Evgenios Vlachos, Aris Salapatas Gkinis, Vasileios Papastergiou, Christos Tsitsipanis, George Giannakopoulos

Abstract:
Sepsis is currently defined as a “life-threatening organ dysfunction caused by a dysregulated host response to infection”. The early detection and prediction of sepsis is a challenging task, with significant potential gains regarding the lives of patient and — as such — should be researched comprehensively. The main goal of this study is to take anonymised and appropriately processed data in order to detect infections which imply future probability for sepsis. In that way, medical practitioners may have the opportunity to treat patient appropriately in a proactive manner. Feature selection techniques were applied in order to define the most important features to feed machine learning models and maximize the performance of the prediction as a binary classification problem. We also aim to highlight the relation of specific clinical input features to the prediction outcome, possibly contributing to an improved, data-driven understanding of this multi-factorial dysfunction. Early findings indicating promising classification performance, with different machine learning algorithms, but also based on appropriate feature engineering, building upon features with a time-sensitive aspect (i.e. features representing different samplings in different positions in time).