Developing an Early Warning System for Sepsis

Sepsis is a life-threatening condition that is caused by infection, and is estimated to affects an estimated 1.7 million adults in the United States and contributes to 265,000 deaths annually. Identifying sepsis before it happens and treating it earlier leads to decreased mortality and decreased lengths of stay. As part of the PhysioNet/Computing in Cardiology Challenge 2019, we developed an ensemble-based approach for the early detection of sepsis in ICU patients. Our final model predicted sepsis using the previous 24 hours of data, and consisted of a combination of two convolutional neural networks and a random forest trained on different subsets of the data. In training our models, we experimented with random undersampling and cluster-based undersampling as a means for addressing severe class imbalance. On validation data, our final model achieved a utility score of 0.432 on hospital A (AUROC: 0.794, AUPRC: 0.101), 0.247 on hospital B (AUROC: 0.816, AUPRC: 0.056), and a utility of 0.375 on combined data from both hospitals (AUROC: 0.809, AUPRC: 0.089). On the heldout test data, the model obtained a utility score of 0.378.