Predicting individual phenytoin dosage.

Most previously suggested methods for predicting phenytoin dosage from steady-state drug levels (Cpss) measured in the clinical setting fail to fully exploit all relevant (population) information. A bayesian prediction method, applicable to any drug, is available. It appropriately combines all types of information. In this paper, we compare the Bayesian method as applied to phenytoin to two other prediction methods (and a baseline, nonfeedback one). Actual doses are compared to predictions in 49 patients. Each method is optimized, as far as possible, for the test data. The comparison favors the Bayesian method. Since each of the other prediction methods for phenytoin can be shown to be a theoretically suboptimal special case of the Bayesian one, the superiority of the latter may be a general phenomenon. Because the pharmacokinetic model linking steady-state phenytoin levels and dosage is so simple, a good approximation of the general Bayesian method can be implemented as a graphical device, or as a program for a programmable calculator. We present and describe both of these approximations.