Marina Cuquerella-Gilabert (1,2,3,4), Sergio Sánchez Herrero (3), Matilde Merino Sanjuan (1,2), Elena Tosca (4), Victor Mangas Sanjuan (1,2), Paolo Magni (4)
(1) Department of Pharmacy and Pharmaceutical Technology and Parasitology, School of Pharmacy, University of Valencia, Valencia, Spain; (2) Interuniversity Research Institute for Molecular Recognition and Technological Development, 46100 Burjassot, Valencia, Spain; (3) Simulation Department, Empresarios Agrupados Internacional S.A., Madrid, Spain; (4) Department of Electrical, Computer and Biomedical Engineering, University of Pavia, Via Ferrata 5, I-27100 Pavia, Italy
Objectives: Therapeutic drug monitoring represents one of the main applications of pharmacometrics in the clinical setting. Bayesian approaches are routinely performed for characterizing individual pharmacokinetic (PK) parameters in order to optimize the benefit/risk balance of a patient. Nonetheless, in real-world clinical settings, strict adherence to predetermined sampling times is frequently unfeasible, and patient sampling may not always be completed. In these instances, it would need to be considered the variations in sampling procedures to ensure that experimental evidence provides informative parameter estimates, enabling the adequate analysis of the forthcoming data. The main aims of this project are: (i) to assess the impact of sampling procedure variation on parameter estimation, and (ii) to evaluate the estimation robustness of different estimation algorithms in all scenarios tested.
Methods: A semi-mechanistic PBPK model (Phys-DAT), developed previously with PhysPK, software for a single oral 40mg citalopram administration was taken as reference [1]. Three estimation algorithms in Python were selected: Powell, a derivative-free method seeking minima using conjugate directions; Nelder-Mead and BGFS, two optimization algorithms that search space without gradients applying each one different approximations. Each estimation method was tested versus six pre-established scenarios, which resulted from the combination of type of sampling (rich or sparse) and three types of individuals (p50th, p5th, p95th). Seven or three randomly selected samples from 0 to 25 h after drug administration were considered for the rich and sparse conditions, respectively. The typical (p50th) and non-typical individuals (p5th and p95th) were obtained from a virtual population of 1.000 individuals taking into account the inter-individual variability on Vd, CL and Peff. For each scenario, 10 PK profiles were generated from a different combination of random samples. Mean Prediction Error (MPE) and coefficient of variation (CV%) of prediction errors were considered for assessing the estimation performance of each algorithm in terms of accuracy and precision, respectively.
Results: The overall performance for characterizing the individual PK parameters (Vd and CL) of a typical individual with rich sampling (7 samples) provided very similar and outstanding results (MPE: 0.973, 0.993, 0.986; 1.001, 1.001, 0.997; and CV (%): 14.78, 20.57, 0.01; 0.08, 0.08, 0.02) across the three estimation algorithms in Python (Powell, Nelder-Mead and BGFS). When assessing the same individual (typical) with sparse sampling (3 samples), differently from the adequate estimations obtained for CL a less accurate estimation of Vd was observed with the Nelder-Mead algorithm (MPE: 0.86). Regarding the non-typical individuals (p5th and p95th), biased (MPE range 0.38 – 2.44) and less precise (CV (%) range 0.05-58.25) estimation of individual PK parameters were obtained independently of the Python algorithm and the sampling conditions (rich or sparse). All the scenarios tested for these non-typical individuals provided MPE larger than the two-fold error (0.5-2), except for the Powell algorithm, which provided MPE within the two-fold error when rich sampling occurred.
Conclusions: These results suggest that sampling richness and individual profile type are two key factors for the accurate and precise estimation of individual PK parameters in clinical practice. According to the results obtained so far, the Powell estimation algorithm provided a more accurate estimation of PK parameters, but the three algorithms tested apparently failed characterizing the individual PK profiles for non-typical patients, irrespectively of the numbers of samples collected.
References:
[1] Cuquerella-Gilabert M, Reig-López J, Serna J, Rueda-Ferreiro A, Merino-Sanjuan M, Mangas-Sanjuan V, Sánchez-Herrero S. Phys-DAT: A physiologically-based pharmacokinetic model for unraveling the dissolution, transit and absorption processes using PhysPK®. Comput Methods Programs Biomed. 2024 Jan;243:107929. doi: 10.1016/j.cmpb.2023.107929. Epub 2023 Nov 18. PMID: 38006685.
Reference: PAGE 32 (2024) Abstr 11006 [www.page-meeting.org/?abstract=11006]
Poster: Methodology - Estimation Methods