Nonparametric population modeling of valproate pharmacokinetics in epileptic patients using routine serum monitoring data: implications for dosage.

Therapeutic drug monitoring (TDM) of valproate (VAL) is important in the optimization of its therapy. The aim of the present work was to evaluate the ability of TDM using model-based, goal-oriented Bayesian adaptive control for help in planning, monitoring, and adjusting individualized VAL dosing regimens. USC*PACK software and routine TDM data were used to estimate population and individual pharmacokinetics of two commercially available VAL formulations in epileptic adult and pediatric patients on chronic VAL monotherapy. The population parameter values found were in agreement with values reported earlier. A statistically significant (P < 0.001) difference in median values of the absorption rate constant was found between enteric-coated and sustained-release VAL formulations. In our patients (aged 0.25-53 years), VAL clearance declined with age until adult values were reached at about age 10. Because of the large interindividual variability in PK behavior, the median population parameter values gave poor predictions of the observed VAL serum concentrations. In contrast, the Bayesian individualized models gave good predictions for all subjects in all populations. The Bayesian posterior individualized PK models were based on the population models described here and where most patients had two (a peak and a trough) measured serum concentrations. Repeated consultations and adjusted dosage regimens with some patients allowed us to evaluate any possible influence of dose-dependent VAL clearance on the precision of total VAL concentration predictions based on TDM data and the proposed population models. These nonparametric expectation maximization (NPEM) population models thus provide a useful tool for planning an initial dosage regimen of VAL to achieve desired target peak and trough serum concentration goals, coupled with TDM soon thereafter, as a peak-trough pair of serum concentrations, and Bayesian fitting to individualize the PK model for each patient. The nonparametric PK parameter distributions in these NPEM population models also permit their use by the new method of 'multiple model' dosage design, which allows the target goals to be achieved specifically with maximum precision. Software for both types of Bayesian adaptive control is now available to employ these population models in clinical practice.