Comparison of four single-point phenytoin dosage prediction techniques using computer-simulated pharmacokinetic values.

The predictive abilities of the following four single-point phenytoin dosage adjustment methods were compared using computer-simulated data: the Bayesian feedback method of Vozeh et al. (B), a linearized version of the Bayesian method (LB), the population-clearance method of Graves et al. (G), and the Rambeck nomogram (R). A series of 512 "subjects" with normally distributed values for volume of distribution, weight, and the Michaelis-Menten variables Vmax and Km were simulated. The steady-state serum concentration (SSSC) resulting from the administration of a standard dose of phenytoin sodium (5 mg/kg/day) was calculated, and "subjects" with SSSCs less than or equal to 12 mg/L or greater than or equal to 17 mg/L were entered in the study. If the concentration was greater than 50 mg/L or the standard dosage exceeded Vmax, the dosage was reduced empirically by 25%. Normally distributed random errors were introduced into the SSSC values to simulate actual patient data. The pharmacokinetic values, dosages, and SSSCs were used for predicting the dosage required to attain an SSSC of 14.9 mg/L. In the unstratified population, the mean error and mean-squared error were lowest for methods G and B, followed by methods LB and R. Methods B and LB gave the highest percentages of satisfactory dosage predictions based on the resultant SSSC value. The performance of all methods was superior at initial SSSCs greater than 8 mg/L.(ABSTRACT TRUNCATED AT 250 WORDS)