Early Prediction of Sepsis by Heterogeneous Event LSTM Model

Bingchan Zhao, Zhide Wei, Tianyi Liu, Jing Xu, Ming Zhang
Peking University


Background: Heterogeneous Event LSTM(HE-LSTM) has been proposed to learn from the joint representation of heterogeneous temporal health records and make clinical endpoint prediction. This study evaluates the performance of HE-LSTM on early prediction of Sepsis when dealing with unbalanced positive samples and negative samples. Methods: we evaluate the model based on full data and shortened data which removes all the data corresponding to label 1, we show that while HE-LSTM works well on the prediction given the full data, it can also perform well on the shortened data, which shows that it is practical in the real-world application for early prediction of Sepsis. Results: after training on 1000 samples with 30% positive samples, HE-LSTM achieves AUROC 0.702, AUPRC 0.053, Accuracy 0.978, F-measure 0.035, Utility 0.017 on our test data, after we mark the patient as positive 6 hours before the ground truth label to encourage early prediction, HE-LSTM achieves AUROC 0.731, AUPRC 0.074, Accuracy 0.972, F-measure 0.075, Utility 0.048 on test data, and we submit the latter entry for Phase 1. When we test HE-LSTM on shortened data, it can also early predict 0.03% of the rows. Conclusions: HE-LSTM performs well on learning from the heterogeneous temporal health records, although it still needs to be improved to deal with the shortened data in real-world medical practice.