Evaluation of the predictive performance of four pharmacokinetic models for propofol.

BACKGROUND: This study has compared the predictive performance of four pharmacokinetic models, two of which are currently incorporated in commercial target-controlled infusion pumps for the administration of propofol.
METHODS: Arterial propofol concentrations and patient characteristic data were available from nine patients who, in a published study, had received a standardized infusion of propofol. Predicted concentrations with 'Diprifusor' (Marsh), 'Schnider', 'Schuttler', and 'White' models were obtained by computer simulation. The predictive performance of each model was assessed overall and over the following phases: rapid infusion (1-5 min), early (1-21 min), maintenance (21-min end-infusion), and recovery (2-20 min post-infusion).
RESULTS: The overall assessment, based on 29-36 samples from each patient, indicated that all four models were clinically acceptable. However, the negligible bias (-0.1%) with the 'Schnider' model was accompanied by overprediction in the rapid infusion phase and underprediction during recovery. This changing bias over time was not detected as 'divergence' when assessed on absolute performance error (APE), (1.4% h(-1)) but became significant (13.2% h(-1)) when based on changes in signed PE over time. The 'Schuttler' model performed well at most phases but overpredicted concentrations during recovery. The White model led to a marginal improvement over 'Diprifusor' and would be expected to reduce the positive bias usually seen with 'Diprifusor' systems.
CONCLUSIONS: In assessing the predictive performance of pharmacokinetic models, additional information can be obtained by analysis of bias at different phases of an infusion. The evaluation of divergence should involve linear regression analysis of both absolute and signed PEs.