Development of a Hybrid Model using Deep Neural Network and XGBOOST for Mortality Prediction of ICU Patients

With the rapid development of various machine learning techniques, this paper proposes a hybrid model using Deep Neural Network (DNN) and XGBoost (XGB) to predict the ICU patient's mortality accurately. For prediction of mortality Physionet challenge 2012 dataset is used, which consists of 8000 records (set A and B) and 42 features. The outcome of this study reveals whether the ICU patient will survive or dies. Since, the dataset has large number of missing values and also highly imbalanced, makes the classification problem more complex. To circumvent these problems, k-Nearest Neighbor (k-NN) imputation is used to handle the missing value and (Synthetic Minority Oversampling Technique) SMOTE for balancing the dataset. Then, the proposed model is trained using 5-fold cross-validation in 80% of the data. The model is tested with rest 20% of the data and various performance measures such as accuracy, sensitivity, specificity, f1-score, score1 and AUC are obtained. The hybrid model (DNN+XGB) is compared with the Gaussian Support Vector Machine and Multinomial Naive Bayes models and reported to be better predictor with AUC of 75.05%.