Safety-Guaranteed Lung-Protective Mechanical Ventilation using Digital Twins and Reinforcement Learning

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Computational Physiology

Background:

• Mechanical ventilation is crucial for respiratory support but carries risks like ventilator-induced lung injury (VILI).
• Existing closed-loop systems often prioritize oxygenation, neglecting safety limits for VILI and other physiological parameters.
• Individualized ventilation control is needed to optimize patient-ventilator interaction and minimize harm.

Purpose of the Study:

• To develop and validate a novel, safety-guaranteed closed-loop mechanical ventilation control strategy.
• To minimize the risk of VILI by optimizing ventilator settings within safe physiological bounds.
• To enhance patient-ventilator synchrony and potentially reduce ventilation duration.

Main Methods:

• A Reinforcement Learning (RL) controller based on a Constrained Markov Decision Process (CMDP) was developed.
• The RL controller utilizes an enhanced Primal-Dual Soft Actor-Critic (SAC) algorithm with dual variables for parameter bounding.
• A high-fidelity computational patient model (digital twin) was used for in silico training of the RL controller.

Main Results:

• The Primal-Dual SAC controller safely converged to optimal ventilator settings.
• The system demonstrated the ability to minimize VILI risk while maintaining physiological parameters within safe limits.
• Simulations based on real patient data confirmed the controller's robust and correct functioning.

Conclusions:

• The proposed AI-driven control strategy offers a safety-guaranteed approach to mechanical ventilation.
• This technique addresses limitations of current closed-loop systems by actively minimizing VILI.
• The digital twin and RL-based controller show promise for personalized and safer mechanical ventilation in clinical practice.