Early Prediction of Sepsis from Clinical Data: The PhysioNet/Computing in Cardiology Challenge 2019

Abstract

The PhysioNet/Computing in Cardiology Challenge fo-cused on the early detection of sepsis from clinical data. A total of 40,336 patient records from two distinct hospital systems were shared with participants while 22,761 patient records from three distinct hospital systems were sequestered as hidden test sets. Each patient record contained up to 40 measurements of vital sign, laboratory, and demographics data for over 2.5 million hourly time windows and over 15 million data points. We used the Sepsis-3 clinical criteria to define the onset time of sepsis. We challenged participants to design automated, open-source algorithms for predicting sepsis 6 hours before clinical recognition of sepsis. We developed a novel, clinical utility-based evaluation metric to assess each algorithm that rewards early sepsis predictions and penalizes late or missed predictions and false alarms. A total of 104 teams from academia and industry submitted a total of 853 entries during the official phase of the Challenge. We accepted 90 abstracts based on Challenge entries for presentations at Computing in Cardiology. We also compared entries to ensure that approaches from different teams remained independent. This article presents our analysis and discusses the implications of the Challenge for early sepsis predictions and related sequential prediction tasks.