Predictions of carbamazepine concentrations using a Bayesian program (PKS System, Abbott): a retrospective evaluation...

This work evaluates the performance of a Bayesian program (PKS System, Abbott) for predicting carbamazepine concentrations in an outpatient population. The retrospective study involved 20 epileptic patients (12 adults and 8 children) receiving carbamazepine monotherapy orally. The program was used to predict measured serum levels after feedback of 0, 1 or 2 steady-state concentrations. A significant negative prediction bias was observed when no feedback concentration was used for estimation. However, the prediction bias (mean prediction error; m.e.) decreased as soon as one feedback concentration was used for estimation. Precision (mean absolute prediction error; m.a.e.) was significantly improved with one feedback concentration and was even better with two concentrations. Likewise, r.m.s.e. (root mean squared error; composite of bias and precision) regularly decreased when the number of feedback concentrations used was increased. Eleven percent of the estimates were unacceptable clinically (prediction error > 2 mg L-1) when 1 feedback concentration was used; less than 3% were unacceptable when two concentrations were used. Thus the performance of the Bayesian dosing program is acceptable when two feedback concentrations are known, and seems able to help the clinician adjust carbamazepine dosage in an outpatient population.