The purpose determines the predictive performance: Comparison of four population pharmacokinetic models of methotrexate.
Anna-Karin Hamberg (1,2) and Sebastian G. Wicha (3)
(1) Dept. of Medical Sciences, Uppsala University and (2) Dept. of Clinical Pharmacology, Uppsala University Hospital, Sweden, (3) Dept. of Clinical Pharmacy, Institute of Pharmacy, University of Hamburg, Germany
Objectives: High-dose methotrexate (HD-MTX) is a highly effective anticancer treatment in cancer patients. In order to support therapeutic drug monitoring of MTX by model-based techniques, we aimed to compare the predictive performance of four published MTX models [1–4] on a clinical dataset on patients with lymphoid malignancies.
Methods: The clinical dataset comprised 12 patients with lymphoid malignancy, contributing 397 MTX samples over up to 8 dosing occasions. Different prediction settings were explored in NONMEM (version 7.3): (i) population prediction using solely the patient covariates, (ii) individual prediction using all available data, (iii) prediction of the remaining dosing occasion using the first sample, (iv) prediction of the second occasion from the first occasion, (v) prediction of the third occasion and (vi) the fourth from the previous two or three occasions. Graphical measures as well as relative bias (rBias) and relative root mean squared error (rRMSE) were used to judge the predictive performance.
Results: In scenarios (i) and (ii), model [2] provided lowest rBias (+23%, -7%) and rRMSE (103%, 43%). For the within-occasion prediction scenario (iii), model [1] was superior (rBias: -51%, rRMSE: 260%) over the other models (abs. rBias: 140%-1714%, rRMSE: 425%-3648%). For the across-occasion prediction scenarios (iv), model [4] provided the least biased prediction of the subsequent occasion (rBias: +2.5%, rRMSE: 149%) compared with the other models (abs. rBias: 16%-31%, rRMSE: 90%-121%). Qualitatively similar results were observed in scenario (v). When the fourth occasion was predicted form the previous data (vi), model [1] was superior (rBias: -66%, rRMSE: 98%) over the other models (abs. rBias: 75%-197%, rRMSE: 113-287%).
Conclusion: The present study highlights the fact that a models predictive performance depends on the use of the model. Good predictive performance in conventionally evaluated scenarios (i) and (ii) may not translate to the same predictive quality if the models are used in the setting of therapeutic drug monitoring, when prediction of future observations or subsequent occasions is warranted.
References:
[1] M. Joerger, A.J.M. Ferreri, S. Krähenbühl, J.H.M. Schellens, T. Cerny, E. Zucca, A.D.R. Huitema. Dosing algorithm to target a predefined AUC in patients with primary central nervous system lymphoma receiving high dose methotrexate. Br. J. Clin. Pharmacol., 73.: 240–247 (2012).
[2] Å.M. Johansson, N. Hill, M. Perisoglou, J. Whelan, M.O. Karlsson, J.F. Standing. A Population Pharmacokinetic/Pharmacodynamic Model of Methotrexate and Mucositis Scores in Osteosarcoma. Ther. Drug Monit., 33.: 711–718 (2011).
[3] A. Nader, N. Zahran, A. Alshammaa, H. Altaweel, N. Kassem, K.J. Wilby. Population Pharmacokinetics of Intravenous Methotrexate in Patients with Hematological Malignancies: Utilization of Routine Clinical Monitoring Parameters. Eur. J. Drug Metab. Pharmacokinet.,: 1–8 (2016).
[4] E. Dombrowsky, B. Jayaraman, M. Narayan, J.S. Barrett. Evaluating performance of a decision support system to improve methotrexate pharmacotherapy in children and young adults with cancer. Ther. Drug Monit., 33.: 99–107 (2011).