. Each point on the plot represents an individual observation, where the x-axis indicates the SHAP value (the impact on the model output), and the colour gradient represents the feature value. This illustrates the directional relationship between the predictors and the target outcome.
Software Usage
The popPK model development and evaluation were done in NONMEM 7.5.0 (ICON Development Solutions, Ellicott City, Maryland, USA) and PsN 5.3.0 (Department of Pharmacy, Uppsala University, Uppsala, Sweden). The statistical analysis and visualisation were programmed through R software (version 4.5.0). The ML models were developed through Python (version 3.12.7), with packages Scikit-learn (version 1.8.0) and XGBoost (version 3.2.0).
Results
Patient's Demographic Information
The demographic information of the internal model development dataset (Copenhagen, Denmark) and the external model evaluation dataset (Utrecht, Netherlands) are shown in Table 1. Only paediatric patients (age defined as <18 years) were included in this study, and the calculated median age in both datasets was around 6 years. The details of the age distribution plot were shown in Supplementary material 1, Figure S1.1. Two-thirds (45/67) of the patients in the Danish dataset were transplanted due to haematological malignant diseases. Other disease-specific information, e.g., diagnosis type, can be found in Supplementary material 2, Table S2.1. In the Danish dataset, 6–7 samples were collected for each patient, while only 4 samples were collected for each patient in the Dutch dataset.
Table 1.
Demographic summary of the datasets for internal model development (Denmark) and external evaluation (the Netherlands)
| Mean ± SD/numbers | Range | |
|---|---|---|
| Internal model development dataset (Copenhagen, Denmark) | ||
| Number of patients/observations | 67/637 | |
| Male/female | 35/32 | |
| Age (years) | 8 ± 5.3 | 0.6–17.4 |
| Body weight (kg) | 31 ± 20.8 | 6.8–99.6 |
| Busulfan dose (mg) | 129 ± 70 | 14–290 |
| By infusion interval | Number of patients/observations | |
| 1st infusion interval | 67/426 | |
| 2nd infusion interval | 31/187 | |
| 3rd infusion interval | 4/24 | |
| Co-administered agents (number of patients) | ||
| ATG | 46 | |
| Cyclophosphamide | 35 | |
| Melphalan | 37 | |
| Etoposide | 1 | |
| Thiotepa | 14 | |
| Fludarabine | 28 | |
| External evaluation dataset (Utrecht, Netherlands) | ||
| Number of patients/observations | 171/1871 | |
| Male/female | 87/84 | |
| Age (years) | 7.2 ± 5.5 | 0.2–17.8 |
| Body weight (kg) | 28.6 ± 20.9 | 3.8–127 |
| Busulfan dose (mg) | 120 ± 62 | 14–336 |
| By infusion interval | Number of patients/observations | |
| 1st infusion interval | 170/676 | |
| 2nd infusion interval | 100/395 | |
| 3rd infusion interval | 32/126 | |
| 4th infusion interval | 169/674 | |
ATG anti-thymocyte globulin, SD standard deviation
Population Pharmacokinetic Model
The final model was a one-compartment model with IIV on CL and volume of distribution (V), and a combined residual error model (Table 2). The goodness-of-fit plots are shown in Supplementary material 1 (Figure S1.2). The pcVPC (Figure S1.3) shows that the final model adequately captures the overall central tendency and variability of the data, although a minor discrepancy between observed and simulated values at the early phase after infusion stopped. The population CL and V for a typical patient with the median weight of 22.8 kg on the first day of infusion were 5.5 L/h (90% CI [5.12, 5.95]) and 17.5 L (90% CI [16.36, 18.84]), respectively. A similar interindividual variability was observed in CL and V, with a coefficient of variation (CV%) of 35.5% and 31.8%, respectively.
Table 2.
Final population pharmacokinetic model parameters
| Parameter | Estimate (RSE%) | Bootstrap 90% confidence interval |
|---|---|---|
|
Fixed effects Clearance model
If 1st time infusion, If 2nd time infusion, If ≥ 3rd time infusion, | ||
| 5.5 (4.7%) | [5.12, 5.95] | |
| 0.83 (8.2%) | [0.72, 0.94] | |
| −0.104 (16.2%) | [−0.14, −0.08] | |
| −0.13 (25.8%) | [−0.19, −0.03] | |
|
Volume of distribution model
| ||
| 17.5 (4.4%) | [16.36, 18.84] | |
| 0.898 (7%) | [0.79, 1.00] | |
|
Random effects Inter-individual variability model (exponential model) | ||
|
0.119 (32.4%) [CV%=35.5%] [Shrinkage = 0.3%] |
[0.06, 0.18] | |
| 0.0923 | [0.03, 0.16] | |
|
0.0961 (49.6%) [CV%=31.8%] [Shrinkage = 1%] |
[0.03, 0.17] | |
| Residual unexplained variability model (combined model) | ||
|
0.00252 (24%) [Shrinkage = 11%] |
[0.0014, 0.0034] | |
|
0.00337 (11.6%) [Shrinkage = 11%] |
[0.00276, 0.00408] | |
CL clearance, CV coefficient of variation, RSE relative standard error, V volume of distribution, WT total body weight (kg) of the patient. The unit of is mg/L.
