External Evaluation of Risperidone Population Pharmacokinetic Models Using Opportunistic Pediatric Data

Abstract

Risperidone is approved to treat schizophrenia in adolescents and autistic disorder and bipolar mania in children and adolescents. It is also used off-label in younger children for various psychiatric disorders. Several population pharmacokinetic models of risperidone and 9-OH-risperidone have been published. The objectives of this study were to assess whether opportunistically collected pediatric data can be used to evaluate risperidone population pharmacokinetic models externally and to identify a robust model for precision dosing in children. A total of 103 concentrations of risperidone and 112 concentrations of 9-OH-risperidone, collected from 62 pediatric patients (0.16-16.8 years of age), were used in the present study. The predictive performance of five published population pharmacokinetic models (four joint parent-metabolite models and one parent only) was assessed for accuracy and precision of the predictions using statistical criteria, goodness of fit plots, prediction-corrected visual predictive checks (pcVPCs), and normalized prediction distribution errors (NPDEs). The tested models produced similarly precise predictions (Root Mean Square Error [RMSE]) ranging from 0.021 to 0.027 nmol/ml for risperidone and 0.053-0.065 nmol/ml for 9-OH-risperidone). However, one of the models (a one-compartment mixture model with clearance estimated for three subpopulations) developed with a rich dataset presented fewer biases (Mean Percent Error [MPE, %] of 1.0% vs. 101.4, 146.9, 260.4, and 292.4%) for risperidone. In contrast, a model developed with fewer data and a more similar population to the one used for the external evaluation presented fewer biases for 9-OH-risperidone (MPE: 17% vs. 69.9, 47.8, and 82.9%). None of the models evaluated seemed to be generalizable to the population used in this analysis. All the models had a modest predictive performance, potentially suggesting that sources of inter-individual variability were not entirely captured and that opportunistic data from a highly heterogeneous population are likely not the most appropriate data to evaluate risperidone models externally.

Keywords: pediatrics; pharmacokinetics; population modeling; precision dosing; risperidone.

FIGURE 1

Comparison of quantitative measures of bias and precision among the models included in the external evaluation analysis. Model (A): Kloosterboer et al., 2021; Model (B): Sherwin et al., 2012; Model (C): Feng et al., 2008; Model (D): Feng et al., 2008 with allometric scaling; and Model E: Thyssen et al., 2010. Precision was evaluated using the mean prediction error (PE) and the root mean square error (RMSE). Bias were evaluated using the mean percent error (MPE) and the mean absolute percent error (MAPE).

FIGURE 2

Population predicted concentrations versus observations for risperidone. Model (A): Kloosterboer et al., 2021; Model (B): Sherwin et al., 2012; Model (C): Feng et al., 2008; Model (D): Feng et al., 2008 with allometric scaling; and Model (E): Thyssen et al., 2010. The dashed black and dashed red lines represent the line of identity and the least-squares regression curve, respectively.

FIGURE 3

Population predicted concentrations versus observations for the metabolite, 9-OH risperidone. Model (A): Kloosterboer et al., 2021; Model (B): Sherwin et al., 2012; Model (C): Feng et al., 2008; and Model (D): Feng et al., 2008 with allometric scaling. The dashed black and dashed red lines represent the line of identity and the least-squares regression curve, respectively.

FIGURE 4

Conditional weighted residuals (CWRES) versus population predictions for risperidone plotted on a log scale. Model (A): Kloosterboer et al., 2021; Model (B): Sherwin et al., 2012; Model (C): Feng et al., 2008; Model (D): Feng et al., 2008 with allometric scaling; and Model (E): Thyssen et al., 2010. The dashed black line corresponds to a CWRES of zero. The solid grey lines correspond to CWRES values of 2 and −2. The dashed red line corresponds to the locally-weighted scatterplot smoothing curve (LOWESS).

FIGURE 5

Conditional weighted residuals versus populations predictions for 9-OH risperidone plotted on a log scale. Model (A): Kloosterboer et al., 2021; Model (B): Sherwin et al., 2012; Model (C): Feng et al., 2008; and Model (D): Feng et al., 2008 with allometric scaling. The dashed black line corresponds to a CWRES of zero. The solid grey lines correspond to CWRES values of 2 and -2. The dashed red line corresponds to the locally-weighted scatterplot smoothing curve (LOWESS).

FIGURE 6

Prediction-corrected visual predictive checks (pcVPCs) of the observed data overlaid on the predictions obtained by performing 1,000 simulations with each risperidone population pharmacokinetic model. Model (A): Kloosterboer et al., 2021; Model (B): Sherwin et al., 2012; Model (C): Feng et al., 2008; Model (D): Feng et al., 2008 with allometric scaling; and Model (E): Thyssen et al., 2010. All pcVPC plots are based on the time after the first dose. The dashed lines represent the 5th, 50th, and 95th percentiles for the observed data, and the gray shaded regions are the 95% prediction interval for the predicted concentrations. The red stars indicate outlying percentiles of the observed data from the prediction interval. The y axis is in log-transformed scale. The x axis represents the time after first recorded dose. A sample that was collected later than 1,000 h after the first recorded dose was omitted from the graphs to improve visualization. The point was within the prediction interval for all of the models tested except for Model A.

FIGURE 7

Prediction-corrected visual predictive checks (pcVPCs) where the observed data are overlaid on the predictions obtained by performing 1,000 simulations with each model for 9-OH risperidone. Model (A): Kloosterboer et al., 2021; Model (B): Sherwin et al., 2012; Model (C): Feng et al., 2008; and Model (D): Feng et al., 2008 with allometric scaling. All pcVPC plots are based on the time after the first dose. The dashed lines are the 5th, 50th, and 95th percentiles for the observed data, and the gray shaded regions represent the 95% prediction interval for the predicted concentrations. The red stars indicate outlying percentiles of the observed data from the prediction interval. The x axis represents the time after first recorded dose. A sample that was collected later than 1,000 h after the first recorded dose was omitted from the graphs to improve visualization. The point was within the prediction interval for all of the models tested.