Early Prediction of Sepsis from Clinical Data Using Deep Neural Network

Wenjie Cai and Danqin Hu
University of Shanghai for Science and Technology


Sepsis is a life-threatening condition when tissue injury occurs in response to infection. The objective of the PhysioNet/Computing in Cardiology Challenge 2019 is the early detection of sepsis using time series physiological data. To get the maximum effect of early detection of sepsis, the labels of the training data that indicate the onset of sepsis were moved 5 hours earlier than the optimal prediction time point. The proposed deep learning model contained two parts. The first part was a small artificial neural network that calculated the weights of 40 kinds of values include vital signs, laboratory values and demographics. And this network learned and generated features that combined all the original data for every time point. The second part of the model used two stacked bidirectional LSTM layers to deal with the generated features in temporal order. And then a time distributed full connected network was applied to get the final predictions. Hyperparameters were finely tuned using five fold cross-validation. In validation set, the overall score of utility was 0.138, the F-measure score was 0.137, the accuracy was 0.877, the AUPRC score was 0.099 and the AUROC score was 0.686. The model still needs to be improved to get better prediction of sepsis from clinical data.