After the forward inclusion and backward exclusion of all candidate covariates, WT was found to have a significant influence on CL and V. This influence was incorporated using an allometric exponent, with the weight relative to the median weight of 22.8 kg, yielding estimates of 0.83 and 0.898 for CL and V, respectively. Additionally, CL, which time-dependence was tested as a covariate, decreased on the second and third infusions. Compared with Day 1, CL decreased by an average of 10.4% on Day 2 and 13% on Day 3.
Limited Sampling Strategy
The LSS, used for the final popPK model to predict AUC, was evaluated across several combinations of sampling times. Table 3 summarises five sampling combination schedules. Generally, all selected LSS schedules demonstrated acceptable bias (MDPE within ± 2.1%) and precision (RMSE < 0.9 mg∙h·L−1, MDAPE < 2.5%) in prediction. The three-sample Schedule 1 (0, 1, 2 h post-infusion) and four-sample Schedule 3 (0, 0.5, 1, 2 h post-infusion) showed lower performance in both bias (MDPE lower than −2%) and precision (RMSE > 0.8 mg∙h·L−1, MDAPE > 2%) than the prediction of other LSS schedules, possibly due to their narrow sampling interval right after the stop of infusion. The combination of four time points in Schedule 4 (0, 0.5, 1, 4 h post-infusion) showed the best predictive performance, with almost no bias (MDPE = −0.24%) and high precision (RMSE = 0.29 mg∙h·L−1, MDAPE = 0.71%).
Table 3.
The prediction performance of clearance when limited samples were available
| Sample schedule | RMSE | MDPE (%) | MDAPE (%) |
|---|---|---|---|
| Schedule 1 (3 samples: 0, 1, 2 h) | 0.84 | −2.04 | 2.11 |
| Schedule 2 (3 samples: 0, 0.5, 4 h) | 0.27 | 0.39 | 0.93 |
| Schedule 3 (4 samples: 0, 0.5, 1, 2 h) | 0.86 | −2.09 | 2.25 |
| Schedule 4 (4 samples: 0, 0.5, 1, 4 h) | 0.29 | −0.24 | 0.71 |
| Schedule 5 (4 samples: 0, 1, 2, 4 h) | 0.24 | −0.63 | 0.63 |
MDAPE median absolute prediction error, MDPE median prediction error, RMSE root mean squared error (mg∙h∙L−1)
Time points are presented as time after the infusion end, where 0 h refers to the time point immediately after the infusion finished
External Evaluation of the Final popPK Model
To assess the predictive performance of the final popPK model, a TDM dataset from Utrecht was used as an external evaluation dataset. In general, the final popPK model predicts the plasma concentration of the external dataset with acceptable negative bias (MDPE = − 4.82%) and precision (MDAPE = 15.16%) (Table 4). Around half of the data were predicted within ± 15% of PE, and more than 80% were predicted within ± 30% of PE. Based on this external evaluation, the final popPK model met all predefined requirements (MDPE ≤ ±15%; MDAPE ≤ 30%; F20 ≥ 35%; F30 ≥ 50%). These criteria were also met when the data were grouped by infusion day (Supplementary Material 2, Table S2.2). Figure S1.4 (Supplementary Material 1) showed the plot of population predictions versus observations on the external dataset.
Table 4.
The final popPK model performance of population prediction and Bayesian forecasting (with three different scenarios) of plasma concentrations in the external Dutch dataset
| MDPE (%) | MDAPE (%) | F15 (%) | F20 (%) | F30 (%) | |
|---|---|---|---|---|---|
| Population prediction evaluation | |||||
| External dataset (all data) | −4.82 | 15.16 | 49.71 | 61.68 | 80.22 |
| Bayesian forecasting evaluation | |||||
| Scenario 1 (Day 1, 0 & 2 h) | 2.55 | 9.19 | 69.02 | 79.26 | 89.17 |
| Scenario 2 (All Day 1 samples) | 2.64 | 9.43 | 68.51 | 78.42 | 89.84 |
| Scenario 3 (All Day 1 & 2 samples) | 3.55 | 9.95 | 71.48 | 81.25 | 89.65 |
Scenario 1: Using plasma concentrations at 0 and 2 h after infusion stopped during the Day 1 infusion (a priori) to predict the concentrations on rest days. Scenario 2: Using all plasma concentrations during the Day 1 infusion (a priori) to predict the concentrations on rest days. Scenario 3: Using all plasma concentrations during Days 1 and 2 infusions (a priori) to predict the concentrations on rest days
F15, F20, and F30 fractions of prediction error within ±15%, ±20%, and ±30%, respectively, MDAPE median absolute prediction error, MDPE median prediction error
The Bayesian forecasting performance of the final popPK model was evaluated through three different scenarios (Table 4). The results showed that individual prediction performance improved for both bias (MDPE = 2.55%) and precision (MDAPE = 9.19%) when using plasma concentrations at 0 and 2 h after infusion stopped during the Day 1 infusion as a priori information (Scenario 1). With these two concentrations, around 70% of the data were predicted within ±15% of PE, and nearly 90% were predicted within ±30% of PE. Interestingly, more concentration points (Scenarios 2 and 3) did not significantly improve this prediction performance over Scenario 1.
