Automated Prediction of Sepsis Onset Using Gradient Boosted Decision Trees

Abstract

In this study, we developed an automatic algorithm that predicts onset of sepsis using hourly clinical data from patients in an ICU setting. We participated as team "Sepsyd" in the PhysioNet/Computing in Cardiology 2019 Challenge and were ranked 2nd with an official final test score of 0.345. Our developed system processed all the clinical input variables provided in the Challenge. We first applied a preprocessing step that applied a log transform to selected variables and imputed missing values of the variables. After preprocessing, a feature set was formed including the 40 preprocessed variables, 34 missing value flags, the changes in the time series in the vital signs variables and the variance of the vital signs variables. Following this, the features of the present hour were combined with the features of the past 5 to 8 hours of data. These combined features were then processed with a gradient boosting tree classifier to estimate the likelihood of a positive sepsis classification at each time step. We compared the utility score of a number of different system configurations using 3-fold cross validation on the training data. Our best system, assessed on the test set, used a maximum tree depth of 4, a look back of 5 hours, and processed the clinical input variables combined with the missing value flags.