Reinforcement Learning to Optimize Ventilator Settings for Patients on Invasive Mechanical Ventilation: Retrospective...

Area of Science:

• Critical Care Medicine
• Artificial Intelligence in Healthcare
• Mechanical Ventilation

Background:

• Improper mechanical ventilation settings are a significant cause of pulmonary injury in critically ill patients.
• Artificial intelligence (AI) offers potential solutions for optimizing mechanical ventilation in intensive care units (ICUs).
• Machine learning algorithms can personalize ventilation strategies by analyzing patient data.

Purpose of the Study:

• To design and evaluate an AI solution for tailoring optimal ventilator strategies for critically ill patients.
• To assess the impact of AI-driven ventilation settings on patient outcomes.

Main Methods:

• A reinforcement learning-based AI solution was developed using observational data from multiple US ICUs.
• The AI agent recommended personalized levels of positive end-expiratory pressure, fraction of inspired oxygen, and tidal volume.
• Off-policy evaluation metrics were used to assess the AI policy's effectiveness.

Main Results:

• The AI solution was evaluated on 21,595 (eICU) and 5,105 (MIMIC-IV) patient ICU stays.
• Estimated hospital mortality rates were lower with the AI policy compared to observed clinical practice.
• The AI policy also improved the proportion of optimal oxygen saturation and mean arterial blood pressure.

Conclusions:

• Customizing mechanical ventilation settings with AI led to lower estimated hospital mortality.
• Reinforcement learning is effective for developing AI models to optimize treatment parameters in complex clinical data.
• Integration into clinical decision support systems and prospective validation are recommended.