Prediction of Sepsis from Clinical Data Using a Convolutional Recurrent Neural Networks

Wang Yongchao1, Bin Xiao2, Xiuli Bi2, Weisheng Li2, Junhui Zhang3, Xu Ma4
1Chongqing University of Posts and Telecommunications, 2Chongqing University of Posts and Telecomunications, 3The First Affiliated Hospital of Chongqing Medical University, 4Human Genetics Resource Center, National Research Institute for Family Planning


Sepsis is a life-threatening condition, and more than 6 million people die from sepsis every year. Therefore, developing an objective and efficient computer-aided tool for early detection of sepsis has becoming a promising research topic. In this paper, we present a novel method for early prediction of sepsis from clinical data by combing Convolutional Neural network (CNN) and Long Short-Term Memory (LSTM). On the one hand, the proposed method takes advantage of CNN model to extract the intrinsic relation between different indicators in clinical data at the same time. On the other hand, LSTM is built in the proposed method to model the temporal dependencies, which only uses the previous information not future information to predict the results. We only used the first seven vital signs in our network and local cross-validation results on training data reveal that the utility score is 0.314.