A novel solution of using deep learning for early prediction cardiac arrest in Sepsis patient: enhanced bidirectional long short-term memory (LSTM)

Cardiac arrest is a common issue in Intensive Care Units (ICU) with low survival rate. Deep learning algorithms have been used to predict cardiac arrest unsuccessfully due to low sensitivity and high false alarm rate. The aims of this research are to decrease the false alarm rate and increase sensitivity and specificity. The proposed system is based on Medical Information Mart for Intensive Care (MIMIC-III) database where two sets of data are created. These two datasets are time-series data and combination of time series and baseline data. Time series dataset is divided into six-time groups. The system model consists of a hybrid model: Multilayer Perceptron (MLP) and enhanced Bidirectional Long Short-Term Memory (LSTM). MLP processes baselines feature like age, sex, chief complaints whereas the bidirectional LSTM is used to handle time series vital signs data from forward and backward direction so that it considers both present and future inputs. The model predicts cardiac arrest up to six hours earlier before the incidence. We achieved better performance for combined dataset where the prediction time window is 1 h. Accuracy, sensitivity, specificity, and Area Under Curve (AUC) equal to 85.7%, 87.7%,84.9%, and 0.86 respectively for the state of art, for proposed solution are 92.6%, 94.3%, 93.6% and 0.94 respectively. The proposed system is reducing the false alarm rate and increasing accuracy, sensitivity, specificity, and the area under curve for the prediction of cardiac arrest using enhanced Bidirectional LSTM model. The problem of missing values, irregularities of time series, and imbalance data set is solved too.