Early Prediction of Sepsis Using Auxiliary Features and Hierarchical LSTM Network

ByeongTak Lee, KyungJae Cho, Oyeon Kwon


In this work, feature augmentation method and hierarchical long-short term memory network are presented for early prediction of sepsis using database collected from two different hospitals which provided by the 2019 Physionet Challenge. Augmented features are consist of value difference between time steps(dD), missingness of origin data(M), the time difference from the last measured time(dT), and NEWS and SOFA score, which are designed to let model easily understand feature information and domain knowledge. To allow the model to learn representations of highly correlated features, auxiliary features with origin data are split into the following four; (1)origin data with M and dT; (2)dD with M and dT; (3)NEWS; (4)SOFA; and fetched to parallel LSTM. The output and input ​​of parallel LSTM with patients' information are utilized ​​in the next layer, followed by the linear classifier and softmax layer. The binary cross-entropy loss is employed as loss function and optimized using Adam method. Resampling and weighted loss are used to handle class imbalance between sepsis and normal labels. The model was trained on a single hospital and tested on the other hospital, and the model selected in the previous test was additionally trained on both A and B dataset. The performance was evaluated on the scoring function given by Physionet Challenge 2019 with AUROC and AUPRC. We obtained AUROC, AUPRC and the score of 0.902, 0.236, and 0.552 from the datasets A and B. Finally, our model is tested on a blind testing dataset and achieved the score of 0.374.