Parrallel LSTM-DNN Fusion Model for Early Prediction of Sepsis in Intensive Care Units

Sepsis is a significant public health problem and a leading cause of death in the world. Sepsis is defined as a life-threatening condition that occurs when the body's response to an infection causes tissue damage, organ failure, or death. To help doctors improve the survival chances of septic patients in intensive care units (ICUs), experts recommend using Neural Networks like Long Short-Term Memory (LSTM) and Deep Neural Network (DNN) for early sepsis prediction. In this study, we used a publicly available dataset sourced from PhysioNet to predict the onset of sepsis six hours before clinical manifestation. To achieve this, we developed a parallel LSTM-DNN fusion architecture. This model organizes the data into samples containing two distinct feature sets. Each sample includes two subsets of features: feature set 1 (FS1) and feature set 2 (FS2). These feature sets are derived from the 40 clinical parameters available in the dataset. FS1 consists of 7 vital signs and age parameters, organized as a 16-hour multivariate clinical time series. FS2 includes 26 laboratory test results, as well as gender and ICU type parameters, presented as single measurements with features aggregated using median values within the same 16-hour window as FS1. A sample with FS1 is used as input data to the LSTM model, while the same sample with FS2 is used as input data to the DNN model. The outputs from the LSTM and DNN models are fused to generate sepsis probability score, with a threshold of 0.5 used to differentiate septic and non-septic cases. Using the 2019 PhysioNet/Computing in Cardiology Challenge dataset, our approach achieved a utility score of 0.20 as defined by challenge organizers and an Area Under the Receiver Operating Characteristic curve of 0.72.