Pieter J. Colin (1), Douglas J. Eleveld (1) and Alison H. Thomson(2)
(1) Department of Anesthesiology, University Medical Center Groningen, University of Groningen, Groningen, The Netherlands. (2) Strathclyde Institute of Pharmacy and Biomedical Sciences, University of Strathclyde, Glasgow, UK.
Objectives: Genetic algorithms are a search heuristic that have been explored in the context of pharmacokinetic pharmacodynamic model selection [1,2], the optimisation of sampling times for PK studies [3] and as alternative structural models to the multi-compartment mammillary models in a machine learning approach to PKPD [4]. They are also well suited to derive combinations of discrete doses and dosing intervals that make up a dosing guideline. Especially when the number of possible combinations rule out the calculation of all possible outcomes. GAs also allow for different constraints to be imposed on the optimization process to safeguard the clinical feasibility of the dosing guideline. In this work we explore the use of a GA for the optimisation of a dosing guideline for intermittent vancomycin administration in adult patients. As a starting point for the optimization we used the Scottish Antimicrobial Prescribing Group vancomycin guideline, which is currently under review.
Methods: A GA was written in R using packages “tidyverse” and “deSolve”. In addition to the SAPG guideline the initial population comprised of 199 candidate solutions. Each solution consisted of 6 loading doses and 9 maintenance doses and dosing intervals. Loading doses were specified for different weight classes and maintenance doses were for different kidney-function classes (the different weight / kidney function class cut-offs were not optimized in this study). Pharmacokinetics were simulated for 10,000 virtual patients according to a recently published pooled population PK model [5]. The fitness of the different solutions, used to drive the selection of the optimal dosing guideline, was defined as the cumulative proportion of patients attaining an AUC0-24h between 400 mg.h.L-1 and 600 mg.h.L-1 throughout the first 72h of therapy. The calculation of the fitness was split up according to BMI, age and estimated creatinine clearance to avoid the optimization being driven by the most populated subgroup. Tournament selection was used to select parent solutions for crossover and mutation. Elitism, 5-fold crossover and pointwise mutation rates were applied with a probability of 1%, 80% and 5%, respectively. The GA was run for 100 generations.
Results: The average fitness for the starting point of the optimization was 1.244 compared to 1.352 for the final optimal solution. Daily target attainment was increased from 0.336, 0.400 and 0.411 to 0.492, 0.445 and 0.432 on days 1, 2 and 3, respectively. The optimal solution shows that loading doses should be increased, irrespective of patient body weight. Also, daily maintenance doses (mg q24h) should be increased for patients with an estimated creatinine clearance below 50 mL.min-1. The increase in target attainment was consistent across the different subgroups of patients in the virtual population. In addition, the optimal guideline shows less variable target attainment across subgroups compared to the initial guideline (10th and 90th percentiles being 0.413 and 0.481 and 0.308 and 0.456 for day 2 respectively).
Conclusions: We have shown that a GA informed by clinical trial simulations is a useful tool to optimize dosing guidelines. The approach allows to impose practical constraints on the optimization process (not addressed in this abstract) and weighting of the fitness across subgroups of patients, leading to robust optimal dosing guidelines which are suitable for implementation in clinical practice. A salient feature of algorithm-based optimization is that the different steps in the process have to be formalized, e.g. choice of patient subgroups, number of dose strengths, PKPD target, …., thereby increasing transparency in the decision making process. The presented approach could improve efficiency for companies embracing the modelling and simulation centred approach to drug development by moving away from M&S-based trial-and-error type optimizations of dosing guidelines.
References:
[1] Sale, M., and Eric A.S. A genetic algorithm based global search strategy for population pharmacokinetic/pharmacodynamic model selection. British journal of clinical pharmacology 79.1 (2015): 28-39.
[2] Bies, R. R., Muldoon, M. F., Pollock, B. G., Manuck, S., Smith, G., & Sale, M. E. A genetic algorithm-based, hybrid machine learning approach to model selection. Journal of pharmacokinetics and pharmacodynamics (2006): 33(2), 195.
[3] Hughes, J.H., Upton R.N., Reuter S.E., Phelps M.A., Foster D.J.R Optimising time samples for determining area under the curve of pharmacokinetic data using non‐compartmental analysis. Journal of Pharmacy and Pharmacology (2019); 71: 1635-44.
[4] Li, Q., Tao H., Wang J., Zhou Q., Chen J., Qin W.Z., Dong L., Fu B., Hou J.L., Chen J., Zhang W.-H.. Warfarin maintenance dose Prediction for Patients undergoing heart valve replacement—a hybrid model with genetic algorithm and Back-Propagation neural network. Scientific reports 8.1 (2018): 9712.
[5] Colin, P. J., Allegaert, K., Thomson, A. H., Touw, D. J., Dolton, M., de Hoog, M., Roberts J. A., Adane E. D., Yamamoto M., Santos-Buelga D., Martín-Suarez A., Simon N., Taccone F. S., Lo Y.L., Barcia E., Struys M.M.R.F., Eleveld D. J. Vancomycin pharmacokinetics throughout life: results from a pooled population analysis and evaluation of current dosing recommendations. Clinical pharmacokinetics (2019): 58(6), 767-780.
Reference: PAGE () Abstr 9251 [www.page-meeting.org/?abstract=9251]
Poster: Oral: Methodology - New Tools