Machine Learning Models Predict the PK Profile
Two classical ML methods (Random Forest and XGBoost) and Lasso regression were trained and optimised to predict busulfan concentration over time (Table 5). Lasso regression uses only linear relationships and does not describe the data well, especially as time after dose increases (Supplementary information 1, Figure S1.5). It had an R2 of only 0.49 on the model development dataset and was not used for external evaluation. Both Random Forest and XGBoost models showed acceptable performance on the model development Danish dataset and were used for external evaluation. For the external Dutch dataset, the Random Forest and XGBoost achieved similar prediction performance (Table 5) compared to the performance of the population prediction of the popPK model (Table 4), in terms of bias (−7.39% and −7.44% for MDPE, respectively), precision (14.84% for both MDAPE), and a similar fraction of points that were predicted within ±15%, ±20%, ±30% of PE (Table 5). However, this performance was worse than the popPK model's individual prediction with MAP (Table 4). Additionally, the ML models slightly under-predicted the concentration points over 6 mg/L, where the Danish dataset used for model development did not have data (Supplementary information 1, Figure S1.4). The distribution of PE across different time points using ML models is shown in Supplementary Information 1 (Figures S1.6 and S1.7).
Table 5.
The performance for predicting plasma concentration over time through different machine learning models
| Random forest (%) | XGBoost (%) | |
|---|---|---|
| Testing dataset (Netherlands) | ||
| MDPE | −7.39 | −7.44 |
| MDAPE | 14.84 | 14.84 |
| F15 | 50.4 | 50.24 |
| F20 | 64.35 | 62.91 |
| F30 | 83.54 | 81.45 |
F15, F20, and F30 fractions of prediction error within ±15%, ±20%, and ±30%, respectively, MDAPE median absolute prediction error, MDPE median prediction error, XGBoost Extreme Gradient Boosting
To provide human-interpretable explanations for ‘black-box’ ML models, SHAP plots of both the Random Forest and XGBoost models are presented in Fig. 1. The SHAP values were calculated for all predictors retained in the final model, including TAD, which is a mechanistically relevant determinant of drug concentration. Both models ranked the TAD, dose, and the day of infusion as the top three most important predictors of plasma concentration. As expected, the TAD and dose were the most influential predictors for both models, demonstrating the widest spread of SHAP values and the greatest impact on prediction variability. Also, as expected, lower TAD or higher dose values are generally associated with positive SHAP values, indicating an increase in predicted outcome (higher concentration). Moreover, the infusion day was the third most contributing feature for both models. A higher value on infusion day (Days 2 or 3) than on Day 1 showed a positive SHAP value, and thus a higher predicted concentration.
Fig. 1.

SHAP summary plot of feature importance in Random Forest (A) and XGBoost (B) models for predicting concentration versus time. Each point represents the SHAP value for a single observation. Features are ordered by overall importance based on mean absolute SHAP values from top to bottom. Colour represents the magnitude of the values of the feature (blue = low, red = high). Age, age (years); Dose, daily busulfan dose (mg/kg); Duration, infusion duration (h); Infusion day, the number of infusions (1, 2, or 3, where Days 3 and 4 will be treated as 3); Sex, gender (0, 1); SHAP Shapley additive explanations, TAD time after dose (h); Weight, body weight (kg)
Figure 2 shows how different models described the concentration-time profiles of two representative patients (Supplementary material 2, Table S2.3) across three infusion days. All models adequately described the profile in the first 6 h after the end of infusion (around TAD 3–9 h). Since the PopPK model is a mechanism-based model, it describes a reasonable terminal phase, for which we have insufficient data in the training dataset. However, the ML models showed large deviations in the terminal phase, with a tendency toward flatter elimination profiles.
Fig. 2.

