IV-030 Sebastian Tandar

Population-level treatment optimization using a multi-objective optimization strategy

Sebastian T. Tandar (1), Linda B.S. Aulin (1,2), J. G. Coen van Hasselt (1)

1. Leiden Academic Centre for Drug Research, Leiden University, Leiden, The Netherlands 2. Centre for Human Drug Research, Leiden, The Netherlands

Objectives:

Optimizing the design of drug treatment schedules can be complex process as it may requires simultaneous consideration of multiple treatment objectives, such as efficacy, toxicity, and drug resistance1. In the past, treatment optimization relied on trial-and-error methods. The advent of pharmacokinetic-pharmacodynamic (PKPD) models significantly enhance the efficiency of this process by enabling treatment evaluation in silico. Nevertheless, the reliance on trial-and-error limits the ability to identify the optimal treatment design2. Thus, while trial-and-error approaches have been successful in identifying fit-for-purpose treatments3, 4, there remains a chance of overlooking an alternative treatment schedule with a potentially improved outcome2, especially when multiple drugs and therapeutic goals must be considered simultaneously.

 

Recently, efforts were made to adopt a more systematic approach to optimize drug treatment schedules2, 5-8. One example proposed a method derived from control theory to optimize treatment schedule for individuals or a typical individual as a surrogate for its corresponding population2. While these development represent important improvements, the potential benefit of optimizing dosing schedule for a population by considering inter-individual variability (IIV) remains to be further explored.

Population attainment of PKPD target(s) has been used as a key metric to guide treatment optimization for a population, even though such approach often relied on a single treatment objective. Such metric takes IIV into consideration and represents the fraction of the population that is expected to reach the treatment target. The metric is analogous to probability of target attainment that is often used to assess antibiotic treatment efficacy9. Here, we propose a novel strategy for dose schedule optimization at the population level by integrating multiple PKPD targets such as efficacy and toxicity limits as optimization objectives. To demonstrate the approach, we optimized ciprofloxacin (CIP)-tobramycin (TOB) combination antibiotic treatment against Pseudomonas aeruginosa infections in intensive care unit (ICU) patients, optimizing for a combination of efficacy, and prevention of toxicity and drug resistance. The combination was chosen based on the previously identified TOB hypersensitivity in CIP-resistant bacteria, which could suppress the emergence of drug resistance10.

Methods:

Dosing optimization framework

Candidate treatment schedules were evaluated based on population attainment of a predefined set of treatment objectives through stochastic simulation (n=1000). For each simulated individual, the cost function of an objective was calculated using a logistic function. Individual cost value was the weighted sum of each component treatment objectives. Here, equal weighing was used across all treatment objectives. The population cost value was the mean of individual cost values. Treatment optimization was conducted using a hybrid of particle swarm optimization (PSO) and limited-memory BFGS with boundary (L-BFGS-B) algorithms. PSO was used to maximize initial exploration of the search space, followed by L-BFGS-B for subsequent local refinement11.

Case study: CIP-TOB dosing optimization

A PKPD model was defined to simulate the progression of bacterial infection during antibiotic treatment. The PD component of the model included compartments accounting for the growth, antibiotic response, and resistance development of sensitive, CIP-resistant, and TOB-resistant bacterial subpopulations that may coexist during an infection. Parameters for the PD model were estimated from experimental data. The PK model was based on previously published population PK models for CIP12 and TOB13 in ICU patients.  The PKPD targets for the treatment were as follows: 1) achieving at least a 3-log10 reduction in endpoint bacterial cell densit[HJv(1] y, 2) maintaining the maximum bacterial cell density below 10^8 cells/mL, 3) capping daily CIP and TOB doses below standard care levels, and 4) keeping CIP [HJv(2] and TOB concentrations below their toxicity limits. The outcome of the population-optimized treatment was compared to that optimized using only the characteristics of the typical individual.

Results: 

In the current use case, optimization with only L-BFGS-B algorithm appeared to be sensitive to local optima, as evidenced by its inconsistent outcomes when different starting points were used. Although the use of PSO tends to slow down the overall process, the algorithm was found to be the most effective for improving initial search space exploration over genetic algorithm and simulated annealing.

In the case study, the treatment schedule optimized at the population level was able to meet all treatment objectives in 94.5% of the simulated individuals. PKPD target attainment rates for endpoint bacterial density, maximum bacterial cell density, and maximum drug [HJv(1] concentrations between 97.8% and 99.2%, which was achieved while utilizing less than half of their respective maximum daily doses used in standard monotherapy. For comparison, the standard CIP and TOB monotherapy were expected to attain the target endpoint bacterial density in 81.0% and 90.5% of the population, respectively.

Treatment optimization based on the typical individual was performed to evaluate the importance of implementing IIV in the optimization process. Although the resulting dosing schedule could meet all treatment objectives for the typical individual, this treatment schedule was only expected to achieve the treatment objectives in 78.2% of the simulated population when IIV was applied. This underscores the potential benefit of implementing IIV during population treatment optimization.

Conclusions:

This study introduces an optimization strategy to optimize dosing schedule at a population level by targeting multiple PKPD targets simultaneously. In the case study, we evaluated the approach through its application in optimizing the dosing schedule of a combination antibiotic treatment, addressing multiple treatment [HJv(1] objectives encompassing efficacy, toxicity, and drug resistance suppression. While control theory-based methods excel in scenarios requiring precise target values, the current approach offers flexibility to incorporate multiple treatment objectives and to target a range of satisfactory outcome. This makes the current method suitable for future studies aiming to optimize treatment outcomes where a range of acceptable target value is preferred over an exact target value. Moreover, our method can be modified to accommodate comprehensive weight assignments – such as those derived from multi-criteria decision analysis – to tune the optimization process to a more clinically pragmatic solutions14.  To summarize, the presented method offers a versatile framework adaptable for various applications by incorporating drug- and disease-specific targets.

