Sepsis Prediction using Continuous and Categorical Features on Sporadic Data

Sepsis is one of the most prevalent causes of mortality in Intensive Care Units (ICUs) and also one of the most expensive health-care problems. Delayed treatment is associated with increase in death and financial burden. This work proposes an early prediction of sepsis validated on Physionet Challenge 2019 dataset. The challenge is to extract continuous, categorical and domain-specific discriminating features from highly sporadic lab data and vital signals. We find that the imputation of extremely isolated data lower the prediction performance. In order to mitigate this, we use a sliding window on sporadic data to generate continuous features which capture the trend. We also devise a binning approach to generate categorical features from the aperiodic data in order to discriminate the deviation from normalcy. Lastly, we observe that a logical fusion of Random Forest and Logit Boost provides optimal performance. Normalized Utility Score (NUS) is used to benchmark the performance of the proposed baselines. Five-fold cross-validation of the best preforming pipeline across the data reveals high median NUS of 0.401.