Published Population Pharmacokinetic Models of Imatinib Perform Poorly on TDM Data from Pediatric Patients

BACKGROUND: Population pharmacokinetic models can potentially provide suggestions for an initial dose and the magnitude of dose adjustment during therapeutic drug monitoring procedures of imatinib. Several population pharmacokinetic models for imatinib have been developed over the last two decades. However, their predictive performance is still unknown when extrapolated to different populations, especially children.

OBJECTIVE: This study aimed to evaluate the predictive performance of these published models on an external real-world dataset containing data from both adults and children.

METHODS: A real-world dataset was collected, containing observations from adult and pediatric patients with Philadelphia chromosome-positive/Philadelphia chromosome-like acute lymphoblastic leukemia and chronic myeloid leukemia (N = 39) treated with imatinib. A systematic review through PubMed was conducted to identify qualified population-pharmacokinetic models for external evaluation (i.e., prediction-based, simulation-based, and Bayesian forecasting diagnostics). Standard allometric scaling was used for models that were developed based on data from adults only.

RESULTS: Fifteen published models were found for evaluation, of which only two were based on data from both children and adults. Prediction-based diagnostics showed that some models had an acceptable level of bias. The model by Shriyan et al. (with allometric scaling) performed best with a median prediction error of 1.24%. However, no models performed well on precision even when allometric scaling was used, where the lowest median absolute prediction error was 37.66% using the model by Schmidli et al. The models by Golabchifar et al. and Schmidli et al. (both with allometric scaling) performed the best of all tested models, with a median prediction error ≤ 15%, median absolute prediction error ≤ 40%, fraction of prediction error within ± 20% (F20) ≥ 0.3, and within ± 30% (F30) nearly 0.4. Simulation-based diagnostics showed that most of the observations outside the 90% prediction interval were from children. Bayesian forecasting showed that the model prediction could be improved using one prior sample, particularly in adults.

CONCLUSIONS: Current models fail to accurately predict imatinib plasma concentrations in our real-world dataset, especially for children. Future pharmacokinetic studies should focus on developing better models for pediatric populations.