Influence of Bayesian optimization on the performance of propofol target-controlled infusion

Background. Target controlled infusion (TCI) systems use population-based pharmacokinetic (PK) models that do not take into account inter-individual residual variation. This study compares the bias and inaccuracy of a population-based vs a personalized TCI propofol titration using Bayesian adaptation. Haemodynamic and hypnotic stability, and the prediction probability of alternative PK models, was studied.

Methods. A double-blinded, prospective randomized controlled trial of 120 subjects undergoing cardiac surgery was conducted. Blood samples were obtained at 10, 35, 50, 65, 75 and 120min and analysed using a point-of-care propofol blood analyser. Bayesian adaptation of the PK model was applied at 60min in the intervention group. Median (Absolute) Performance Error (Md(A) PE) was used to evaluate the difference between bias and inaccuracy of the models. Haemodynamic (mean arterial pressure [MAP], heart rate) and hypnotic (bispectral index [BIS]) stability was studied. The predictive performance of four alternative propofol PK models was studied.

Results. MdPE and MdAPE did not differ between groups during the pre-adjustment period (control group: 6.3% and 16%; intervention group: 5.4% and 18%). MdPE differed in the post-adjustment period (12% vs. -0.3%), but MdAPE did not (18% vs. 15%). No difference in heart rate, MAP or BIS was found. Compared with the other models, the Eleveld propofol PK model (patients) showed the best prediction performance. Conclusions. When an accurate population-based PK model was used for propofol TCI, Bayesian adaption of the model improved bias but not precision.