Early Sepsis Detection using Machine Learning and Neural Networks

Sepsis can cause overwhelming changes that impair multiple organs, leading to their failure and sometimes even having fatal results. Low blood pressure, difficulty in breathing, fever, mental confusion, and fast heart rate are some of the symptoms. Infections acquired in health care settings also frequently result in Sepsis. These infections affect hundreds of millions of patients globally every year and are one of the most recurrent adverse events during care delivery. Clinical conditions can rapidly deteriorate as healthcare-associated disorders are often impervious to antibiotics. During critical scenarios such as an intensive care unit, Early diagnosis and timely medical treatment of sepsis is crucial to give the patient prompt treatment since there is an increase in the mortality rate as each hour passes away in critical care. This paper proposes a method to predict the inception of sepsis 6 hours in advance using various machine learning and deep learning models and presents a comparative study of the same. The Medical Information Mart for Intensive Care III (MIMIC3) dataset was used to test the traditional machine learning methods such as RandomForest(RF), XGBoost, and also Deep Learning techniques such as neural networks and Autoencoders with XGBoost. The dataset contains an extensive range of parameters that are associated with laboratory, vitals, and demographics of patients which help them classify as sepsis and non-sepsis patients. A data pipeline was successfully created to clean the data, impute missing values and perform various feature engineering techniques. Our best performing model is a Deep Neural Network with an AUC ROC score of 0.888.