DEAR: Dual-Level Self-attention GRU for Online Early Prediction of Sepsis

Yu Zhao, Yike Wu, Mo Liu, Xiangrui Cai, Ying Zhang, Xiaojie Yuan

Abstract: Sepsis is one of the leading causes of death in intensive care units (ICUs). Online early prediction of sepsis has the potential for application to support clinicians. Despite the great success of deep neural networks in modeling electronic health records (EHRs), many architectures are incapable to be applied to online early prediction scenarios due to two major limitations. First, they overlook the earlier signs of the disease which are vital for the early prediction of sepsis. Second, they are unable to provide interpretation of prediction results. To tackle the above limitations, we propose a Dual-level sElf-Attention Gated Recurrent Unit Networks model, DEAR. On the one hand, DEAR is able to directly identify and strengthen the important time steps from the precedent memory. Specifically, the cell of DEAR straightforwardly fusion the history of both feature level and temporal level into the current step. On the other hand, DEAR provides interpretability with multi-head attention. Experimental results in the real-world sepsis dataset demonstrate that our model outperforms state-of-the-art methods in terms of both utility and balanced accuracy. In addition, the visualization of multi-head attention weights also indicates that DEAR reveals the importance of different time steps in the early prediction of sepsis onset.