Arsenii Zats1,3, Andrew Brandon2, Tadao Tamura1,3, Monika Berezowska1, Shelby Barnett2, Gareth Veal2, Alwin Huitema4, Mirjam van de Velde5, Alaric Taylor-Roffey1,3, Jugal Suthar
1Vesynta, 2Translational and Clinical Research Institute, Newcastle University Centre for Cancer, 3University College London, 4Department of Pharmacology, Princess Máxima Center for Pediatric Oncology, 5 Emma Children’s Hospital, Pediatric Oncology, Amsterdam UMC
Introduction: Vincristine is a chemotherapeutic agent used to treat pediatric and adult cancers, including leukemia and lymphoma. Its narrow therapeutic index and high inter-individual pharmacokinetic (PK) variability complicate dose optimisation and increase toxicity risk [1]. Accurate population pharmacokinetic (PopPK) models are essential to predict vincristine concentrations and optimise dosing regimens. However, existing models trained on specific populations often perform inconsistently due to variabilities in model structure and inter-site procedural differences [2,3]. Model averaging addresses model structure uncertainty by combining weighted predictions from multiple models, while dataset augmentation improves model generalisability by integrating data from various sources [3,4]. The objective of this study was to explore model averaging and dataset augmentation for vincristine PopPK modelling and to compare their predictive performances against specialised single models on set-aside unexposed validation data. Methods: An augmented dataset was created by combining vincristine data from four studies, including Ewing’s sarcoma (n = 39, VIDE patients) [5], TDM (n = 102) [6], UKALL (n = 31) [7], and VINCA trial (n = 37) [8]. Data from the PINOCCHIO study (n = 56) [1] was not included in preliminary analysis. Study inconsistencies were addressed by standardising units, harmonising covariates, and handling missing data and outliers. Data was split using covariate stratification to ensure representativeness, assessed by the Kolmogorov-Smirnov test. This resulted in a training set (149 patients), an internal validation set (38 patients), and a reserved unexposed set (17 patients) for model performance comparison. A baseline model was selected from Nijstad et al., given its robust performance and comprehensive three-compartment structure incorporating tubulin binding and multiple dosing occasions [1]. The model was enhanced by adding inter-site variability (ISV) on the central (V1) and peripheral (V2) volumes of distribution and implementing the M3 method for handling data below the limit of quantification (BQL). Model parameters were estimated using NONMEM® 7.4 on the augmented dataset. The model was internally validated using bootstrap, SIR and VPC. Model averaging was implemented by combining predictions from transcribed 6 vincristine models Nijstad et.al. [1], Barnett et.al. [9], van de Velde et.al. [8], Moore et.al. [10], Guilhaumou et.al. [11], and Igarashi et.al. [12]), with each model’s contribution weighted inversely to its root mean square error (RMSE), following the methodology described by Uster et.al. [2]. Predictive performances of the enhanced augmented model, the model averaging approach, and individual models were evaluated using the reserved unexposed dataset. Metrics included relative bias (rBias), RMSE, and the percentage of predictions within 90-110% of the observed values. Results: Adding ISV and incorporating BLQ handling (M3 method) significantly improved the dataset-augmented model fit compared to the original Nijstad et.al. model. This adjustment lowered the objective function value (OFV) from 2715.20 to 2453.86, leading to reductions in Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) values. Overall, the dataset-augmented model and model averaging approach both substantially improved predictive accuracy compared to individual models. RMSE was 25.98% for model averaging and 26.24% for the dataset-augmented model. Other individual models had considerably higher RMSE values, ranging from 31.8% to 136.6%. Relative bias (rBias) was lower for model averaging (-0.77%) and dataset augmentation (-0.41%) compared to individual models, where rBias ranged from -12.5% to +38.7%. The percentage of predictions within 90-110% of observed values was 41.0% for model averaging and 43.6% for dataset augmentation, outperforming individual models, ranging from 12.8% to 24.3%. The composite ranking score (integrating RMSE, rBias, and prediction accuracy) was lowest for model averaging and dataset augmentation, indicating superior performance. Conclusion: Both dataset augmentation and model averaging demonstrated improvements upon the baseline model, in terms of accuracy, reducing error and bias. Both methods also outperformed the specialised models for our representative unexposed real world data. While their performance was similar, dataset augmentation provided more stable predictions with less variability, suggesting greater generalisability.
[1] Nijstad AL et al. A population pharmacokinetic modeling approach to vincristine in children. Pharm Res. 2022;39(10):2487-2495. doi: 10.1007/s11095-022-03364-1. [2] Uster DW et al. A model averaging approach improves predictive performance in model-informed precision dosing: vancomycin as a case study. Clin Pharmacol Ther. 2021;109:175-183. doi: 10.1002/cpt.2065. [3] Svensson E et al. Integration of data from multiple sources for nevirapine population pharmacokinetics. Br J Clin Pharmacol. 2012;74(3):465-76. doi: 10.1111/j.1365-2125.2012.04205.x. [4] Agema BC et al. Selecting the best pharmacokinetic models for a priori model-informed precision dosing with model ensembling. Clin Pharmacokinet. 2024;63:1449-1461. doi: 10.1007/s40262-024-01425-9. [5] Chemotherapy for the treatment of recurrent and primary refractory Ewing sarcoma. ISRCTN Registry. ISRCTN36453794. 2014. doi: 10.1186/ISRCTN36453794 . [6] Therapeutic drug monitoring in children with cancer. ISRCTN10139334. Available at: https://doi.org/10.1186/ISRCTN10139334 . [7] United Kingdom Trial for children and young adults with Acute Lymphoblastic Leukaemia and Lymphoma 2011. ISRCTN64515327. Available at: https://doi.org/10.1186/ISRCTN64515327 [8] van de Velde ME et al. Vincristine-induced peripheral neuropathy in pediatric oncology: The VINCA trial. Cancers (Basel). 2020;12(12):3745. doi: 10.3390/cancers12123745. [9] Barnett S et al. Vincristine dosing and therapeutic drug monitoring in neonates and infants. Eur J Cancer. 2022;164:127-136. doi: 10.1016/j.ejca.2021.09.014. [10] Moore AS et al. Vincristine pharmacodynamics and pharmacogenetics in children: a population modeling approach. J Paediatr Child Health. 2011;47:875-882. doi: 10.1111/j.1440-1754.2011.02103.x. [11] Guilhaumou R et al. Population pharmacokinetics and pharmacogenetics of vincristine in pediatric solid tumors. Cancer Chemother Pharmacol. 2011;68:1191-1198. doi: 10.1007/s00280-010-1541-4. [12] Igarashi T et al. Population pharmacokinetics and exposure-response analysis of vincristine in malignant lymphoma. Cancer Chemother Pharmacol. 2021;87(4):501-511. doi: 10.1007/s00280-020-04220-y.
Reference: PAGE 33 (2025) Abstr 11690 [www.page-meeting.org/?abstract=11690]
Poster: Drug/Disease Modelling - Paediatrics