A Reinforcement Learning (RL)-Motivated Simulation Framework for Evaluating Vancomycin Dosing Strategies.

Area of Science:

• Pharmacokinetics and Pharmacodynamics
• Computational Biology
• Machine Learning in Medicine

Background:

• Achieving and maintaining therapeutic vancomycin levels is critical for treatment efficacy and minimizing toxicity.
• Existing guidelines for vancomycin dosing are based on empirical data, but optimal theoretical strategies under diverse conditions are not fully understood.

Purpose of the Study:

• To develop a novel reinforcement learning (RL) based simulation framework for optimizing vancomycin dosing strategies.
• To integrate clinical guidelines into an RL reward system using the area under the time-concentration curve (AUC).

Main Methods:

• Developed a deep learning two-compartment pharmacokinetic model (PK-RNN-2CM).
• Utilized patient-specific data to generate ground truth time-concentration curves.
• Simulated vancomycin dosing strategies under various conditions, including noise perturbations to mimic real-world variability.
• Employed 24-hour AUC and Root Mean Square Error (RMSE) as evaluation metrics.

Main Results:

• Both low-dosing and high-dosing AUC targets showed comparable performance in noise-free simulations.
• The low-dosing strategy achieved higher AUC reward scores in noisy conditions.
• The high-dosing strategy demonstrated greater stability under noisy conditions.

Conclusions:

• The developed RL-based simulation framework provides a new approach for optimizing vancomycin dosing.
• This methodology can help refine dosing strategies to improve patient outcomes in vancomycin therapy.
• The framework's ability to incorporate clinical guidelines and simulate real-world variability offers valuable insights for therapeutic drug monitoring.