An Algorithm to Generate Virtual Patients Cohort for a Model of Solid Tumor Treatment
Borisov I.(1), Metelkin E.(1), Kolesova G.(1), Demin O.(1)
(1) InSysBio CY
Introduction: Conventional QSP modelling approach implies fitting a model to series of mean data values, thus obtaining parameters’ values, which describe a certain “reference patient”. Variability of possible patient responses can be modeled by a cohort of “virtual patiens” (VP). The goal of generating such a cohort is to reproduce results of clinical trials, which are typically expressed in terms of summary statistics. A number of methods were proposed to generate VP cohorts [1] and address different aspects of this problem. However, the goal to reproduce different statistical measures by a VP cohort becomes more challenging if the cohort has to be simultaneously fitted to multiple statistical results of different therapies.
Objectives: The goal of this study is to propose an algorithm to efficiently generate a cohort of virtual patients, which meets statistical measures typically reported by a clinical trial protocol in the field of solid tumor treatment. The main focus of the algorithm is the case of multiple observed measures, which should be reproduced by a VP cohort for a number of therapies.
Methods: In case of oncology models for solid tumor, reported outcome statistics typically include percentage of complete and partial responders (CR, PR), percentage of patients with progressive disease (PD) and stable disease (SD), and median values for “time to event” observables, such as duration of response (DoR), time to response (TTR), time to progression (TTP), progression-free survival (PFS). The algorithm proposes a step-wise approach for generating a VP cohort and follows the general workflow proposed by Allen et al.[2]. At first plausible patients’ population is generated and divided into groups of outcomes for different therapies. Linear Programming algorithm finds fractions of patients from each group, which exactly meet percentage of responders for each therapy. Then patients are sampled according to the found fractions to fit statistics of location and spread. The computations were conducted in Julia language, which provides a framework to solve both optimization and sampling problems [3].
Results: The proposed algorithm allows the modeler to generate VP cohort, which fits reported summary statistics for a number of clinical trials The application of the algorithm is illustrated by a use-case of solid tumor model with a number of outcomes for each therapy. The performance of the algorithm is measured by how close it meets the experimentally reported statistical quantities. The results highlight the ability of algorithm to exactly meet percentage of responders for multiple therapies and set the tolerance level for fitting statistics of location and spread.
Conclusions: The proposed algorithm allows a modeler to generate a representative VP cohort, which can reproduce results of multiple clinical trials. The stages of the algorithm are illustrated by a solid tumor model use-case, but its application is not limited to the field of oncology and can be potentially extended to a broader range of modelling problems, where VP generation is on-demand.
References:
[1] Kolesova G, Stepanov A, Lebedeva G, Demin O. Application of different approaches to generate virtual patient populations for the quantitative systems pharmacology model of erythropoiesis. J Pharmacokinet Pharmacodyn. 2022 Oct;49(5):511-524. doi: 10.1007/s10928-022-09814-y.
[2] Allen RJ, Rieger TR, Musante CJ. Efficient Generation and Selection of Virtual Populations in Quantitative Systems Pharmacology Models. CPT Pharmacometrics Syst Pharmacol. 2016 Mar;5(3):140-6. doi: 10.1002/psp4.12063.
[3] Bezanson J, Karpinski S, Shah VB, Edelman A. Julia: A fast dynamic language for technical computing. arXiv preprint arXiv:12095145. 2012.