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:
One of the major side effects of cytotoxic anticancer treatment is neutropenia, a severe reduction of neutrophils that puts patients at high risk of life-threatening infections. Reliable predictions of chemotherapy-induced neutropenia can help to early identify patients at risk as well as patients at subtherapeutic doses. Predicting the neutrophil time course can also support individualised dosing schedules. Novel measurement devices in digital health care, allowing for frequent neutrophil monitoring at home, require the application of new recursive data processing methods that enable decision support in ongoing treatment. Recursive data processing is well established in the field of meteorology in the form of sequential data assimilation (DA) methods that are used to improve model-based weather forecasts as new data becomes available. The objective of this study was to investigate the application of sequential DA techniques in the field of systems pharmacology. Based on the example of neutropenia, induced by the cytotoxic anticancer drugs docetaxel or paclitaxel, the benefits of sequential DA methods should be examined.
Methods:
In contrast to batch processing of data, sequential DA methods provide a framework to recursively process data [1]. This results in iterative cycles of forecasts based on mathematical models and combining these computer-generated forecasts with measurement data in real time. In these cycles the posterior is recursively updated based on Bayes’ formula. For applications in systems pharmacology, particle filter algorithms are particularly suitable, as they are sequential DA methods that allow for non-Gaussian error models and nonlinear structural models. The particle filter distribution approximates the current posterior distribution, which integrates all prior information up to the current data point.
For the simulation study, we used prior knowledge from a population analysis [2] of a clinical study for initialising the particle filter. The forecast was generated using the gold-standard model for neutropenia [3], for the update patient-specific data was simulated and the updated posterior functioned as prior for the next update step. For the same setting a maximum a-posteriori (MAP) estimation was performed in prior work [5], which allows for comparison of results. MAP estimation, the current state-of-the-art in systems pharmacology, provides access to the mode of the posterior distribution and is, in its basic form, non-recursive. To account for long-term effects over several treatment cycles, we used an extended neutropenia model [4] that describes bone marrow exhaustion for the anticancer drug paclitaxel.
Results:
Based on the mean of the posterior distribution as point estimate for the particle filter, both approaches have comparable root mean squared errors (RMSEs) in the parameter estimates and in the summary statistics, e.g. time to recovery or grade of neutropenia. In contrast to MAP estimation, the particle filter also quantifies the uncertainty of these statistics by estimating the full posterior. This leads to much more informative results as the probabilities for all possible outcomes are provided and not only the most likely one. Additionally, in this recursive framework, only the current data point is needed for the update step as all prior information is already included in the previous posterior distribution. The complexity of the optimization problem in the MAP estimation, however, increases with the number of measurements. This makes the particle filter much more efficient for multiple cycle therapy and suitable for real-time implementation, which allows to support decision-makers in ongoing treatment.
Conclusions:
Using neutropenia as example, we demonstrated that sequential DA methods provide an efficient framework to recursively process patient data. With the development of novel digital healthcare devices this is becoming more and more important in systems pharmacology. In particular, we showed that the particle filter enables patient-specific predictions about the time-course of neutrophil counts, which help to identify patients at risk for super- or subtherapeutic doses and support adaptations for subsequent treatment cycles. The comprehensive uncertainty quantification, sequential data processing and easy-to-interpret results are crucial for rational and individual decision-making in oncology.
References:
[1] Reich and Cotter, CUP 2015
[2] Kloft et al., Clin Cancer Res. 2006
[3] Friberg et al., J Clin Oncol. 2002
[4] Henrich et al., J Pharmacol Exp Ther. 2017
[5] Netterberg et al.,Cancer Chemother Pharmacol. 2017
Reference: PAGE 27 (2018) Abstr 8432 [www.page-meeting.org/?abstract=8432]
Poster: Methodology - Estimation Methods