Corinna Maier (1,2), Niklas Hartung (1), Charlotte Kloft (3) and Wilhelm Huisinga (1)
(1) Institute of Mathematics, University of Potsdam, Germany (2) Graduate Research Training Program PharMetrX: Pharmacometrics & Computational Disease Modelling, Freie Universitaet Berlin and University of Potsdam, Germany (3) Department of Clinical Pharmacy and Biochemistry, Institute of Pharmacy, Freie Universitaet Berlin, Germany
Objectives:
A major aspect of therapeutic drug or biomarker monitoring (TDM) is to combine monitoring data with prior knowledge for model-based predictions of individualised therapy. These predictions are used to identify patients at risk for toxic or subtherapeutic concentrations and to give recommendations for subsequent dose adaptations. Bayesian forecasting tools typically only use the most probable model parameters for predicting therapy outcome without quantifying the associated prediction uncertainties, which include essential information about the probabilities of possible outcomes. A particularly critical example in which TDM could be beneficial is cytotoxic chemotherapy with neutropenia as the most frequent dose-limiting side effect. Bone marrow toxicity can expose patients to life-threatening infections, but also serves as surrogate for efficacy. It is therefore appealing to use neutrophil counts as a biomarker to guide individualised dosing [1]. For neutrophil guided dose selection, the neutropenia time-course for clinically possible dosing regimens must be predicted based on patient-specific neutrophil counts.
Methods:
Bayesian forecasting typically uses maximum a-posteriori (MAP) estimation, which determines the mode of the posterior distribution, i.e. the most probable parameters given patient-specific data. Such an analysis was e.g. performed in some prior work to control neutropenia [2]. In our study, data assimilation (DA) methods are presented which enable a comprehensive uncertainty quantification by deriving the full posterior distribution, namely the Sampling Importance Resampling (SIR) algorithm, the Metropolis-Hastings (MH) algorithm and the particle filter (PF). We systematically compare the properties of these fully Bayesian approaches with a local (normal) approximation and MAP estimation in terms of the accuracy of point estimates, the quality of uncertainty quantification and computational efficiency. We also investigate the benefits of deriving the full posterior distribution for informed decision support in the context of chemotherapy-induced neutropenia. For these purposes, we performed simulation studies in which we combined prior knowledge from clinical studies for the anticancer drugs docetaxel [3] and paclitaxel [4] with simulated patient data.
Results:
We identified that MAP estimation has several limitations for model-informed precision dosing, since it provides only a point estimate without quantification of associated uncertainties. In addition, the MAP parameter estimate depends on the parametrisation of the model, e.g. log-transformation, and the predicted outcome does not necessarily correspond to the most probable outcome. We also observed that a local (normal) approximation to the shape of the posterior distribution could yield misleading results, e.g. underestimation of the risk of severe neutropenia, which can have serious consequences and is highly undesirable in a high-risk environment. The fully Bayesian approaches provided increased accuracy of point estimates compared to MAP estimation and a comprehensive uncertainty quantification. The full posterior distribution also enables probabilistic statements about all possible outcomes for a patient, e.g. the probability of all neutropenia grades 0-4 or of possible recovery days. The particle filter additionally provides a sequential framework for efficient processing of monitoring data during ongoing treatment: Data does notneed to be stored for future analysis and the implementation of inter-occasion variability is facilitated as the information of previous occasions is already encoded in the current posterior distribution.
Conclusions:
Fully Bayesian methods offer crucial advantages over MAP estimation for TDM. They provide more accurate point estimates, a comprehensive uncertainty quantification and probabilistic statements about different possible outcomes, all of which are important aspects for decision makingin clinical care. Dose selection can then be based on the simultaneously available risk of toxic as well as of subtherapeutic regimes in order to find an effective as well as safe dose for the individual patient. Therefore, fully Bayesian approaches have the potential to improve patient care in various therapeutic areas. As new digital monitoring devices enable the frequent and non-invasive collection of patient data during treatment, sequential approaches to TDM will offer attractive opportunities in future.
References:
[1] Di Maio et al., Nat. Rev. Clin. Oncol. 2006
[2] Netterberg et al.,Cancer Chemother Pharmacol. 2017
[3] Kloft et al., Clin Cancer Res. 2006
[4] Henrich et al.,J Pharmacol Exp Ther. 2017
Reference: PAGE 28 (2019) Abstr 8843 [www.page-meeting.org/?abstract=8843]
Poster: Methodology - Estimation Methods