Early Prediction of Sepsis via SMOTE Upsampling and Mutual Information Based Downsampling

Sepsis is a life-threatening response to infection that can lead to tissue damage, organ failure and death. The early prediction of sepsis is important, as it reduces undesirable patient outcomes associated with late-stage septic shock. However, effective early prediction is challenging, because the data is often heavily imbalanced against positive sepsis diagnosis. If the class imbalance is not addressed, models trained will tend to overfit in favour of the majority class, leading to degraded performance on the minority class. In this paper, we suggest a two-step method which consists of a mutual information based downsampling algorithm and a Synthetic Minority Over-sampling Technique (SMOTE), in order to effectively perform early prediction of sepsis. Our team, Kent Ridge AI (ranked 77th), obtained a utility score of -0.164 on the full test set by using the proposed two-step method. Additionally, we report crossvalidation results and identify several methods to improve performance.