A Machine Learning Approach to Predict the Sepsis Status: Analyzing the Connection Between Relevant Laboratory Values and Other Physiological Measurements Obtained in Intensive Care Unit

Abstract

Sepsis is a life-threatening disease caused by the body's response to an infection, which leads to tissue damage, organ failure and death. According to literature, about 1.7 million Americans each year suffer from sepsis with approximately 270,000 fatal cases, and more than one third of mortalities in hospitals are associated with sepsis. It is estimated that worldwide about 30 million people suffer from sepsis each year, and 6 million patients with sepsis die. Above all, the cost of sepsis care is extremely high and constitutes a health risk mostly in developing countries. Having regard to the above, early diagnosis and therapy of sepsis play a key role in increasing the chance of cure for many patients. In this regard, many clinicians and researchers have proposed new definitions for sepsis (Singer et al. 2016), however the subject of early diagnosis methods and their practical usability and limits for sepsis remains to a large extent unexplored. Therefore, a cost-effective self-diagnostic technic has been proposed in this work, which is based on the XGBoost algorithm and deep convolution neural network model for sepsis prediction. Septic data were used for the model training, and the highest accuracy and the area under the ROC curve (AUC) obtained to evaluate the performance of the model were 96.49% and 70.78%, respectively. The novelty of this study lies in the investigations that unlike other studies are performed using vital signs determined independently from laboratory measurements. It has shown that the proposed model is faster and can provide more accurate results, and thus has great potential to be used as an efficient technique in predicting sepsis.