AntibioSim

Gymnasium environments for reinforcement learning in antimicrobial stewardship and antibiotic treatment optimization.

AntibioSim provides four Gymnasium-compatible environments that model key clinical decision points in antibiotic therapy: drug selection, dose optimization, therapy switching, and ward-level resistance control. Each environment is grounded in established pharmacokinetic/pharmacodynamic (PK/PD) models and bacterial dynamics from the antimicrobial stewardship literature.

Environments

Installation

pip install antibiosim

For training:

pip install antibiosim[train]

Quick Start

import gymnasium as gym
import antibiosim

env = gym.make("antibiosim/AntibioticSelection-v0")
obs, info = env.reset()

for _ in range(14):
    action = env.action_space.sample()
    obs, reward, terminated, truncated, info = env.step(action)
    if terminated or truncated:
        break

Domain Models

• Pharmacokinetics: One-compartment and two-compartment PK models (Drusano, 2004)
• Pharmacodynamics: Emax and sigmoidal Emax models (Regoes et al., 2004)
• Bacterial dynamics: Logistic growth with antibiotic kill (Austin et al., 1999)
• Resistance: Two-population susceptible/resistant dynamics (Levin & Bonten, 2004)

Baseline Agents

• Random: Uniform random action selection
• Heuristic: Clinical guideline-based rules (Barlam et al., 2016)
• PPO: Proximal Policy Optimization via Stable-Baselines3

Training

# Train PPO on all environments
python -m antibiosim.training.train_all

# Train a single environment
antibiosim-train --env antibiosim/DoseOptimization-v0

Citation

@software{dhia2026antibiosim,
  author = {Dhia, Hass},
  title = {AntibioSim: Gymnasium Environments for Reinforcement Learning in Antimicrobial Stewardship},
  year = {2026},
  url = {https://github.com/HassDhia/antibiosim},
  version = {0.1.0}
}

License

MIT License. Copyright (c) 2026 Hass Dhia, Smart Technology Investments Research Institute.