Pulmonary gas exchange evaluated by machine learning: a computer simulation.

Area of Science:

• Pulmonary physiology and critical care medicine
• Computational modeling and machine learning applications

Background:

• Quantifying pulmonary gas exchange in intensive care units (ICUs) is crucial for patient management.
• Traditional methods may not fully capture the complexity of multi-compartment pulmonary blood flow.
• Machine learning (ML) offers potential for advanced analysis of complex physiological data.

Purpose of the Study:

• To investigate if ML analysis of ICU monitoring data can quantify pulmonary gas exchange.
• To assess the accuracy of ML in estimating multi-compartment ventilation/perfusion (V/Q) parameters.
• To compare 'two-point' versus 'single-point' ML estimation strategies.

Main Methods:

• A 21-compartment V/Q model simulated pulmonary blood flow under various physiological conditions.
• Machine learning ensembles ('stacked regressor') were trained and validated on simulated data.
• ML models estimated shunt, log standard deviation (SD), and mean V/Q using arterial blood gases from two different inspired oxygen fraction (FiO2) settings.

Main Results:

• Two-point ML estimates of shunt, log SD, and mean V/Q demonstrated high accuracy (R² ≈ 1.00, slope ≈ 1.00, intercept ≈ 0.00).
• Kernel density and Bland-Altman plots confirmed excellent agreement between estimated and true V/Q parameters.
• Single-point ML estimates, using only one FiO2 setting, showed lower accuracy (R² = 0.77–0.89).

Conclusions:

• ML analysis of blood gas, indirect calorimetry, and cardiac output data can effectively quantify pulmonary gas exchange.
• Utilizing data from two distinct FiO2 settings significantly enhances the accuracy of ML-based V/Q modeling.
• This approach holds promise for detailed pulmonary function assessment in critical care settings.