Early Prediction of Sepsis Utilizing Multi-branches Multi-tasks Hybrid Deep Learning Model

Sepsis is a global leading cause of death, with a mortality rate of up to 50% for septic shock. Early diagnosis and treatment can prevent the majority of cases and deaths. The information-abundant setting within intensive care units and the prompt growth of medical data analytics have generated significant interest in developing intelligent models to predict acute medical conditions such as septic shock. This paper proposes a novel multi-branch multi-tasks hybrid deep neural networks for early predicting the onset of sepsis. We design different hybrid models to handle physiological signals of different frequencies. For low-frequency signals, we utilize a novel temporal convolution neural network model with multi-scale attention mechanism for extracting global features (Branch 1). For high-frequency signals, we utilize a novel multi-view convolutional neural network with bidirectional recurrent neural network and temporal attention mechanism for extracting local features (Branch 2). In addition, we introduce the concept of multi-task learning to enhance the inference capability of the deep learning model. The first sub-task is to determine whether a patient will experience sepsis onset (a classification problem), and the second sub-task is to predict how much time remains before sepsis onset for a patient (a regression problem). These two sub-tasks are interrelated, and through parallel learning of both sub-tasks, we can learn the correlation between them to further improve predictive ability. This approach aims to provide early solutions to mitigate the risk of sepsis onset.