Abstract
Vincristine is a cornerstone of pediatric chemotherapy, but its use is limited by vincristine-induced peripheral neuropathy (VIPN). Paradoxically, African children tolerate higher vincristine doses with minimal neurotoxicity, raising questions about the pharmacological mechanisms and model generalizability across populations. In this study, we re-estimated vincristine pharmacokinetic (PK) and PK-pharmacodynamic (PKPD) models using data from both Dutch and Kenyan pediatric cohorts and designed five simulation scenarios to assess the impact of Type I (informative censoring) and Type II (population-specific effect modifier) selection bias on model predictions. A three-compartment PK model with saturable distribution and a proportional-odds PKPD model with Markov elements jointly described vincristine disposition and VIPN risk. Kenyan children showed lower clearance but markedly reduced PD sensitivity, resulting in negligible predicted VIPN even at higher doses. Incorporating informative censoring improved internal validity by capturing the observed dropout dynamics, while mechanistic extrapolation using CYP3A5, ABCB1, and CEP72 genotype distributions increased external validity and aligned model predictions better with empirical outcomes. Clinically, these findings support population-specific vincristine dosing strategies and suggest that reduced neurotoxicity in African children reflects lower neuronal susceptibility rather than reduced systemic (plasma) exposure. Methodologically, this work demonstrates how unrecognized selection bias can distort PKPD extrapolations across populations and highlights the value of integrating mechanistic and genetic information within model-based frameworks to improve the safety and external validity of model-informed dosing strategies beyond the original study populations.