Concentration–time profiles by different models for two representative patients across three treatment days. Predicted concentrations from the PopPK model (simulation-based, green), Random Forest (blue), and XGBoost (red) are shown for two representative patients. Shaded areas and dashed lines represent PopPK prediction intervals. The dotted lines are the 5th, 10th, 25th, 75th, 90th, and 95th percentiles at every hour of the simulated data (1000 replicates for each patient), and the solid lines are the median. Time after dose (TAD) is shown on the x-axis and concentration on the y-axis
Machine Learning Models Predict the AUC
Two ML algorithms (i.e., Random Forest and XGBoost) were also employed to predict the popPK-calculated AUC for the Danish dataset with four time-point LSS (Fig. 3). All four-point LSS demonstrated acceptable predictive performance, with R2 > 0.9, using both algorithms. The schedule with narrow sampling at the beginning of the post-infusion period (0, 0.5, 1, 2 h post-infusion) showed the worst performance across both algorithms. Additionally, the schedule with spread sampling (0, 1, 4, 6 h post-infusion) performed best (R2 was 0.945 and 0.95 in Random Forest and XGBoost, respectively) across all LSS schedules, and its R2 scores in both algorithms were similar to the all-time-point model respectively. In Fig. 3, schedules with a 4-h point after infusion ended generally performed better than other schedules. In some cases, the 6-h post-infusion time point is inconvenient in clinical practice and should be avoided. Considering this, the schedule (0, 1, 2, 4 h post-infusion) performed best with Random Forest, and the schedule (0, 0.5, 1, 4 h post-infusion) performed best with XGBoost. The RMSE of both algorithms predicting AUC with different schedules is shown in Figure S1.8 (Supplementary information 1).
Fig. 3.

Predictive performance (R2) of AUC for Random Forest (A) and XGBoost (B) models across limited sampling strategies. The gold bars and dashed horizontal lines represent performance using all available time points, while blue (A) and pink (B) bars represent reduced sampling schemes. AUC area under the curve
Discussion
In this study, a new popPK model for IV infusion of busulfan in our cohort of paediatric patients was developed using retrospectively collected data. A one-compartment model adequately described our data. The final popPK model achieved an acceptable performance through an external evaluation with a Dutch dataset. Total body weight was identified as a contributing factor to the variability through an allometric relationship. Our model estimated the typical CL to be 5.5 L/h for a patient weighing 22.8 kg during the first infusion. According to our model, a standard 70 kg patient has a CL of 13.95 L/h, which is slightly higher than 10.45 L/h by Bognar et al. [30] and 12.4 L/h by McCune et al. [3] for patients of the same body weight, but slightly lower than the 14.35 L/h from a recent model [31]. A comparison across models [3, 30–32] is provided in Table S2.4 (Supplementary material 2). Furthermore, as in previous research [2, 33–35], we also observed a time-dependent decrease in CL. Compared to Day 1, the CL decreased by around 10.4% and 13% on Days 2 and 3, respectively. A systematic review summarises 12 busulfan models with time-varying CL, and they estimate that the CL on Days 2 to 4 is 6–13.5% (median 7.6%) lower than that on Day 1 [36]. The decrease in busulfan clearance might be explained by interference of busulfan metabolites with GST-mediated conjugation and depletion of glutathione [37]. The supportive care drugs were not tested in our covariate search, but a systematic review summarises that only a few studies report fludarabine and the fludarabine/clofarabine combination to affect CL, whereas no other drugs were identified as significant covariates by population PK methods [36].
Using fewer samples to make an accurate prediction is an advantage of employing popPK models in TDM. From this study, TDM sample schedules with 4 points (0, 0.5, 1, 4 or 0, 1, 2, 4 h post-infusion) were shown to adequately predict the AUC of busulfan through the final popPK model and were recommended as potential LSS schedules. The three-sample schedule (0, 0.5, 4 h post-infusion) also showed a satisfactory prediction of AUC. However, it could be problematic to further reduce the number of samples for LSS to two, as a single compromised sample will lead to unreliable predictions. Thus, for better clinical practice, four samples (at least three) are recommended for LSS. The suggested LSS schedules avoided the 6-h sample point, thereby shortening the overall turnaround time of the procedure. From a clinical point of view, this is a major improvement in the procedure, as the analysis must be completed and evaluated the same day as blood sampling to adjust the dose for the next day. Moreover, most patients undergo transplantation for malignant diseases and are therefore already exposed to intensive treatment. As many of these patients are children, minimising blood sampling is important due to their limited blood volume and to reduce the overall procedural burden. A user-friendly software was developed to use our model for Bayesian dose optimisation (Supplementary material 3).
In our results, the ML models did not extrapolate the terminal kinetics phase, unlike the mechanistic models (Fig. 2). The mechanistic structure of pharmacometric models allows them to extrapolate beyond the data used for model development, which is a key advantage when analysing sparse datasets. Pharmacometric methods can be used to simulate various scenarios beyond initial data, commonly used for clinical trial simulations in dose selection. Whereas the parameters of pharmacometric models can be interpreted directly in biological terms, the ML parameters can only be understood mathematically and have limited biological interpretation. Methods like the SHAP summary plot can improve the interpretability of ML models. In our results, the SHAP plots show the infusion day as an important feature for predicting plasma concentration, consistent with the popPK results. A later infusion day may result in a higher concentration, which could be explained by a decrease in CL in the latter part of the infusion, as popPK showed. But one can also explain this as the given dose might be higher on Days 2 and 3 of infusion, as the ML models lack physiologically explainable parameters.
