Using machine learning to optimize antibiotic combinations: dosing strategies for meropenem and polymyxin B against carbapenem-resistant Acinetobacter baumannii

Abstract

Objectives: Increased rates of carbapenem-resistant strains of Acinetobacter baumannii have forced clinicians to rely upon last-line agents, such as the polymyxins, or empirical, unoptimized combination therapy. Therefore, the objectives of this study were: (a) to evaluate the in vitro pharmacodynamics of meropenem and polymyxin B (PMB) combinations against A. baumannii; (b) to utilize a mechanism-based mathematical model to quantify bacterial killing; and (c) to develop a genetic algorithm (GA) to define optimal dosing strategies for meropenem and PMB.

Methods: A. baumannii (N16870; MIC_meropenem = 16 mg/L, MIC_PMB = 0.5 mg/L) was studied in the hollow-fibre infection model (initial inoculum 10⁸ cfu/mL) over 14 days against meropenem and PMB combinations. A mechanism-based model of the data and population pharmacokinetics of each drug were used to develop a GA to define the optimal regimen parameters.

Results: Monotherapies resulted in regrowth to ~10¹⁰ cfu/mL by 24 h, while combination regimens employing high-intensity PMB exposure achieved complete bacterial eradication (0 cfu/mL) by 336 h. The mechanism-based model demonstrated an SC₅₀ (PMB concentration for 50% of maximum synergy on meropenem killing) of 0.0927 mg/L for PMB-susceptible subpopulations versus 3.40 mg/L for PMB-resistant subpopulations. The GA had a preference for meropenem regimens that improved the %T > MIC via longer infusion times and shorter dosing intervals. The GA predicted that treating 90% of simulated subjects harbouring a 10⁸ cfu/mL starting inoculum to a point of 10⁰ cfu/mL would require a regimen of meropenem 19.6 g/day 2 h prolonged infusion (2 hPI) q5h + PMB 5.17 mg/kg/day 2 hPI q6h (where the 0 h meropenem and PMB doses should be 'loaded' with 80.5% and 42.2% of the daily dose, respectively).

Conclusion: This study provides a methodology leveraging in vitro experimental data, a mathematical pharmacodynamic model, and population pharmacokinetics provide a possible avenue to optimize treatment regimens beyond the use of the 'traditional' indices of antibiotic action.

Keywords: Acinetobacter baumannii; Antibiotic resistance; Combination therapy; Genetic algorithm; Machine learning; Mechanism-based model; Meropenem; Pharmacodynamics; Pharmacometrics; Polymyxin.