Early Sepsis Prediction Using LSTM Recurrent Neural Network

Zhengling He1, Xianxiang Chen2, Zhen Fang2, Chenshuo Wang2, Li Jiang2, Zhongkai Tong2, Zhongrui Bai2, Yichen Pan2, Yueqi Li2
1University of Chinese Academy of Sciences, 2Institute of Electronics, Chinese Academy of Sciences


Early prediction of sepsis can help to identify risks in time and take necessary measures to prevent more dangerous situations from occurring. In this challenge, we regard the prediction of sepsis as a time series prediction problem, and construct a deep structrue based on recurrent neural network (RNN), a deep learning method, to automatically learn potential features based on training data sets to predict the occurrence of sepsis. Specifically, the model consists of 3 long short-term memory (LSTM) layers and 4 fully connected (FC) layers with dropout strategy, the segments with the timestep of 4 hours are used as the input of this model to learn the dependence on time steps, random undersampling and weighted loss function are applied to tackle with the problem of class imbalance. The performance of the proposed model was validated on officially published datasets A and B (including 40,336 subjects) by 5-fold cross validation, the preliminary results showed the mean sensitivity was 0.72, specificity was 0.79 and the normalized utility score was 0.48.