You are here

Fast Parameter Identification for Personalized Pharmacokinetics



OBJECTIVE: To develop a novel and fast computing scheme for constructing personalized pharmacokinetic models. The scheme must rely on (i) a limited set of measurements for each individual patient and (ii) a knowledge base of existing well-calibrated models for a collection of diverse individual in order to approximate in silico the structure of metabolic interactions for a given individual patient by solving a parameter identification problem. 

DESCRIPTION: The field of pharmacokinetics (PK) is dedicated to the study of the concentration dynamics of substances administered to a living organism. In population PK, models of variability among a population receiving clinically relevant doses of a substance of interest are constructed as a function of observable traits such as demographic, pathophysiological, and therapeutic-relevant features, such as body weight, excretory and metabolic functions. These models are used to understand how certain traits in the population affect the dynamics of metabolic interactions. Individual-specific (or personalized) PK models have a different use as they can assist in designing personalized therapies so as to improve effectiveness and to avoid severe side-effects given individual patient characteristics in drug response (see for example, [2] and [3]). However, developing individual specific PK models is a demanding task that may require extensive in vivo sampling or intensive computational effort because the parameter identification problem is underdetermined (see e.g. [4]). In silico approaches may provide a more cost-efficient way of developing personalized PK models. For example, a PK model for a given individual maybe constructed from a knowledge base of well-calibrated models for a set of representative individuals in a population. Using a limited set of measurements obtained from a new individual (pursuant a controlled dose of the substance of interest) one may aim to find the combination (or ensemble) of existing models that best fits the data. The latter task is tantamount to solving a stochastic optimization problem in which the objective is to maximize a measure of model fit by choosing parameter values that are a convex combination of the parameter values in the collection of well-calibrated individual models. This process can be seen as constructing an avatar, i.e. an artificial representation of the new individual as a function of the existing knowledge base. Nonetheless, this stochastic optimization problem is quite challenging from a computational standpoint. First, evaluating the fit of a given choice of parameters may require non-negligible computational time. Several model runs (for the same parameter combination) must be obtained in order to reduce the effects of noise and thus obtain an adequate estimation of the measure of fit. Running this kind of model repeatedly is an extremely time consuming task. Secondly, the model simulation error is not necessarily well-behaved, i.e. it may not be zero-mean and may also exhibit significant correlation. Finally, since the underlying model is akin to a blackbox, the optimization problem associated to model identification is not necessarily convex. The proposed work should result in a relatively simple to use app that could guide the personalizing of drug therapies with military applications such as severe bleeding control and PTSD treatment. 

PHASE I: The STTR performer must conceive, implement and test a new in silico approach to constructing personalized PK models. The proposed design must jointly use a limited set of measurements for a given individual patient and a knowledge base of existing well-calibrated models for a collection of diverse individuals in a population sample. For phase I, the knowledge base will be simulated but in Phase II, the STTR performer must develop the testbed and gather a data set. Personalized model identification must be done by finding the combination of parameter values in the collection of well-calibrated individual models that maximizes model fit with respected the limited set of measurements. Implementation and testing of the proposed scheme must be conducted with a well-known PK model (e.g. glucose metabolism). Given that running PK models could be extremely time consuming, the proposed scheme must be distributed and cloud-based or cluster-based. In this computing infrastructure distributed stochastic gradient algorithms could be implemented in order to speed up model identification. A prototype computing scheme for constructing personalized PK models must be developed in the form of a software library that will be application agnostic and flexible enough to accommodate alternate in-silico technologies. Thus, after selecting a PK/PD application area, the overall applicability of the technology must be demonstrated by adapting the algorithmic scheme to a different type of PK model. In light of the fact that modeling error could be significant and not necessarily well-behaved, a comprehensive methodology for assessing of the robustness of the proposed computational scheme must be proposed. 

PHASE II: The first task in this phase consists of developing an experimental testbed needed to construct a knowledge base of PK models. This knowledge base must correspond to representative population of adult individuals from which extensive database of measurements will be gathered. Data collection must be done in an unobtrusive manner avoiding burdens, distractions, or alterations to participants typical lifestyle. Thus, the data collected would reflect typical human behavior. The data must be recorded in such a manner that subjects cannot be identified, directly or through identifiers linked to the subjects. In addition to numerical accuracy of the fit, standard clinical metrics should be used to evaluate the quality of the models by comparing fitted model to datasets in order to understand how clinically relevant are the numerical inaccuracies obtained during the fitting process. A second task in this phase, consists of analyzing the performance of the scheme as it relates to bias and/or correlation in modeling error. In case modeling performance is greatly affected, a suite of alternative variations to the computing scheme proposed must be identified and duly supported with empirical evidence. Finally, the last task consists of developing apps that would make use of the personalized PK models obtained. For example, a fully personalized model can be used to educate patients on how to manage their particular condition more effectively. This could be done for instance by (i) creating (or repurposing) a phone app to record relevant information, (ii) communicating recorded information to a cloud service for computing the personalized PK model, and (iii) developing training modules based on the personalized PK model for the use of practitioners and patients. Other example apps could be used by physicians seeking to understand better (individualized) dosing strategies. A system can take the form of a web-services physicians can access to collect data and submit information to a central facility for model identification, use the identified model to suggest therapeutic use and improved dosing to physicians (see for example [1]). 

PHASE III: In Phase III, the STTR performer's software will be available for military and civilian use. The FDA has recently accepted in silico trials as supporting evidence for approving new drugs and/or medical devices. In this context, the value of in-silico clinical studies is directly related to the quality and consistency of data used to generate and test these models. We expect the STTR performer will lay out the foundations for obtaining FDA approval for potential future applications of the software. For example, the STTR performer will strictly rely on standardized clinical trial data (in compliance with FDA). We envision that the team that develops the software will market it for Government laboratory use, and negotiate commercial licensing with commercial and academic markets. As an alternative, any or all of these artifacts might be released into the open source community through organizations such as the Open Source Electronic Health Record Alliance (OSEHRA) or Open Health IT Tools or similar organizations for open sources licensing. Based on negotiations with the types of government and commercial organizations cited, it is possible that hybrid commercial and open source licensing could occur. In the case where these artifacts are released into the open source community, the STTR awardee would need to develop and provide a plan to state how it would sell additional consulting, software implementation and/or training services around their workflow model, technical implementation guidelines, and/or software controls. 


1: Patek D. Lv D., Campos-Nañez E. and Breton M. (2016) Retrospective Optimization of Daily Insulin Therapy Parameters. Proceedings of 11th IFAC Symposium on Dynamics and Control of Process Systems Trondheim, Norway.

2: Konagaya A. Towards an In Silico Approach to Personalized Pharmacokinetics (2012) in Molecular Interactions A. Meghea, Ed. In Tech. pp. 263-282.

3: Kovatchev B., Breton M., Dalla Man C. and Cobelli C. (2009) In Silico Preclinical Trials: A Proof of Concept in Closed-loop Control of Type I Diabetes. Journal of Diabetes Science and Technology. Vol. 3, pp. 44-55.

4: Aoki Y., Hayami K., De Sterck H., Konagaya A. (2014). Cluster Newton Method for Sampling Multiple Solutions of Underdetermined Inverse Problems: Application to a Parameter Identification Problem in Pharmacokinetics, SIAM Journal of Scientific Computing, Vol. 36 No. 1, pp. 14-44.


KEYWORDS: Pharmacokinetics, Model Identification, Precision Medicine, Stochastic Optimization 

US Flag An Official Website of the United States Government