Development of covariate-informed model pooling method to predict the correct starting dose
Bram C. Agema (1, 2), Stijn L.W. Koolen (1, 2), Ron H.J. Mathijssen (1), Brenda C.M. de Winter (2,3), Birgit C.P. Koch (2,3) , Sebastiaan D.T. Sassen (2,3)
(1) Dept. of Medical Oncology, Erasmus MC Cancer Institute, Erasmus University Medical Center; Rotterdam, The Netherlands (2) Dept. of Hospital Pharmacy, Erasmus University Medical Center; Rotterdam, The Netherlands, (3) Rotterdam Clinical Pharmacometrics Group; Rotterdam, The Netherlands
Objectives:
An increasing number of population pharmacokinetic (POP-PK) models have been developed which can predict plasma concentrations based solely on patient characteristics. However, the a priori predictive value of these models is mostly low due to a large margin of error and a high degree of inter-individual variability. In addition, it is difficult to determine which models perform best in the intended target population. To counter this problem, we developed an algorithm that pools different POP-PK models tailored to patient characteristics which we call covariate-informed model pooling (CIMP). Our aim is to decrease the proportion of patients with plasma levels outside the therapeutic window by giving a predicted starting dose. We used imatinib as a case study for this analysis.
Methods:
A retrospective study was performed in 62 patients who were treated with imatinib for gastro-intestinal stromal tumors (GIST) at the Erasmus Medical Center. Steady-state imatinib plasma concentrations, laboratory values and patient characteristics were collected.
17 models describing imatinib PK were identified in literature. Five models were excluded due to missing covariate data (α1-glycoprotein levels and genotyping) or insufficient information on model structure. The remaining 12 models were subsequently used to generate model-predicted imatinib concentrations for each patient. Patients were thereafter stratified in tertiles using the covariates incorporated in the models (age, weight, albumin, haemoglobin and white blood cell count). For each tertile of patients and for each covariate the median relative bias and the root mean square error (RMSE) were determined per model. Each model was subsequently scored and given an importance score based on the relative bias and RMSE and a penalty to be determined in sensitivity analyses.
Thereafter, each tertile of each covariate was also given a score based on the amount of predictions/observations outside of the 80% - 125% interval, [NH1] [BA2] divided by the total amount of predictions/observations outside the interval for each covariate. This was multiplied with the prediction of each model and the score for the model.
Using this prediction, sensitivity analyses for both the penalty as well as the interval were performed. These were evaluated visually using density plots and numerically with the relative bias and RMSE of this new prediction.
After obtaining the final predictions using CIMP, the predicted plasma concentrations were extrapolated towards a steady state trough time using the half-life[NH3] [BA4] , as is current practice in our hospital. Using these predicted extrapolated trough concentrations, a dose recommendation was obtained with steady-state trough levels between1100 – 2200 ng/mL as target [1]. As no upper target limit is known for GIST patients in literature, we doubled the threshold and used this as an upper limit. When a patient was predicted to be under- or overexposed, the algorithm would augment or decrease the dose in steps of 100 mg to increase exposure. Using the extrapolated observed concentrations the result of the dosing advices were simulated
The methodology was validated using data from 100 patients treated in the intensive care unit that received vancomycin.
Results:
The sensitivity analyses showed that the penalty for bias and RMSE was best set to 500. In addition, no relevant influence of changing the 80 – 125% interval was observed.
Using the CIMP dose recommendations, the amount of patients outside the therapeutic window of 1100 - 2200 ng/mL was reduced by 36.4% (from 53.2% to 33.9%). When only using the best model the amount of patients outside the therapeutic window was 48.4% and when using the equal importance score for each model, this was reduced to 43.5%.
The CIMP prediction resulted in a relative bias of -4.9% and a relative RMSE of 38.8%. This was better compared to pooling all models equally (bias: -3.1%, RMSE: 45.2%) or using the best model after an external evaluation (bias: 10.1%, RMSE: 54.5%).
When tested in the vancomycin cohort, the CIMP methodology showed similar results.
Conclusions:
We showed that CIMP is a feasible approach for predicting drug plasma concentrations before sampling, using all available PK models including a weighting factor, which outperforms more conventional methods. When these predictions are used to perform dose recommendations, the amount of patients below the threshold decreased by 34.6%.
Yu H, Steeghs N, Nijenhuis CM, Schellens JHM, Beijnen JH, Huitema ADR. Practical guidelines for therapeutic drug monitoring of anticancer tyrosine kinase inhibitors: Focus on the pharmacokinetic targets. Clin Pharmacokinet. 2014;53(4):305–25.