Development and validation of a reinforcement learning algorithm to dynamically optimize mechanical ventilation in...

Area of Science:

• Critical Care Medicine
• Artificial Intelligence in Healthcare
• Pulmonary Engineering

Background:

• Mechanical ventilation is crucial for critically ill patients but optimizing settings is complex.
• Current ventilation strategies may not be sufficiently individualized, potentially impacting patient outcomes.
• Dynamic adjustment of ventilation parameters is needed to improve patient prognoses.

Purpose of the Study:

• To develop and evaluate VentAI, a reinforcement learning algorithm for dynamic mechanical ventilation optimization.
• To assess VentAI's performance in suggesting personalized ventilation regimes for critically ill patients.
• To compare VentAI's efficacy against standard clinical care in reducing mortality.

Main Methods:

• Developed VentAI using a Markov decision process with Q-learning and a reward system.
• Extracted patient data fingerprints (44 features) as multidimensional time series.
• Validated and tested VentAI on large datasets of mechanical ventilation events from ICU admissions.

Main Results:

• VentAI demonstrated significantly higher estimated performance returns (83.3-84.1) versus standard care (51.1).
• VentAI recommended more frequent adjustments in positive end-expiratory pressure (PEEP) and tidal volume (Vt) within optimal ranges.
• VentAI reduced the use of high fraction of inspired oxygen (FiO2) levels, favoring moderate concentrations.

Conclusions:

• VentAI offers a reproducible, high-performance approach to dynamically optimize individualized mechanical ventilation.
• The algorithm's dynamic strategy may benefit critically ill patients requiring mechanical ventilation.
• Reinforcement learning holds promise for improving mechanical ventilation management in intensive care units.