Prediction and simulation of PEEP setting effects with machine learning models.

Area of Science:

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

Background:

• Mechanical ventilation is crucial for critically ill patients.
• Positive end-expiratory pressure (PEEP) management is complex and impacts patient outcomes.
• Current PEEP titration methods may not fully utilize available patient data.

Purpose of the Study:

• To develop and validate a machine learning (ML) model for automated PEEP adjustment.
• To utilize routinely measured data for predicting ventilation success parameters.
• To establish a novel, data-driven approach for PEEP titration in intensive care units (ICUs).

Main Methods:

• Retrospective observational study design.
• Utilized data from 51,811 mechanically ventilated patients across multiple US ICUs (MIMIC-III and eICU databases).
• Developed a multi-tasking neural network to predict arterial blood gases and respiratory compliance.

Main Results:

• The ML model achieved superior performance across all prediction tasks.
• Accurate 45-minute ahead predictions for partial pressures of oxygen (PaO2) and carbon dioxide (PaCO2), and respiratory system compliance.
• Demonstrated model utility through simulated case scenarios for PEEP adjustment.

Conclusions:

• Machine learning offers a promising, cost-effective method for PEEP titration.
• The developed model can predict key ventilation parameters using existing ICU data.
• Access to comprehensive ICU patient data is vital for enhancing predictive model accuracy.