An Ensemble of Bagged Decision Trees for Early Prediction of Sepsis

Abstract

Sepsis is a serious medical condition caused by the body's response to an infection. Early prediction and treatment of sepsis are critical. In response to the PhysioNet/CinC Challenge 2019, we developed an algorithm for early prediction of sepsis. Three datasets provided by the challenge are from ICU patients in three separate hospitals, two of which are publicly available to the participants, but the third is hidden and used for scoring. Data are highly unbalanced and contain many missing values. Each patient's data comprises hourly collected samples of 40 features. We preprocessed the data by a plausibility filter eliminating the outliers, forward filling of the missing data and replacing the remaining by population mean, and standardization of the numerical data. We developed an ensemble of bagged decision trees with a highly unbalanced misclassification cost to predict the sepsis for each sample of features in a patient. The classifier was trained on the first hospital dataset and validated on the second hospital dataset. A total of 15 important features and a set of hyperparameters were selected in an iterative training approach. Nine entries were submitted for evaluation of the utility score on a subset of the hidden dataset. The entry with the best utility score (0.335) was selected for running on the full test dataset and the final utility score was ??.