Busulfan (Bu) is an alkylating agent used in chemotherapy regimens alongside drugs such as cyclophosphamide (CY) and fludarabine (Flu) for hematopoietic stem cell transplantation (HSCT). Owing to concerns about the long-term effects of total-body irradiation in children, Bu-based conditioning regimens have been widely applied in the conditioning of pediatric hematopoietic stem cells. However, Bu has a narrow therapeutic window, and its pharmacokinetic characteristics show significant inter-individual variability, which is particularly pronounced in children (Marsit et al., 2020). Insufficient drug exposure is associated with a higher rate of transplant failure or relapse, while excessive exposure is associated with increased toxicity and transplant-related mortality (Bartelink et al., 2016). Notably, the efficacy and adverse drug reactions of Bu are closely related to the area under the concentration-time curve (AUC) of its blood concentration, hence therapeutic drug monitoring (TDM) is often required to achieve personalized drug administration (Rasor et al., 2019; Sweiss et al., 2020; Bognàr et al., 2024). The limited sampling strategy (LSS) is a method that involves collecting a small number of samples to measure drug concentrations, using pharmacokinetic models to determine the optimal sampling times, and then assessing drug exposure and combining population pharmacokinetics to estimate the AUC. This approach minimizes blood draws, making it particularly suitable for young infants and toddlers (Sibbald et al., 2006). Although some transplant centers have established models for the LSS to predict the AUC values of Bu, the differences in patient ethnicity, age, disease types, dosing regimens, and blood sampling schemes included in different models result in poor generalizability of these models. Therefore, based on the need for clinical TDM, we aimed to develop a LSS that is more suitable for predicting Bu exposure in Chinese children undergoing HSCT at our hospital.

The efficacy and adverse drug reactions of Bu are closely related to the AUC. As a cytotoxic drug, Bu has a narrow therapeutic window and significant inter-individual pharmacokinetic variability, particularly in children. Therefore, TDM is often required to achieve personalized dosing. The pharmacokinetic data from this study align with those from a previous study (Cremers et al., 2002) for the first dose administration (population estimates: CL = 0.29 L/h/kg and Vd = 0.84 L/kg). However, it has been reported that there is significant intra-individual variability in the CL of Bu when used in children prior to HSCT, with approximately 80% of patients experiencing a decrease in CL (Marsit et al., 2020). Since our study only included CL values from the first Bu dose, these did not represent the CL at steady state. Some studies have suggested that age is the primary factor affecting Bu exposure in children (Cremers et al., 2002; Faraci et al., 2018; Nakamura et al., 2008). Other studies have demonstrated a close logarithmic linear relationship between body weight and Bu CL, suggesting that weight-based dosing might be more appropriate for pediatric patients (Nguyen et al., 2004). Unlike other published studies, the predominant disease in our study was severe aplastic anemia, a non-malignant condition, and the proportion of younger children was relatively low, which may have led to certain differences in the pharmacokinetic data compared to other studies.

Although some studies have successfully established models to estimate the AUC_0-t values of Bu using the LSS method during the pre-experimental phase, our attempts to validate some of these established models yielded unsatisfactory results, with significant percentage errors (PE). As is well known, the classic pharmacokinetic calculation of AUC is performed by summing up trapezoidal areas, and a greater number of blood sampling points can yield a more accurate AUC value. Therefore, we collected 11 blood samples per patient, which allowed a more precise reflection of drug exposure in children, thereby facilitating external validation and model generalization. In pediatric patients, Bu is administered as a continuous 2-h intravenous infusion, and the longer infusion tubing leads to a delay in the time to reach peak concentration (Neroutsos et al., 2021). We studied and compared the blood sampling schemes of other LSS models. To better capture the C_max value, we designated three blood sampling time points at 120, 135, and 150 min post-administration. The results of our study indicate that the majority of patients had the highest blood concentration 135 min after administration. In addition, significant fluctuations in blood concentrations around C_max were evident. If peak concentration is only measured at two time points (120 and 150 min), the maximum peak concentration may be underestimated. In the present study, blood samples for Bu were collected from a separate venous fluid pathway to limit the influence of drugs on the fluid pathway. Therefore, the design of the blood sampling scheme is particularly important for accurately calculating the AUC (Utano et al., 2021; Teitelbaum et al., 2020; Watanabe et al., 2015). Multiple linear regression (MLR) is a statistical method of regression analysis used to analyze and model the linear relationship between a dependent variable and multiple independent variables. This approach helps assess the effect of each independent variable on the dependent variable and allows for predictions. The application of MLR models to predict drug AUC is becoming increasingly popular (Li et al., 2022; Xiang et al., 2021).

