Early Diagnosis and Prediction of Sepsis Shock by Combining Static and Dynamic Information Using Convolutional-LSTM

Abstract

Deep neural network models, especially Long Short Term Memory (LSTM), have shown great success in analyzing Electronic Health Records (EHRs) due to their ability to capture temporal dependencies in time series data. In this paper, we proposed a general deep neural network framework which incorporates two additional components with the aim of improving LSTM. The first component, a Convolutional Neural Network (CNN), is added before LSTM to obtain local characteristics of EHRs. The second component, a fully connected neural network (FC), introduces static information (e.g., age) to LSTM, which is applied to handle dynamic information (e.g., lab result). The medical condition we aim to predict is septic shock – it is the most advanced complication of sepsis and is due to severe abnormalities in circulation and/or cellular metabolism. Our proposed framework was evaluated for two experimental tasks: visit level early diagnosis (left align) and event level early prediction (right align). Our results show that for visit level early diagnosis, by incorporating both CNN and static information, our framework consistently outperforms the original LSTM. For event level early prediction, the same outcome is observed when predicting < 5 hours into the future, however, when predicting ≥ 5 hours into the future, the addition of the CNN component alone obtains the best results.