ventisim 0.1.0

Gymnasium environments for reinforcement learning in mechanical ventilation

ventisim

Gymnasium Environments for Reinforcement Learning in Mechanical Ventilation

VentiSim provides standardized Gymnasium environments for developing and benchmarking reinforcement learning algorithms applied to mechanical ventilation management. The environments model lung mechanics, gas exchange, and PEEP-dependent alveolar recruitment with physiologically grounded parameters derived from the ARDS literature.

Installation

pip install ventisim

With training dependencies (stable-baselines3, PyTorch):

pip install ventisim[train]

Full development install:

pip install ventisim[all]

Quick Start

import gymnasium as gym
import ventisim  # registers environments

env = gym.make("ventisim/TidalVolumeControl-v0")
obs, info = env.reset(seed=42)

for step in range(200):
    action = env.action_space.sample()
    obs, reward, terminated, truncated, info = env.step(action)
    if terminated or truncated:
        obs, info = env.reset()

env.close()

Environments

Environment	Paradigm	Obs Dim	Act Dim	Key Challenge
ventisim/TidalVolumeControl-v0	Pressure control	7	1	Deliver target VT (6 mL/kg IBW) while keeping plateau < 30 cmH2O
ventisim/PEEPOptimization-v0	PEEP titration	8	1	Balance recruitment vs. overdistension for optimal PaO2
ventisim/FullVentilatorManagement-v0	Multi-parameter	10	4	Simultaneously optimize oxygenation, ventilation, and lung protection

All environments support three difficulty tiers (easy, medium, hard) controlling inter-patient variability, compliance drift, and measurement noise.

Architecture

Models Layer
  SingleCompartmentLung  -- RC-circuit lung mechanics (compliance, resistance)
  GasExchangeModel       -- Alveolar gas equation + shunt + PEEP recruitment
  VirtualPatient         -- ARDS severity profiles (Berlin definition)

Environments Layer (Gymnasium API)
  TidalVolumeControl-v0         -- 1D action: inspiratory pressure
  PEEPOptimization-v0           -- 1D action: PEEP change
  FullVentilatorManagement-v0   -- 4D action: FiO2, PEEP, RR, P_insp

Agents Layer
  Random                 -- Uniform random sampling (lower bound)
  Heuristic              -- Clinical protocol baselines (proportional control)
  PPO                    -- Proximal Policy Optimization via stable-baselines3

Training

Train all agents and generate results:

python train_all.py

Train a single environment:

ventisim-train --env ventisim/TidalVolumeControl-v0

Run benchmarks (random + heuristic):

ventisim-benchmark

Paper

If you use VentiSim in your research, please cite:

@software{dhia2026ventisim,
  author    = {Dhia, Hass},
  title     = {VentiSim: Gymnasium Environments for Reinforcement Learning
               in Mechanical Ventilation},
  year      = {2026},
  publisher = {Smart Technology Investments Research Institute},
  url       = {https://github.com/HassDhia/ventisim},
}

License

MIT License. See LICENSE for details.

Contact

Hass Dhia -- Smart Technology Investments Research Institute

[e-mail]