Early Prediction of Sepsis using Machine Learning
Anuraag Shankar, Mufaddal Diwan, Snigdha Singh, Husain Nahrpurawala, Tanusri Bhowmick
Abstract
Sepsis is a fatal disease caused by infection. It has a significantly high mortality rate, particularly for patients in the ICU. The early and accurate detection of Sepsis is crucial as delayed treatment causes a sharp increase in the mortality rate. The proposed research aims to develop a classifier that accurately predicts Sepsis up to six hours before the clinical diagnosis of the disease. This is achieved using the patient's EMR, vital signs and demographics. The research shows several imputation techniques and proposes a new filling algorithm known as Mixed Filling. The main features contributing to the classifier's predictions have been described, thereby making the model more interpretable for medical personnel. Six models namely Random Forest, Logistic Regression, Light Gradient Boosting Machine, eXtreme Gradient Boosting, Neural Network and Long Short-Term Memory have been investigated for the classification of patients. The evaluation metrics that have been obtained are unprecedented and can be extremely useful for the timely and accurate prediction of Sepsis.