The early prediction of neonates mortality in Intensive Care Unit

Predicting mortality in Intensive Care Units (ICU) is an important issue because resources are very expensive and limited compared to the needs. Besides, decision-making is critical in the part of hospitals. Many scoring systems and machine learning models have emerged in previous years in this context. But the combination of an efficient prediction in the earliest hours of admission to the ICU remains a great and open challenge. Neonates are patients aged under 28 days. Neonatal ICU mortality prediction after the two first hours of admission using the machine learning tools is the main goal of this work. Data used includes information such as laboratory test results, demographic data and vital sign measurements. The proposed solution is composed of three steps: (1) The first step, is feature selection based on feature importance and recursive feature elimination. (2)The second step is patients' classification into mortal and alive using an ensemble of algorithms in order to keep the best performing ones. Tested algorithms were Classification and Regression Trees (CART), K-Nearest Neighbors (KNN), Linear Discriminant Analysis (LDA), Logistic Regression (LR), Naive Bayes (NB), Random Forest (RF) and Support Vector Machine (SVM). LDA was the best performing one with an accuracy = 0.947 and AUROC = 0.97.(3)The last step is mortality time prediction using the Galaxy-Random Forest method (f-score=0.871).