Model predictive control with Bayesian updates (MPC) is more robust to model misspecification, compared to standard Bayesian control (sEBE) for Therapeutic Drug Management (TDM). Investigation in a cohort of 315 patients receiving tacrolimus during the first 14d after renal transplantation.

Faelens R (1), Luyckx N (2), Leirens Q (2), Kuypers D (3), Bouillon T(1)

(1) Drug Delivery and Disposition, Department of Pharmaceutical and Pharmacological Sciences Catholic University of Leuven, Belgium; (2) SGS Exprimo, Mechelen, Belgium; (3) Department of Nephrology, University Hospital Leuven, Belgium.

Objectives: Model qualification (transfer between populations), treatment of interoccasion variability (IOV) and downweighing of observations are relevant issues in TDM [1]. MPC, a technique derived from process control theory, could help relax the requirement for a “perfect” model: if the “states” of the model can be frequently updated, MPC is relatively robust against model misspecification. As a proof of concept for MPC in TDM, we evaluated predictive ability for tacrolimus concentrations between physicians, MPC and sEBE, using both a model trained on the evaluation dataset and a (misspecified) model trained on a different dataset.

Methods: Model building was performed independently on two datasets: (A) 100 patients with a rich profile of 10 samples over 12 h at d7 post-transplant [2], and (B) 315 patients with daily trough samples 0-14 days post-transplant [3]. As the intended use of the models was to inform TDM, no extensive covariate search was performed. Monolix2018R2 was used [4].

The models were implemented in TDMore, a framework for model based dose adaptation and simulation currently under development. Dataset B was used for prospective evaluation. sEBE was performed on all available concentration measurements at each respective time. MPC performs piece-wise estimation. For the first measurement, the individual parameters are obtained as with sEBE, using the population typical values and population IIV as prior. Predicted ODE compartment states at the observation time and corresponding individual parameters are then stored in memory. For the subsequent measurement, the prediction starts at the previously stored ODE compartment states. EBE for this measurement only is performed using the previous individual parameters and population IIV as prior, the latter being a heuristic decision to allow for sufficient flexibility during future updates.

Predictive performance was characterized as IPRED/DV, and summarized as fraction within a target range between 0.88 to 1.11 corresponding to the target range for tacrolimus at our institution immediately after transplantation (12-15mcg/L)). To evaluate physician-based dosing, it was assumed that physicians predict their chosen doses will hit the target of 13.5 mcg/L. Since (mis)prediction of future samples translates inversely into dosing decisions, an inference regarding dose adjustments can be made (underprediction -> overdosing and vice versa).

Results: Model building on Dataset A identified a 2-cpt model with lagged oral absorption (Model A). CYP3A5 and Weight were included as covariates. Model building on dataset B (trough levels only) identified a 1cpt model with saturable increase of the elimination rate over the observation period (Model B). This implies that Model A is misspecified, as it does not capture this trend. Prospective evaluation is summarized in Table 1 (d2, d4 and d10 after transplant).

Table 1: Probability of target attainment (PTA) and 95% binomial proportion confidence interval.

Model predictions outperform physician predictions in almost all cases, except using misspecified model A with sEBE. Since the procedure for MPC and sEBE does not differ for the first observation, target attainment is identical on day 2. On later days, sEBE and MPC perform equally well using model B. Using misspecified model A, MPC performs as well as when using model B, and outperforms sEBE.

Conclusions: These results are preliminary and require confirmation with simulated and historical datasets. However, in the case of perhaps inevitable model misspecification in real world situations, MPC is a viable alternative to sEBE. Specifically for tacrolimus during the first 14d after renal transplantation, MPC outperforms physicians regardless of model misspecification.

References:
[1] Keizer RJ, Ter Heine R, Frymoyer A, Lesko LJ, Mangat R, Goswami S. Model-Informed Precision Dosing at the Bedside: Scientific Challenges and Opportunities. CPT Pharmacometrics Syst Pharmacol. 2018; 12: 785-787.
[2] Kuypers DR, Claes K, Evenepoel P, Maes B, Coosemans W, Pirenne J, Vanrenterghem Y. Time-related clinical determinants of long-term tacrolimus pharmacokinetics in combination therapy with mycophenolic acid and corticosteroids: a prospective study in one hundred de novo renal transplant recipients. Clin Pharmacokinet. 2004; 43: 741-62.
[3] Vanhove T, Hasan M, Annaert P, Oswald S, Kuypers DRJ. Pretransplant 4β-hydroxycholesterol does not predict tacrolimus exposure or dose requirements during the first days after kidney transplantation. Br J Clin Pharmacol. 2017; 83: 2406-2415.
[4] Monolix version 2018R2. Antony, France: Lixoft SAS, 2018. http://lixoft.com/products/monolix/

Reference: PAGE 28 (2019) Abstr 9076. [www.page-meeting.org/?abstract=9076.]

Poster: Clinical Applications

PDF poster/presentation (click to open)