Early Detection of Sepsis Using LSTM Neural Network with Electronic Health Record

Abstract

Early identification of sepsis may help in identifying possible risks and take the necessary actions to prevent more severe situations. We employed a recurrent neural network with Long Short-Term Memory (LSTM) and machine learning to identify the sepsis in its early stage. Sepsis can become a life-threatening disorder caused by the body’s response to infection, which results in tissue destruction, organ failure, and death. Every year, around 30 million people get sepsis, with one-fifth of them dying as a result of the disease. Early detection of sepsis and prompt treatment can often save a patient’s life. With the use of a Deep neural network, predict whether or not a patient has Sepsis Disease based on his or her ICU data. The objective of this study is to use physiological data to detect sepsis early. Patients’ data, such as vital signs, laboratory results, and demographics, are used as inputs. For the inference phase, we employed an LSTM to determine the best training hyperparameters and probability threshold. In this paper, an LSTM-based model for predicting Sepsis in ICU patients is proposed. We created a data pipeline that cleaned and processed data while also identifying relevant predictive characteristics using RF and LR approaches and training LSTM networks. With an AUC-ROC score of 0.696, RF is our top conventional classifier.