Ensemble Machine Learning Model for Mortality Prediction Inside Intensive Care Unit

Abstract

Intensive care unit (ICU) admits the most seriously ill patients requiring extensive monitoring. Early mortality prediction is a crucial issue in intensive care. If a patient's likelihood of survival or mortality is predicted early enough, the patient will give proper and timely care to save the patient's life. In recent decades various severity scores and machine learning models have been developed for mortality prediction. By contrast, mortality prediction is still an open challenge. The main objective of this study is to provide a new framework for mortality prediction based on an ensemble classifier. The proposed ensemble classifier amalgamates five different classifiers: Linear discriminant analysis, decision tree, multilayer perceptron, Knearest neighbor, and logistic regression. Data is vertically divided according to expert medical opinion into six feature sets, choosing the most accurate classifier for each subset of features. Framework evaluated benchmark data from Medical Information Mart for Intensive Care (MIMIC III) database using the first 24 h collected for each patient. The performance was validated using standard metrics include Precision, Recall, F-score, Area under the roc curve. Results show F-score 91.02% with precision 92.34% recall 90.14% for MLP classifier, achieved F-score of 93.7%, precision of 96.4%, recall of 91.1%, and AUROC of 93.3% for proposed ensemble classifiers which outperforms using classical classifiers. The results show the validity of the proposed system to be an assistant system for physicians in ICU.