Prediction of weaning in mechanically ventilated sepsis patients using interpretable machine learning methods

Area of Science:

• Critical Care Medicine
• Machine Learning in Healthcare
• Respiratory Therapy

Background:

• Mechanical ventilation weaning is a significant challenge in intensive care units (ICUs).
• Sepsis often necessitates mechanical ventilation due to Acute Lung Injury/Acute Respiratory Distress Syndrome (ALI/ARDS).
• Existing weaning prediction models may be outdated due to evolving clinical definitions.

Purpose of the Study:

• To develop and validate a machine learning model for predicting successful weaning from mechanical ventilation in septic patients.
• To utilize the standardized Weaning Index (WIND) framework for model development.
• To ensure the model reflects current clinical practices for weaning septic patients.

Main Methods:

• Analysis of data from the MIMIC-IV database.
• Identification of risk factors using univariate analysis and feature selection via LASSO regression and Recursive Feature Elimination (RFE).
• Development and evaluation of predictive models (XGB, RF, GBM) using AUC and F1 scores, with SHAP values for feature importance analysis.

Main Results:

• A cohort of 3774 patients was analyzed.
• Feature selection identified 12 key predictors from an initial 46 variables.
• The XGB model demonstrated superior performance with AUCs of 0.849 (training), 0.838 (internal), and 0.825 (external validation).
• Key predictors included mean airway pressure, ICU length of stay, and lactate levels.

Conclusions:

• An interpretable and accurate XGB-based model was developed to predict mechanical ventilation weaning outcomes in septic patients.
• The model provides a valuable tool for clinical decision-making in ICUs.
• This approach supports evidence-based practice in respiratory care for critically ill septic patients.