Predictability of individualized dosage regimens of carbamazepine and valproate mono- and combination therapy.

WHAT IS KNOWN AND OBJECTIVE

Many investigators agree that appropriate rational utilization of therapeutic drug monitoring (TDM) with Bayesian feedback dosage adjustment facilitates epilepsy treatment with carbamazepine (CBZ) and/or valproate (VPA) by increasing the seizure control and safety, as well as by reducing treatment costs. In previous works we have developed and used in clinical practice population pharmacokinetic (PK) models of different dosage forms for VPA and post-induction CBZ behaviour, as well as for combined therapy with CBZ plus another 'old' antiepileptic drug (AED). An important step of external validation is to evaluate how well a procedure of Bayesian individualizing AED dosage regimens based on a proposed population PK model and sparse TDM data 'works', and how helpful it is in real practical clinical settings. The aim of this study was to evaluate the predictability of individualized dosage regimens for monotherapy with CBZ in the post-induction period or with VPA, as well as for CBZ and VPA given as combination therapy based on TDM data of epileptic patients and the earlier developed population models.

METHODS

Four groups of TDM data were analysed using the USC*PACK software for PK/PD analysis: 556 predictions for adult epileptic patients on CBZ monotherapy, 662 predictions for VPA monotherapy, 402 predictions of CBZ serum levels and 430 predictions of VPA serum levels for adult epileptic patients on CBZ+VPA combination therapy. Statistical characteristics of the prediction errors (PE) and weighted PE were used to estimate bias and precision of predictions. Intraindividual and interoccasional variability of predictions were also estimated.

RESULTS AND DISCUSSION

This study demonstrated that in most cases of CBZ and VPA monotherapy and combination therapy, predictions of future AED concentrations based on the earlier developed population PK models, TDM data and patient-specific maximum a posteriori probability Bayesian posterior parameter values provided clinically acceptable estimates. Statistical analysis of the residuals demonstrated that the distributions of residual and weighted residual were close to the normal distribution (Kolmogorov-Smirnov test, P > 0·05) and their mean values did not differ statistically significant from zero (no statistically significant bias, P > 0·05) for all groups of predictions. The observed decreased quality of predictions of VPA concentrations during VPA+CBZ combination therapy, especially when CBZ dosages were changed, might well be explained by their PK interactions. For all groups, in linear regression analysis, the observed trend of decreasing of the prediction quality over various future prediction time horizons was considered statistically significant (P < 0·05). Prediction of serum levels further in future was less precise than those closer to the present for a 1·5- to 3·5-year observation period. No bias in predictions was associated with the time horizons.

WHAT IS NEW AND CONCLUSION

Our validation results suggest good predictive performance of the population models developed earlier, and quite acceptable predictions of future AED serum levels for individualized dosage regimens of CBZ and VPA therapy in real clinical settings.