Early Sepsis Prediction Using Ensemble Learning with Features Extracted from LSTM Recurrent Neural Network

Early prediction of sepsis can help to identify potential risks in time and help take necessary measures to prevent more dangerous situations from occurring. In PhysioNet/Computing in Cardiology Challenge 2019, we integrate Long Short Term Memory (LSTM) recurrent neural network and ensemble learning to achieve early sepsis prediction. Specifically, we tackle the problem of class imbalance and data missing firstly, and then we manually extract features according to the prior knowledge from the medical field. In addition, we regard the prediction of sepsis as a time series prediction problem and pre-train LSTM-based models as feature extractors to obtain the "deep" features on time series that might be related to the onset of sepsis. Manual features and "deep" features are then used to train prediction models under the framework of ensemble learning, including Extreme Gradient Boosting (XGBoost) and Gradient Boosting Decision Tree (GBDT) regressor. The final normalized utility score our team (UCAS_DataMiner) have obtained was 0.313 on full hidden test set.