Utilizing Informative Missingness for Early Prediction of Sepsis

Abstract

Aims: Physicians have to routinely make crucial decisions about patients’ health in the ICU. Sepsis affects about 35% of ICU patients, killing approximately 25% of the afflicted. In this paper, we aim to predict the occurrence of sepsis early by studying the missingness of physiological variables and using it with the overall trends in data.Methods: We chose XGBoost as our base model and tried several variations by changing hyperparameters, window sizes and imputation methods. To further improve the model, we used masking vectors to represent the missingness of features in the dataset. Additional modifications include shifting the Sepsis Label to earlier time steps and tuning the classification probability threshold to further improve the model’s performance.Results: The XGBoost model with a sliding window of size 5, no imputation, utilizing informative missingness of all temporal variables and trained on labels shifted by 3 hours before toptimal, achieved a Utility Score of 0.337 on the full test set. We identified as "CTL-Team" in the challenge and were officially ranked 5th on the basis of this score.