On the other hand, previous mechanistic knowledge used for pharmacometrics might introduce bias into the analysis if it is misinterpreted. The development of pharmacometric models is a repetitive and time-consuming process that often requires running many models to obtain the final model (e.g., to develop the final busulfan model here, around 110 models were run). Conversely, ML models are timesaving regardless of the model development process. The built-in feature selection for the ML algorithms enables a single run for model development, in contrast to the repeated model runs required for covariate analysis in pharmacometric models. Difficulties for data-driven machine learning algorithms arise from the need for rich and high-quality data to develop robust models [14, 15]. From our results, the ML models could adequately predict the data through external evaluation. These ML models were developed on a dense dataset with only 67 children. However, it needs to be noted that most of the points of the external dataset here were in the first 3 h post-infusion, which have plenty of data in the training dataset. Also, the ML models did not describe the PK profile well, either during the infusion phase or the terminal phase of drug elimination, as they were not able to extrapolate to areas where no data were available. This might indicate that sampling covering the entire timespan is needed for the ML models to describe the full pharmacokinetic profile. In this study, both ML algorithms showed the potential ability to predict AUC when using four time-point LSS schedules. The schedules recommended by the PopPK model (0, 0.5, 1, 4 h post-infusion) or (0, 1, 2, 4 h post-infusion) also demonstrated acceptable AUC prediction performance when using the ML algorithm. These results demonstrated the potential of using ML models to predict busulfan pharmacokinetic exposure (AUC) with LSS. However, the ML algorithms in this study required concentrations collected at predefined time points, whereas the PopPK method accounts for the actual sampling time of each concentration measurement. Although sampling generally follows predefined schedules in clinical practice, the PopPK may offer more flexibility when deviations from nominal sampling times occur.
Although our model underwent external evaluation using data from an independent centre, an important limitation is that model development relied on a relatively small retrospective dataset from a single institution. While the model accurately reflects the characteristics of our local cohort, its performance may not generalise to populations with different demographic or clinical profiles. To enhance broader applicability, larger prospective, multicentre, and ideally international, studies (particularly within clearly defined conditioning regimens) are needed to establish models with robust external validity. Additionally, a previous study demonstrated that busulfan clearance achieves approximately 95% of the adult value by around 2.5 years of postnatal age [3]. However, the number of children younger than 2.5 years in our model development dataset was limited, which precluded reliable estimation of a maturation function. Therefore, the final model should be applied with caution in children aged < 2.5 years. Another limitation of our study is that there was no standardised sampling on Days 2 and 3. Patients who had samples taken on Days 2 and 3 were usually those patients who did not achieve the target AUC on the first day of infusion, as the clinicians were generally more concerned with dose increments than reductions. This might introduce an overestimation of the reduction in CL, in line with the fact that we found a slightly larger CL reduction (10.4%) with repeated infusions than reported in the literature (Median 7.6%). Additionally, the evaluation dataset was more concentrated on the time points around 0–3 h post-infusion. Thus, the model performance on predicting the concentration of the terminal phase was not well evaluated externally.
Conclusion
This study showed a comparison between the performance of a popPK model and ML models for predicting busulfan concentrations. A popPK model was developed and externally evaluated based on two different paediatric datasets, respectively. A time-varying CL was observed, with CLs at the second and third infusions lower than on the first infusion day. Limited sampling strategies with sampling at 0, 0.5, 1, 4 h or 0, 1, 2, 4 h after the end of infusion could adequately predict the exposure of busulfan in our patients. In addition, we found that two commonly used ML algorithms (Random Forest and XGBoost) could achieve similar prediction performance of the plasma concentrations of busulfan compared to a traditional popPK model within the timeframe where data was available. However, to fully describe the pharmacokinetic profile with the ML model, a sampling of data covering the entire timespan was needed. Thus, popPK models will still be the preferred method for TDM on busulfan in the clinic.
Supplementary Information
Below is the link to the electronic supplementary material.
Acknowledgements
We want to kindly acknowledge C.A. Lindemans, R. Admiraal and T. Bognar from the University Medical Centre Utrecht and the Princess Máxima Centre for Paediatric Oncology, both in Utrecht the Netherlands for sharing data for validation purposes upon reasonable request. Furthermore, R. Admiraal is acknowledged for fruitful discussions. The authors would like to thank all participants in this study. The authors would like to thank the lab technicians for their sample analysis.
Funding
Open access funding provided by Copenhagen University.
