Time-Specific Metalearners for the Early Prediction of Sepsis

Motivation: Accounting for complex clinical dynamics in sepsis patients while aiming at an automated analysis of hourly (non-)validated data is challenging. The algorithm has to deal with imprecise, incorrect and incomplete data in addition to being time aware.Methods: We aimed to build time-specific stacked ensembles and a non-specific XGBoost learner to predict sepsis 6 hours prior to the sepsis onset. The models were trained on a triple split of 40,336 ICU stays taken from the training sets of the 2019 PhysioNet/CinC Challenge. Data was cleaned and features were built based on rolling windows including several clinical scores and criteria, such as shock index, qSOFA, SOFA, SIRS, NEWS, cNEWS. Model performance was evaluated using task-specific utility functions. Furthermore, variable importance was assessed.Results and conclusion: Although no official score was obtained in the Challenge as team Sepsis2G, we found normalized utility score of 0.394 for our non-specific XG-Boost model on a held out subset of the training data. The threshold selection was displaced in time-specific meta-learners leading to an inferior performance. Most important variables included the assumed presence of ventilation, white blood cell count, partial thromboplastin time, blood urea nitrogen and rolling quantiles of the temperature. Partial SOFA-scores, cNEWS, and the shock index showed major importance in the ICU admission phase.