Model‐based dosing with concentration feedback as an integral part of personalized hydroxycarbamide management
Year: 2018
Type: letter
Abstract: We read with great interest the manuscript by Estepp and colleagues where the authors evaluated model-based strategies for predicting hydroxycarbamide (hydroxyurea) exposure in children with sickle cell anaemia (SCA) 1. The authors conclude that a Bayesian dosing approach using sparse sampling as proposed by our group 2 ‘did not measurably add to the prediction accuracy compared to various population-based dosing algorithms’, which is not consistent to findings reported in the literature 3-5. Therefore, we conducted a predictive performance analysis with similar Bayesian scenarios for external validation as described by Estepp et al. 1 Estepp et al. 1 suggested that the Bayesian approach (Scenarios 3–5) only slightly improved prediction accuracy (as shown in their Figure 3), even when all the concentration measurements were used as feedback (Scenario 5 with up to 12 measurements from each subject). They concluded that their dosing model using only body weight as a covariate (Scenario 1) and without inclusion of patient-specific PK data is as predictive as individualized dosing based on posterior Bayesian estimation. Bayesian feedback utilizes patient specific real-time concentration measurements to obtain the most accurate estimate of drug exposure for dose individualization 6. With detailed and patient-specific information used as feedback as in Scenario 5, the predictions should be largely informed by the individual data and be close to the noncompartmental analysis (NCA) estimates that were used for comparison. However, when the PK model assumes small inter-individual variability (IIV) of parameters, Bayesian parameter estimates will remain close to the population mean. Indeed, the IIV of parameter estimates used for Bayesian estimation by Estepp et al. 7 were unusually small as they reported – coefficients of variation (CVs) for oral clearance (CL/F) and volume of distribution (V/F) were 6.9% and 9.4% respectively, with other IIV values fixed to zero (as shown in 7 Table S3). In the same manuscript, the authors reported that the CV of hydroxycarbamide clearance analysed by the noncompartmental analysis was 15% (as shown in 7 Table 2). In addition, the variability of hydroxycarbamide clearance in other studies has been consistently reported in the range of 15%–30% 1, 2, 8, 9. A low variability in clearance in the population PK model (i.e. the Bayesian prior) would have caused significant shrinkage of individual estimates towards the population mean 10, 11 thus defeating the value of Bayesian estimation for accurate individual predictions. In our evaluation, we compared three population PK models: 1) the model described by Estepp et al. (Model 1) 1, 7; 2) a model similar to Model 1 but with a higher CV of 15% for both clearance and volume of distribution (Model 2); and 3) the population PK model by Dong et al. (Model 3) 2. As external data source for the validation, we used PK data from the HUSTLE trial (ClinicalTrials.gov NCT00305175) 9 since we did not have access to the PK data in the current study. Similar scenarios as described in the Estepp study 1 were tested. For Scenarios 1 & 3, concentrations at 1.5 h were simulated using the population PK model 2 given that the HUSTLE study did not collect a 1.5-h sample. For Scenario 4, we used a model predicted 3-h concentration (Scenario 4.1), or a 4-h concentrations as an alternative scenario (Scenario 4.2). In Figure 1A–C, we summarize how much of the variability in AUC (reported as R2) is explained with a regression equation (Scenario 1) and Models 1–3 (Scenarios 2–5). Contrary to what Estepp et al. 1 suggested in their Figure 3, feedback of concentration data does lower the variability and improve the precision of the predictions, and most significantly when all data points are used. As shown in Figure 1A, the correlations between predictions and observations are more tightly clustered around the identity line in Scenario 5 than they are in Scenario 1, and the coefficient of determination (R2) increased from 0.43 to 0.83, indicating an overall 93% increase in prediction accuracy. This is in line with Bayesian theory. Of note, our validation data set had seven PK sampling time points as the maximum per subject. When using the 12 concentrations per subject as in the Estepp study 1, we expect that predictions would have been even better. The coefficients of determination further improve when Models 2 and 3 are used as priors. Panel D shows box-and-whisker plots in a similar fashion as the data reported in Estepp et al. 1 summarizing the prediction errors in AUC estimates with the regression equation and Models 1–3. Prediction errors are largest with the equation and with the population model without concentration feedback, and decrease with each additional concentration feedback scenario. These results are very different from the predictive performance results presented in the manuscript by Estepp et al. 1 Our evaluation shows that Bayesian estimations incorporating patient-specific concentration measurements outperform the one sample regression equation. In addition, the analysis indicates that a sparse sampling strategy using 3 timed samples provides improved predictions compared to population predictions when a moderate level of IIV (15%) was applied. The coefficient of determination (R2) improved from 0.43 to 0.63 or 0.82 (Figure 1B). These results suggest that a dosing algorithm developed based on a population PK model (with body weight as the only covariate for the parameters) can generate a reasonable starting dose in most patients, but would require subsequent dose individualization using posterior Bayesian estimation to facilitate further tailoring of the hydroxycarbamide dose to precisely reach the concentration and exposure targets for each patient. In summary, we found significant improvements of predictive performance using Bayesian estimation when the population PK model assumes realistic variability estimates. We disagree with Estepp et al. 1 that the accuracy of population-based dosing without any patient-derived PK samples is sufficient to target AUCs in individual patients. Our analysis confirms that a model-based approach with sparse sampling feedback remains an attractive strategy for precision dosing of hydroxycarbamide in paediatric patients with SCA and is currently being prospectively evaluated 12, 13. Key protein targets and ligands in this article are hyperlinked to corresponding entries in http://www.guidetopharmacology.org, the common portal for data from the IUPHAR/BPS Guide to PHARMACOLOGY 14, and are permanently archived in the Concise Guide to PHARMACOLOGY 2017/18. All authors have completed the Unified Competing Interest form at www.icmje.org/coi_disclosure.pdf (available on request from the corresponding author) and declare: no support from any organization for the submitted work; no financial relationships with any organizations that might have an interest in the submitted work in the previous 3 years; no other relationships or activities that could appear to have influenced the submitted work. The HUSTLE trial was supported by R01-HL090941 from NHLBI (REW).
DOI: 10.1111/bcp.13585
Language: English
Field: Medicine
Domain: Health Sciences
Open Access status: green
Award: G674879958
OpenAlex ID: W2803085691