Declarations
Ethics approval
The study was conducted in accordance with the Declaration of Helsinki. Permission was obtained from the Team for Journal Data (R-24059676), Research Ethics and Research Support, Capital Region of Denmark, to access and process relevant patient health information and other confidential personal data from medical records for research purposes without obtaining individual informed consent from the included patients or their legal guardians. Data from the University Medical Centre Utrecht and the Princess Máxima Centre for Paediatric Oncology, both in Utrecht, the Netherlands, have been shared for validation purposes upon reasonable request by C.A. Lindemans, R. Admiraal and T. Bognar. The Dutch data were sampled after ethical committee approval under number 11/063-k by the Medical Ethical Committee in Utrecht.
Conflicts of interest
Kjeld Schmiegelow received speaker and/or Advisory Board Honoraria from Illumina (2021), Jazz Pharmaceuticals (2021, 2023) and Servier (2021, 2023, 2025). Kjeld Schmiegelow received speaker fees from Amgen (2021, 2023, 2025) and Medscape (2021), educational grants from Servier (2021, 2023, 2025) and SERB (2025), and research grants from Novo Nordisk Foundation (2020, 2022). Kjeld Schmiegelow has stocks in Novo Nordisk. All other authors have no competing interests to declare that are relevant to the content of this article.
Availability of data and material
Due to GDPR and institutional privacy regulations, the clinical data underlying this study cannot be made publicly available. However, the data that support the findings of this study are available from the corresponding author upon reasonable request.
Author contributions
T.Y., M.T., K.S., and T.M.L. conceptualised the study. T.Y., M.T., T.N.M., H.H.U., J.N., M.J.P., R.A., and A.S.B.R. curated the data. T.Y. performed the formal analysis, developed the methodology and software, validated and visualised the data, and wrote the original draft of the manuscript. M.T., T.N.M., H.H.U., J.N., M.J.P., A.S.B.R., K.S., and T.M.L. provided resources and reviewed and edited the manuscript. K.S. and T.M.L. acquired the funding, administered the project, and supervised the study. All authors read and approved the final version of the manuscript.
Funding
This research is funded by the University of Copenhagen and the Copenhagen University Hospital.
Consent to participate
Not applicable.
Consent for publication
Not applicable.
Code availability
The code is available from the corresponding author upon reasonable request.
Footnotes
Kjeld Schmiegelow and Trine Meldgaard Lund shared last authorship.
References
- 1.Lawson R, Paterson L, Fraser CJ, Hennig S. Evaluation of two software using Bayesian methods for monitoring exposure and dosing once-daily intravenous busulfan in paediatric patients receiving haematopoietic stem cell transplantation. Cancer Chemother Pharmacol. 2021;88:379–91. 10.1007/s00280-021-04288-0. [DOI] [PubMed] [Google Scholar]
- 2.Bartelink IH, Boelens JJ, Bredius RGM, Egberts ACG, Wang C, Bierings MB, et al. Body weight-dependent pharmacokinetics of busulfan in paediatric haematopoietic stem cell transplantation patients: towards individualized dosing. Clin Pharmacokinet [Internet]. 2012;51:331–45. 10.2165/11598180-000000000-00000. [DOI] [PubMed] [Google Scholar]
- 3.McCune JS, Bemer MJ, Barrett JS, Scott Baker K, Gamis AS, Holford NHG. Busulfan in infant to adult hematopoietic cell transplant recipients: a population pharmacokinetic model for initial and Bayesian dose personalization. Clin Cancer Res. 2014;20:754–63. 10.1158/1078-0432.CCR-13-1960. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Dix SP, Wingard JR, Mullins RE, Jerkunica I, Davidson TG, Gilmore CE, et al. Association of busulfan area under the curve with veno-occlusive disease following BMT. Bone Marrow Transplant [Internet]. 1996;17:225–30. [PubMed] [Google Scholar]
- 5.Shukla P, Goswami S, Keizer RJ, Winger BA, Kharbanda S, Dvorak CC, et al. Assessment of a model-informed precision dosing platform use in routine clinical care for personalized busulfan therapy in the pediatric hematopoietic cell transplantation (HCT) population. Front Pharmacol. 2020;11:888. 10.3389/fphar.2020.00888. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Bartelink IH, Lalmohamed A, van Reij EML, Dvorak CC, Savic RM, Zwaveling J, et al. Association of busulfan exposure with survival and toxicity after haemopoietic cell transplantation in children and young adults: a multicentre, retrospective cohort analysis. Lancet Haematol [Internet]. 2016;3:e526–36. 10.1016/S2352-3026(16)30114-4. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Trame MN, Bergstrand M, Karlsson MO, Boos J, Hempel G. Population pharmacokinetics of busulfan in children: increased evidence for body surface area and allometric body weight dosing of busulfan in children. Clin Cancer Res. 2011;17:6867–77. 10.1158/1078-0432.CCR-11-0074. [DOI] [PubMed] [Google Scholar]