References:

  1. Ribba, B., Bräm, D. S., Baverel, P. G., & Peck, R. W. (2022). Model enhanced reinforcement learning to enable precision dosing: A theoretical case study with dosing of propofol. CPT: pharmacometrics & systems pharmacology, 11(11), 1497–1510. https://doi.org/10.1002/psp4.12858
  2. Bachmann, F., Koch, G., Bauer, R. J., Steffens, B., Szinnai, G., Pfister, M., & Schropp, J. (2023). Computing optimal drug dosing with OptiDose: implementation in NONMEM. Journal of pharmacokinetics and pharmacodynamics, 50(3), 173–188. https://doi.org/10.1007/s10928-022-09840-w
  3. Wang, C., Zhang, C., Li, X., Zhao, S., He, N., Zhai, S., & Ge, Q. (2021). Dose Optimization of Vancomycin for Critically Ill Patients Undergoing CVVH: A Prospective Population PK/PD Analysis. Antibiotics (Basel, Switzerland), 10(11), 1392. https://doi.org/10.3390/antibiotics10111392
  4. Lee, S. M., Yang, S., Kang, S., & Chang, M. J. (2021). Population pharmacokinetics and dose optimization of vancomycin in neonates. Scientific reports, 11(1), 6168. https://doi.org/10.1038/s41598-021-85529-3
  5. Guo, T., van Hest, R. M., Zwep, L. B., Roggeveen, L. F., Fleuren, L. M., Bosman, R. J., van der Voort, P. H. J., Girbes, A. R. J., Mathot, R. A. A., Elbers, P. W. G., & van Hasselt, J. G. C. (2020). Optimizing Predictive Performance of Bayesian Forecasting for Vancomycin Concentration in Intensive Care Patients. Pharmaceutical research, 37(9), 171. https://doi.org/10.1007/s11095-020-02908-7
  6. De Carlo, A., Tosca, E. M., Fantozzi, M., & Magni, P. (2024). Reinforcement Learning and PK-PD Models Integration to Personalize the Adaptive Dosing Protocol of Erdafitinib in Patients with Metastatic Urothelial Carcinoma. Clinical pharmacology and therapeutics, 10.1002/cpt.3176. Advance online publication. https://doi.org/10.1002/cpt.3176
  7. Luo, W., Tan, X., & Shen, J. (2023). Optimal treatment strategy for cancer based on mathematical modeling and impulse control theory. Axioms, 12(10), 916. https://doi.org/10.3390/axioms12100916
  8. Aghaee, M., Ledzewicz, U., Robbins, M., Bezman, N., Jay Cho, H., & Moore, H. (2023). Determining optimal combination regimens for patients with multiple myeloma. European journal of pharmaceutical sciences : official journal of the European Federation for Pharmaceutical Sciences, 187, 106492. https://doi.org/10.1016/j.ejps.2023.106492
  9. Selig, D. J., Kress, A. T., Nadeau, R. J., & DeLuca, J. P. (2023). Beta-Lactam Probability of Target Attainment Success: Cefepime as a Case Study. Antibiotics (Basel, Switzerland), 12(3), 444. https://doi.org/10.3390/antibiotics12030444
  10. Hernando-Amado, S., Laborda, P., Valverde, J. R., & Martínez, J. L. (2022). Mutational background influences P. aeruginosa ciprofloxacin resistance evolution but preserves collateral sensitivity robustness. Proceedings of the National Academy of Sciences of the United States of America, 119(15), e2109370119. https://doi.org/10.1073/pnas.2109370119
  11. Shutao Li, Mingkui Tan, Tsang, I. W., & Kwok, J. T. (2011). A Hybrid PSO-BFGS Strategy for Global Optimization of Multimodal Functions. IEEE transactions on systems, man, and cybernetics. Part B, Cybernetics : a publication of the IEEE Systems, Man, and Cybernetics Society, 41(4), 1003–1014. https://doi.org/10.1109/TSMCB.2010.2103055
  12. Xie, F., Wang, Y., Peng, Y., Cheng, Z., & Li, S. (2021). Pharmacokinetic/pharmacodynamic evaluation of tobramycin dosing in critically ill patients: the Hartford nomogram does not fit. The Journal of antimicrobial chemotherapy, 76(9), 2335–2341. https://doi.org/10.1093/jac/dkab164
  13. Guo, T., Abdulla, A., Koch, B. C. P., van Hasselt, J. G. C., Endeman, H., Schouten, J. A., Elbers, P. W. G., Brüggemann, R. J. M., van Hest, R. M., & Dutch Antibiotic PK/PD Collaborators (2022). Pooled Population Pharmacokinetic Analysis for Exploring Ciprofloxacin Pharmacokinetic Variability in Intensive Care Patients. Clinical pharmacokinetics, 61(6), 869–879. https://doi.org/10.1007/s40262-022-01114-5
  14. Bellanti, F., van Wijk, R. C., Danhof, M., & Della Pasqua, O. (2015). Integration of PKPD relationships into benefit-risk analysis. British journal of clinical pharmacology, 80(5), 979–991. https://doi.org/10.1111/bcp.12674

Reference: PAGE 32 (2024) Abstr 11158 [www.page-meeting.org/?abstract=11158]

Poster: Methodology - New Modelling Approaches

PDF poster / presentation (click to open)