An Ensemble Learning Algorithm for ICU Patient Mortality Prediction

Prediction of patient mortality in Intensive Care Units (ICU) can aid the prevision of timely medical intervention and allocation of vital resources to those patients who are at the greatest risk of dying and for the provision of suitable interventions to save their lives. There is a lack of current research in an accurate, robust and timely solution that can handle complex imbalanced ICU data with significant missing values. This study examines the use of ensemble algorithms to produce reliable results in the prediction of ICU patient mortality by using patient’s medical history data. Four popular single classifiers; Decision Trees (DT), Naive Bayes (NB), Logistic Regression (LR) and Support Vector Machines (SVM) and three ensemble classifiers; Light Gradient Boosting Machine (LGBM), Random Forest (RF) and Stacking are implemented in this research. Experiments are conducted with and without feature selection on a test set that handles data imbalance and missing values. The results indicate that the LGBM model without feature selection outperformed the state of the art approaches in terms of accuracy (0.97) and Area Under the Curve (AUC) (0.97). It was found that automated data-driven features selection did not improve the model performance if there was no prior domain knowledge