Theoretically, drug concentrations may fluctuate during the infusion period due to dynamic changes in absorption and distribution. After the infusion is complete, drug concentrations may stabilize or begin to decline. The calculation of CL is based on the elimination phase of the drug, which can more accurately reflect the drug’s clearance characteristics. However, our established MLR model showed that the best-adjusted r ² values in the equations included the independent variable C₆₀, suggesting that the blood concentration halfway through infusion may have some predictive value for subsequent changes in blood concentration. Therefore, our data processing experience suggests that including samples collected during the infusion period may enhance the predictive accuracy of MLR equations.

The Jackknife method, also known as “leave-one-out cross-validation,” assesses the stability of a statistic on data subsets by repeatedly excluding each observation and recalculating the statistic. This is useful for robustness analysis of parameter estimations in MLR models. Bootstrapping is a nonparametric method that estimates the sampling distribution of a statistic by resampling with replacements from the original data. This method allows for a more accurate estimation of model performance indicators (such as regression coefficients) because it considers variability in the sample data. Using both the Jackknife and Bootstrap methods for the internal validation of MLR models can provide a more comprehensive, robust, and precise model assessment, especially when dealing with small samples, non-normally distributed data, and complex data characteristics. The combined use of these methods can enhance the reliability and effectiveness of model validation. The ICC assesses consistency by calculating correlation coefficients, which are suitable for evaluating the consistency of multiple raters or multiple measurements, while BA plots are useful in assessing consistency by comparing the average differences and limits of agreement between two measurement methods, suitable for evaluating the consistency between two different measurement techniques or raters. In consistency evaluation, combining both methods for joint assessment can effectively avoid the limitations of a single-method evaluation, providing more representative results.

In this study, we conducted an external validation of the model using three patients and found good consistency with the internal validation results, which also met the clinical requirements for AUC accuracy. Subsequently, all children undergoing HSCT with Bu received AUC_0-360 predictions using model 10 with four blood sampling points, which reduced manpower and patient costs while providing effective personalized medication data. Nevertheless, our study has certain limitations. Because we used a more intensive blood sampling scheme that included times covering the absorption, distribution, metabolism, and elimination phases of the drug, the number of children enrolled in the study, sample size for external validation, and the proportion of infants and young children were all relatively small. Moreover, our data collection was limited to a specific region, which may restrict the generalizability of our results. Additionally, owing to funding and time constraints, we did not conduct long-term follow-up studies to provide clinical outcomes for children based on different AUC levels of Bu. We will address these issues in subsequent research.

In conclusion, this study employed MLR to establish a LSS model, and the results indicated that when the number of sampling points is 2–4, it can accurately predict the AUC_0-360 values of Bu. The model was internally validated using both the Jackknife and Bootstrap methods, and methodological consistency was evaluated using the ICC and BA methods. When the number of sampling points was 2–4, the accuracy and stability of the model improved with an increase in the number of sampling points. From the external validation data, it was observed that the predictive performance was better when the number of sampling points was 3–4; thus, it is recommended to use either three or four blood sampling points for prediction in practical applications. In summary, the LSS model, as a mathematical-pharmacological method with strong operability, demonstrates good clinical application value for TDM of Bu in patients undergoing HSCT.