Bayesian regression analysis of non-steady-state phenytoin concentrations: evaluation of predictive performance

Michaelis-Menten saturable pharmacokinetics confound the determination of appropriate phenytoin maintenance doses. This study retrospectively evaluated the performance of an IBM-PC/XT computer program applying Bayesian regression to the "explicit solution to the Michaelis-Menten equation." Zero to five non-steady-state phenytoin serum concentrations were used to predict either non-steady-state concentrations at least 10 days in the future (n = 49) or steady-state concentrations (n = 20). Non-steady-state concentration prediction precision (% mean absolute error) using 0-5 non-steady-state feedbacks was 137%, 62%, 39%, 31%, 25%, and 15%, respectively, and steady-state concentration prediction precision was 446%, 47%, 50%, 44%, 21%, and 13%, respectively. Elimination of subjects receiving concurrent drugs known to induce phenytoin metabolism significantly improved predictions based on population priors; however, performance improvements were not apparent after two serum level feedbacks. The program provided clinically acceptable predictions with four or more feedbacks. Refinement of population parameters and optimal sampling times should further improve performance.