- 8.Palmer J, McCune JS, Perales M-A, Marks D, Bubalo J, Mohty M, et al. Personalizing busulfan-based conditioning: considerations from the American Society for Blood and Marrow Transplantation Practice Guidelines Committee. Biol Blood Marrow Transplant. 2016;22:1915–25. 10.1016/j.bbmt.2016.07.013. [DOI] [PubMed] [Google Scholar]
- 9.Ciurea SO, Andersson BS. Busulfan in hematopoietic stem cell transplantation. Biol Blood Marrow Transplant. 2009;15:523–36. 10.1016/j.bbmt.2008.12.489. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Geddes M, Kangarloo SB, Naveed F, Quinlan D, Chaudhry MA, Stewart D, et al. High busulfan exposure is associated with worse outcomes in a daily i.v. busulfan and fludarabine allogeneic transplant regimen. Biol Blood Marrow Transplant. 2008;14:220–8. 10.1016/j.bbmt.2007.10.028. [DOI] [PubMed] [Google Scholar]
- 11.European Medicines Agency. Busulfan Fresenius Kabi: EPAR-Product Information [Internet]. 2025. https://www.ema.europa.eu/en/medicines/human/EPAR/busulfan-fresenius-kabi#product-info. Accessed 11 Jul 2026
- 12.Long-Boyle JR, Savic R, Yan S, Bartelink I, Musick L, French D, et al. Population pharmacokinetics of busulfan in pediatric and young adult patients undergoing hematopoietic cell transplant: a model-based dosing algorithm for personalized therapy and implementation into routine clinical use. Ther Drug Monit. 2015;37:236–45. 10.1097/FTD.0000000000000131. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Del Valle-Moreno P, Suarez-Casillas P, Mejías-Trueba M, Ciudad-Gutiérrez P, Guisado-Gil AB, Gil-Navarro MV, et al. Model-informed precision dosing software tools for dosage regimen individualization: a scoping review. Pharmaceutics. 2023;15:1859. 10.3390/pharmaceutics15071859. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.Huang S, Xu Q, Yang G, Ding J, Pei Q. Machine learning for prediction of drug concentrations: application and challenges. Clin Pharmacol Ther. 2025;117:1236–47. 10.1002/cpt.3577. [DOI] [PubMed] [Google Scholar]
- 15.Keutzer L, You H, Farnoud A, Nyberg J, Wicha SG, Maher-Edwards G, et al. Machine learning and pharmacometrics for prediction of pharmacokinetic data: differences, similarities and challenges illustrated with rifampicin. Pharmaceutics. 2022;14:1530. 10.3390/pharmaceutics14081530. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Rawal A, Ou J, Zhu H, Sauna Z, Tegenge MA. Utilization of machine learning approaches for drug clearance prediction and population pharmacokinetic covariate analysis. Clin Transl Sci. 2025;18:e70359. 10.1111/cts.70359. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Woillard J-B, Labriffe M, Prémaud A, Marquet P. Estimation of drug exposure by machine learning based on simulations from published pharmacokinetic models: the example of tacrolimus. Pharmacol Res. 2021;167:105578. 10.1016/j.phrs.2021.105578. [DOI] [PubMed] [Google Scholar]
- 18.Woillard J-B, Labriffe M, Debord J, Marquet P. Tacrolimus exposure prediction using machine learning. Clin Pharmacol Ther. 2021;110:361–9. 10.1002/cpt.2123. [DOI] [PubMed] [Google Scholar]
- 19.McCune JS, Navarro SL, Baker KS, Risler LJ, Phillips BR, Randolph TW, et al. Prediction of busulfan clearance by predose plasma metabolomic profiling. Clin Pharmacol Ther. 2023;113:370–9. 10.1002/cpt.2794. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Lawson R, Staatz CE, Fraser CJ, Hennig S. Review of the pharmacokinetics and pharmacodynamics of intravenous busulfan in paediatric patients. Clin Pharmacokinet. 2021;60:17–51. 10.1007/s40262-020-00947-2. [DOI] [PubMed] [Google Scholar]
- 21.Desire S, Mohanan EP, George B, Mathews V, Chandy M, Srivastava A, et al. A rapid & sensitive liquid chromatography-tandem mass spectrometry method for the quantitation of busulfan levels in plasma & application for routine therapeutic monitoring in haematopoietic stem cell transplantation. Indian J Med Res [Internet]. 2013;137:777–84. [PMC free article] [PubMed] [Google Scholar]
- 22.Bognàr T, Bartelink IH, van der Elst KCM, Kingma JS, Smeijsters EH, Lindemans CA, et al. Busulfan exposure target attainment in adults undergoing allogeneic hematopoietic cell transplantation: a single day versus a multiple day therapeutic drug monitoring regimen. Transplant Cell Ther. 2024;30:1007.e1-1007.e10. 10.1016/j.jtct.2024.07.015. [DOI] [PubMed] [Google Scholar]
