A hybrid machine learning approach for early mortality prediction of ICU patients

Hospitals face many pressures, including limited budgets and resources. The intensive care unit (ICU), has attracted much attention from the medical community. Patients in the ICU are monitored continuously and their physiological data is consistently measured. This provides a valuable opportunity to analyze significant clinical data. Predicting the mortality of patients in the ICU is crucial. By using this prediction, intensivists identify patients who will not benefit from receiving treatment in the ICU and focus more on the care of patients who will benefit from this treatment. To date, various scoring systems and machine learning models have been developed to predict mortality in intensive care units. In this paper, our goal is to provide a model that can predict mortality for up to 24 h after ICU admission. This research has been conducted using Medical Information Mart for Intensive Care III (MIMIC-III) database. Relevant data have been extracted, preprocessed, and prepared for data mining analysis. The proposed framework has two algorithmic stages. In the first step, time-series data within 24 h of admission is fed to a convolutional neural network with optimized hyperparameters. In the second step, by using the output of the previous stage and a filtering strategy for temporal features, a new data set is created and is fed to an Xgboost algorithm, for final classification. The proposed framework outperforms severity of illness scores and machine learning models within 24 h of admission to the ICU and attains a ROC AUC of 0.863 (± 0.004).