Prediction of Sepsis from Clinical Data Using Long Short-Term Memory and eXtreme Gradient Boosting

Developing an objective and efficient computer-aided tool for early detection of sepsis has become a promising research topic. In this paper, we present two methods for early prediction of sepsis from clinical data. One is neural network-based method and the other is eXtreme Gradient Boosting (XGBoost) based method. Considering the temporal relationship between clinical data from sepsis patients in the ICU, we built a Long Short-Term Memory (LSTM) network to extract the intrinsic relation between different indicators in clinical data and meanwhile model the temporal dependencies, which only uses the previous information not future information to predict the results. Neural networks have made great achievements in unstructured data, such as image processing and speech processing, while traditional machine learning methods are better at processing structured data than neural networks. Thus, we trained an XGBoost model on the pre-processed data for improving the prediction accuracy. In official phase, we only used the first seven vital signs in our network, on test set A, the LSTM-based method has the utility score is 0.267 and the score of XGBoost-based method is 0.392. We submit the latter method as the final entry and the official final test utility score is 0.313. Our team name is CQUPT_Just_Try, and the ranking is 15th.