DEVELOPMENT AND ASSESSMENT OF DIFFERENT BALANCING TECHNIQUES AND DEEP LEARNING BASED EARLY MORTALITY PREDICTION MODELS FOR ICU IMBALANCE DATA

Predicting ICU patient’s mortality is an important area of research to assist the clinical staff in decision-making, and subsequently make more exact strategies in recognizing high mortality risk patients. Handling huge clinical data is still a challenge. These data have different issues such as high dimensions, many missing values, imbalanced data, time-series and data recorded irregularly. This paper mainly focuses on developing four different deep learning models: deep neural network (DNN), deep long-short term memory (DLSTM), deep bidirectional long-short term memory (DBLSTM) and deep gated recurrent unit (DGRU) using two standard datasets, the Physionet challenge 2012 and WiDS datathon 2020 to predict the ICU patients mortality. During the simulation study, missing values are handled using k-NN imputation in Physionet and mean imputation in WiDSdatathon, then balancing of the data is done by employing synthetic minority oversampling technique (SMOTE), cost sensitive learning (CSL) and generative adversarial network (GAN). Feature extraction is done by using discrete wavelet transform (DWT) in WiDS datathon. From the simulation study, it is demonstrated that the (SMOTE+DGRU) has obtained AUC, F1-score and accuracy of 0.8081, 0.7964, 0.8081 respectively in 90 epochs for Physionet challenge 2012. Whereas (SMOTE+DBLSTM) has provided AUC, F1-score and accuracy of 0.8724, 0.8739, 0.8724 respectively for same epochs for WiDSdatathon. In overall, it is observed that the SMOTE balancing technique is performing better in comparison to CSL and GAN.