(168b) Machine Learning Can Help Design Effective Antimicrobial Dosing Regimens

Objective: Addressing severe bacterial infections continues to be a major challenge, often requiring the use of combination therapy, which entails administering two or more antibiotics that work through different mechanisms. This research investigates the use of machine learning (ML) models to forecast the effectiveness of combination therapies at inhibiting bacterial growth. By incorporating longitudinal data on bacterial load, captured through optical density measurements during time-kill assays, and fitting this data to a mathematical framework, our goal was to improve dosing strategies for combination treatments. Here we processed data from prior work, where we exposed Acinetobacter baumannii to two antibiotic combinations: ceftazidime/amikacin (CAZ/AMK) and ceftazidime/avibactam (CAZ/AVI), tested under both synchronous and asynchronous dosing schedules[1]. Given the difficulty of first principles based methods of mathematical modeling to use for dosing regimen design, our objective was to test whether machine learning methods can help with the task. The ML-driven strategy enhanced the extraction of crucial pharmacodynamic parameters, yielding more precise predictions of the average kill rate (D) in comparison to the bacterial growth rate (K_g). When D/K_g > 1, the bacterial population diminishes, but when D/K_g < 1, it indicates a rise in the bacterial population.

Methods: The following machine learning methods were used to train models based on the data mentioned above: Gaussian Process Regression (GPR), Random Forest Regression, Neural Network (MLP Regressor), Nearest Neighbors Regression (KNN), Linear Regression, Gradient Boosted Trees, and Decision Tree. All models capture the effect of antibiotic concentrations on the kill rate for the least susceptible subpopulation of the infecting bacterial population. The training data were produced in previous work, where kill rate estimates at distinct pairs of antibiotic concentrations were produced. Each trained model is then used to forecast the effectiveness of combination therapies at inhibiting bacterial growth under clinically relevant pharmacokinetics. The forecast is based on a simple comparison of the bacterial growth rate (K_g) to the average kill rate (D) induced by the antibiotic combination. Selected model forecasts are compared to actual experimental data, obtained in vitro in a hollow fiber infection model (HFIM).

Results: The aggregate recommendation by the machine learning methods provides good predictions on the effectiveness of corresponding dosing regimens observed at the HFIM.

Conclusions: The results presented here demonstrates how various machine learning (ML) models can be used to enhance combination therapies aimed at resistant pathogens, which can ultimately lead to improved management of bacterial infections. This innovative ML-based analytical framework offers greater precision, accounts for uncertainty, and provides actionable insights for optimizing dosing regimens in combination therapies.

References:

1.Eales, B. M. et al. Rapid design of combination antimicrobial therapy against Acinetobacter baumannii. Comput Chem Eng 192, 108884 (2025).