- 23.Keizer RJ, Jansen RS, Rosing H, Thijssen B, Beijnen JH, Schellens JHM, et al. Incorporation of concentration data below the limit of quantification in population pharmacokinetic analyses. Pharmacol Res Perspect. 2015;3:e00131. 10.1002/prp2.131. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Irby DJ, Ibrahim ME, Dauki AM, Badawi MA, Illamola SM, Chen M, et al. Approaches to handling missing or “problematic” pharmacology data: pharmacokinetics. CPT Pharmacomet Syst Pharmacol. 2021;10:291–308. 10.1002/psp4.12611. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Zhao C-Y, Jiao Z, Mao J-J, Qiu X-Y. External evaluation of published population pharmacokinetic models of tacrolimus in adult renal transplant recipients. Br J Clin Pharmacol. 2016;81:891–907. 10.1111/bcp.12830. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Mao J-J, Jiao Z, Yun H-Y, Zhao C-Y, Chen H-C, Qiu X-Y, et al. External evaluation of population pharmacokinetic models for ciclosporin in adult renal transplant recipients. Br J Clin Pharmacol. 2018;84:153–71. 10.1111/bcp.13431. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Leven C, Ménard P, Gouin-Thibault I, Ballerie A, Lacut K, Ollier E, et al. Transferability of published population pharmacokinetic models for apixaban and rivaroxaban to subjects with obesity treated for venous thromboembolism: a systematic review and external evaluations. Pharmaceutics. 2023;15:665. 10.3390/pharmaceutics15020665. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Li R, Sun F, Feng Z, Zhang Y, Lan Y, Yu H, et al. Evaluation and application of population pharmacokinetic models for optimising linezolid treatment in non-adherence multidrug-resistant tuberculosis patients. Eur J Pharm Sci. 2024;203:106915. 10.1016/j.ejps.2024.106915. [DOI] [PubMed] [Google Scholar]
- 29.Lundberg S, Lee S-I. A unified approach to interpreting model predictions. 2017; http://arxiv.org/abs/1705.07874
- 30.Bognàr T, Bartelink IH, Egberts TCG, Rademaker CMA, Versluys AB, Slatter MA, et al. Association between the magnitude of intravenous busulfan exposure and development of hepatic veno-occlusive disease in children and young adults undergoing myeloablative allogeneic hematopoietic cell transplantation. Transplant Cell Ther. 2022;28:196–202. 10.1016/j.jtct.2022.01.013. [DOI] [PubMed] [Google Scholar]
- 31.Balis FM, Rieger E, Bunin NJ, Gardiner J, Shaw LM, Olson TS, et al. New approach to busulfan dosing in infants and children based on a population pharmacokinetic analysis. Cancer Chemother Pharmacol. 2025;95:32. 10.1007/s00280-025-04757-w. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32.Langenhorst JB, Boss J, van Kesteren C, Lalmohamed A, Kuball J, Egberts ACG, et al. A semi-mechanistic model based on glutathione depletion to describe intra-individual reduction in busulfan clearance. Br J Clin Pharmacol. 2020;86:1499–509. 10.1111/bcp.14256. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33.Takahashi T, Illamola SM, Jennissen CA, Long SE, Lund TC, Orchard PJ, et al. Busulfan dose recommendation in inherited metabolic disorders: population pharmacokinetic analysis. Transplant Cell Ther. 2022;28:104.e1-104.e7. 10.1016/j.jtct.2021.11.018. [DOI] [PubMed] [Google Scholar]
- 34.Rhee S-J, Lee JW, Yu K-S, Hong KT, Choi JY, Hong CR, et al. Pediatric patients undergoing hematopoietic stem cell transplantation can greatly benefit from a novel once-daily intravenous busulfan dosing nomogram. Am J Hematol. 2017;92:607–13. 10.1002/ajh.24734. [DOI] [PubMed] [Google Scholar]
- 35.Lawson R, Staatz CE, Fraser CJ, Ramachandran S, Teague L, Mitchell R, et al. Population pharmacokinetic model for once-daily intravenous busulfan in pediatric subjects describing time-associated clearance. CPT Pharmacomet Syst Pharmacol. 2022;11:1002–17. 10.1002/psp4.12809. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 36.Takahashi T, Jaber MM, Brown SJ, Al-Kofahi M. Population pharmacokinetic model of intravenous busulfan in hematopoietic cell transplantation: systematic review and comparative simulations. Clin Pharmacokinet. 2023;62:955–68. 10.1007/s40262-023-01275-x. [DOI] [PubMed] [Google Scholar]
- 37.Combarel D, Tran J, Delahousse J, Vassal G, Paci A. Individualizing busulfan dose in specific populations and evaluating the risk of pharmacokinetic drug-drug interactions. Expert Opin Drug Metab Toxicol. 2023;19:75–90. 10.1080/17425255.2023.2192924. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
