Development and validation of machine learning models for predicting extubation failure in patients undergoing cardiac surgery: a retrospective study

Patients with multiple comorbidities and those undergoing complex cardiac surgery may experience extubation failure and reintubation. The aim of this study was to establish an extubation prediction model using explainable machine learning and identify the most important predictors of extubation failure in patients undergoing cardiac surgery. Data from 776 adult patients who underwent cardiac surgery and were intubated for more than 24 h were obtained from the Medical Information Mart for Intensive Care (MIMIC)-IV database. The primary endpoint was extubation failure according to the WIND criteria, with 205 patients experiencing extubation failure. The data was split into a training set (80%) and a test set (20%). The performance of the XGBoost algorithm was the highest (AUC 0.793, Mean Precision 0.700, Brier Score0.150), which was better than that of logistic regression (AUC 0.766, Mean Precision 0.553, Brier Score0.173) and random forest (AUC 0.791, Mean Precision 0.510, Brier Score 0.181). The most crucial predictor of extubation failure is the mean value of the anion gap in the 24 h before extubation. The other main features include ventilator parameters and blood gas indicators. By applying machine learning to large datasets, we developed a new method for predicting extubation failure after cardiac surgery in critically ill patients. Based on the predictive factors analyzed, internal environmental indicators and ventilation characteristics were important predictors of extubation failure. Therefore, these predictive factors should be considered when determining extubation readiness.