BACKGROUND:Weaning from mechanical ventilation remains a critical challenge in intensive care units. Machine learning has shown potential in supporting clinical decisions during this process.
OBJECTIVE:Sepsis frequently leads to ALI/ARDS, requiring mechanical ventilation. However, with evolving definitions of weaning, many existing predictive models have become outdated. This study aimed to develop a predictive model based on the standardized WIND framework to accurately predict successful weaning in septic patients under current clinical practices.
METHODS:Data from the MIMIC-IV database were analyzed. Univariate analysis identified risk factors for extubation outcomes, and feature selection was performed using LASSO regression with 10-fold cross-validation and recursive feature elimination (RFE). Predictive models, including XGB, RF, and GBM, were evaluated based on AUC and F1 score. SHAP values were used to assess feature importance.
RESULTS:A total of 3774 patients were included. Univariate analysis showed that the failed weaning group had longer ICU stays, higher ventilator settings, and elevated levels of blood urea nitrogen, blood glucose, creatinine, SOFA scores, lactate, and platelet count (P < 0.05). Feature selection reduced 46 variables to 12 key predictors. The XGB model performed best, with AUC values of 0.849, 0.838, and 0.825 for the training, internal, and external cohorts, respectively. SHAP analysis identified mean airway pressure, ICU length of stay, and lactate as the most influential predictors.
CONCLUSION:We developed an interpretable, accurate XGB-based model to predict weaning outcomes in septic patients.