Equation 3
\begin{equation} \text{MADPE}_{i}(\%)=median\ \{|\text{PE}_{\text{ij}}|,\ j=1,\ldots,\ N_{i}\}\nonumber \\ \end{equation}
To go further in the evaluation of these models, the predictive performance of Bayesian forecasting was assessed according to Broeckeret al. (12). Calculations of MDPE and MDAPE enable the comparison of different model-predicted and observed concentrations.
Fractionated databases were created to assess the impact of observed tobramycin concentrations on the predictive performance: (a) complete database, (b) removal of all observed peaks, (c) removal of all residuals, (d) removal of all dosages after Day3, (e) removal of all observed concentration (a priori prediction). (c) and (d) were applied only for patients with more than 2 doses.
An MDPE between –20% and 20% and an MDAPE ≤ 30% are considered acceptable criteria for bias and inaccuracy (13).
The predictive performance was further evaluated by simulating 500 tobramycin concentrations per time point using model PK parameters as fix and the $SIM of NONMEM®. To assess the forecasting error between the observed and the simulated dataset, the 5th, 50th and 95th percentile of the simulated data over the time dose (Predicted Corrected Visual Predictive Check [pcVPC]) was plotted. With the same simulated dataset, the distribution of the Normalized Prediction Distribution Error (NDPE) was also plotted. Normal distribution of NPDEs were checked visually by using the histogram of NPDEs frequency.