woundsim 0.1.0

Gymnasium-compatible reinforcement learning environments for wound healing treatment optimization

WoundSim

Gymnasium-compatible reinforcement learning environments for wound healing treatment optimization

---

WoundSim provides four Gymnasium-compatible RL environments for wound healing treatment optimization, each wrapping a validated ODE model from the wound healing literature. All model parameters are sourced from peer-reviewed publications with inline # SOURCE: comments. The package includes random, clinical heuristic, and PPO baselines with full Stable-Baselines3 compatibility.

Installation

pip install woundsim              # Core (numpy, scipy, gymnasium)
pip install woundsim[rl]          # + SB3, PyTorch for RL training
pip install woundsim[all]         # Everything

Development install:

git clone https://github.com/HassDhia/woundsim.git
cd woundsim
pip install -e ".[all]"

Quick Start

import gymnasium as gym
import woundsim
from woundsim.agents.heuristic import HeuristicAgent

env = gym.make("woundsim/WoundMacrophage-v0", difficulty="medium")
agent = HeuristicAgent(env, env_id="woundsim/WoundMacrophage-v0")
obs, info = env.reset(seed=42)

done = False
total_reward = 0.0
while not done:
    action, _ = agent.predict(obs)
    obs, reward, terminated, truncated, info = env.step(action)
    total_reward += reward
    done = terminated or truncated

print(f"Episode reward: {total_reward:.2f}")
env.close()

Environments

Environment	Model	State Dim	Action Dim	Source
WoundMacrophage-v0	Zlobina macrophage polarization	5	1	Zlobina & Gomez (2022)
WoundIschemic-v0	Simplified Xue-Friedman ischemic	6	2	Xue, Friedman & Sen (2009)
WoundHBOT-v0	Flegg HBOT angiogenesis	4	2	Flegg et al. (2009, 2015)
WoundDiabetic-v0	Extended diabetic inflammation	7	3	Waugh & Sherratt (2006)

WoundMacrophage-v0

Five-variable ODE model of macrophage M1/M2 polarization from Zlobina et al. (2022). The agent controls a polarization treatment signal to balance debris clearance (M1) with tissue regeneration (M2).

• Observation: [debris, M1, M2, granulation, new_tissue] (normalized to [0,1])
• Action: [polarization_signal] in [0, 1]
• Reward: Penalizes debris and treatment cost; rewards tissue growth
• Difficulties: easy (a0=0.3), medium (a0=0.6), hard (a0=0.9 + noise)

WoundIschemic-v0

Six-variable model simplified from Xue & Friedman (2009). The agent controls revascularization intensity and growth factor application to heal ischemic wounds.

• Observation: [wound_area, oxygen, VEGF, macrophages, fibroblasts, ECM]
• Action: [revascularization, growth_factor] in [0, 1]
• Difficulties: mild, moderate, severe

WoundHBOT-v0

Four-variable HBOT angiogenesis model from Flegg et al. (2009, 2015). The agent controls hyperbaric oxygen session parameters to promote vascularization and wound closure.

• Observation: [capillary_tips, sprouts, oxygen, wound_area]
• Action: [intensity, duration_fraction] in [0, 1]
• Difficulties: acute, chronic, non-healing

WoundDiabetic-v0

Seven-variable diabetic wound model incorporating glucose-insulin dynamics from Waugh & Sherratt (2006) with macrophage polarization from Zlobina et al. (2022). The agent must simultaneously manage wound treatment and glycemic control.

• Observation: [wound, debris, M1, M2, glucose, insulin, ECM]
• Action: [polarization, growth_factor, insulin_dose] in [0, 1]
• Difficulties: well-controlled, moderate, uncontrolled

Architecture

woundsim/
  src/woundsim/
    models/        # ODE wound healing models (Zlobina, Xue-Friedman, Flegg)
    envs/          # Gymnasium environments wrapping each model
    agents/        # Random, heuristic, PPO agent implementations
    training/      # Training infrastructure with shared configs
    benchmarks/    # Systematic benchmark evaluation runner
  tests/           # 173 unit and integration tests
  paper/           # LaTeX paper and publication figures
  results/         # Training results (JSON)

Training

# Train PPO on all environments and evaluate against baselines
python train_all.py

# Generate publication-quality figures from training results
python generate_figures.py

Paper

The accompanying paper is available at:

• PDF (GitHub)

Citation

If you use WoundSim in your research, please cite the software:

@software{dhia2026woundsim_software,
  author = {Dhia, Hass},
  title = {WoundSim: Gymnasium-Compatible Reinforcement Learning Environments for Wound Healing Treatment Optimization},
  year = {2026},
  publisher = {Smart Technology Investments Research Institute},
  url = {https://github.com/HassDhia/woundsim}
}

To cite the accompanying paper:

@misc{dhia2026woundsim,
  author = {Dhia, Hass},
  title = {WoundSim: Gymnasium-Compatible Reinforcement Learning Environments for Wound Healing Treatment Optimization},
  year = {2026},
  howpublished = {\url{https://github.com/HassDhia/woundsim/blob/main/paper/woundsim.pdf}},
  institution = {Smart Technology Investments Research Institute}
}

License

MIT License. See LICENSE for details.

Contact

Hass Dhia - Smart Technology Investments Research Institute

• Email: [e-mail]
• Web: smarttechinvest.com/research