Investigation on Ensemble Models for Mortality Prediction in Intensive Care Unit Patients

Abstract

Background: The intensive care unit (ICU) is a crucial component of the hospital. Allocating resources according to the needs of patients in the ICU is vital for the quality of care. Predicting mortality in this unit can assist nurses and doctors in allocating optimal resources for patients. Objectives: The present study aims to compare the performance of bagging and boosting methods in predicting the mortality of patients admitted to the ICU using demographic, clinical, and laboratory information. Methods: Starting in February 2020, we conducted a study analyzing the demographic, clinical, and laboratory characteristics of 2,055 adult patients admitted to the ICU of a selected hospital over one year. We employed Random Forest (RF), LightGBM (LGBM), and XGBoost (XG) models to compare their accuracy in predicting outcomes. To ensure data integrity, we utilized the interquartile range (IQR) to identify and remove outliers and excluded rows with missing values. Our study also highlighted the significance of various patient characteristics on mortality rates and utilized logistic regression to calculate odds ratios with a 95% confidence interval. Results: The study indicated that the accuracy of the RF model is 0.91, while LGBM and XG both achieved an accuracy of 0.93. We also compared them using the receiver operating characteristic (ROC) curve, with RF (area = 0.91), LGBM (area = 0.94), and XG (area = 0.94). It can be concluded that LGBM and XG had almost the same performance. Conclusions: Based on the accuracy of traditional scoring methods in past studies, we found that machine learning methods have higher accuracy. In this study, the performance of ensemble models was reported to be better than individual models used in previous studies. Furthermore, when comparing ensemble methods (bagging and boosting), boosting techniques (LGBM, XG) demonstrated similar performance and were superior to the bagging strategy (RF).