Comparison of machine learning models for the prediction of mortality of patients with unplanned extubation in intensive care units

Unplanned extubation (UE) can be associated with fatal outcome; however, an accurate model for predicting the mortality of UE patients in intensive care units (ICU) is lacking. Therefore, we aim to compare the performances of various machine learning models and conventional parameters to predict the mortality of UE patients in the ICU. A total of 341 patients with UE in ICUs of Chi-Mei Medical Center between December 2008 and July 2017 were enrolled and their demographic features, clinical manifestations, and outcomes were collected for analysis. Four machine learning models including artificial neural networks, logistic regression models, random forest models, and support vector machines were constructed and their predictive performances were compared with each other and conventional parameters. Of the 341 UE patients included in the study, the ICU mortality rate is 17.6%. The random forest model is determined to be the most suitable model for this dataset with F₁ 0.860, precision 0.882, and recall 0.850 in the test set, and an area under receiver operating characteristic (ROC) curve of 0.910 (SE: 0.022, 95% CI: 0.867–0.954). The area under ROC curves of the random forest model was significantly greater than that of Acute Physiology and Chronic Health Evaluation (APACHE) II (0.779, 95% CI: 0.716–0.841), Therapeutic Intervention Scoring System (TISS) (0.645, 95% CI: 0.564–0.726), and Glasgow Coma scales (0.577, 95%: CI 0.497–0.657). The results revealed that the random forest model was the best model to predict the mortality of UE patients in ICUs.