Direct comparison of three methods for predicting digoxin concentrations

Three methods of determining digoxin population pharmacokinetic parameters were compared for their abilities to predict 118 measured serum digoxin concentrations (SDCs) in 49 patients. NONMEM software (version IV) was used to generate a residual and a weighted residual for each measured-concentration-predicted-concentration pair. Prediction error analysis was done by a maximum likelihood technique that accounted for several within-patient measures. Data analysis also included graphic observation of weighted residuals (WRES) and calculation of the mean WRES and median absolute prediction error. A further parallel analysis was also carried out on subpopulations with and without concurrent quinidine and congestive heart failure (CHF). Method III was without bias in all subpopulations studied and had the smallest WRES in all populations. Method I was without bias in the overall population, however, it underpredicted SDCs in patients receiving quinidine and in those with CHF. Method II underpredicted SDCs in the overall population, those receiving quinidine, and in patients without CHF. There were no between-method differences in precision as assessed by absolute prediction error.