Early prediction of sepsis using LSTM networks

Abstract

Sepsis is a severe life-threatening disease that causes millions of deaths every year. As a significant public health problem, the annual costs of hospital treatment of sepsis patients are rising. Without the treatment of antibiotics for sepsis, the risk of mortality increases with each additional hour. It is desirable to diagnose sepsis as early as possible, ideally earlier than the doctors can identify it with traditional methods. Many existing machine learning approaches show excellent results for onset sepsis prediction. However, it is important to perform early prediction of sepsis (i.e., hours before it takes place), especially in the case of intensive care unit patients. In this thesis, I train a Long Short-Term Memory (LSTM) network for the early prediction of sepsis using an existing dataset published as part of the PhysioNet/Computing in Cardiology Challenge in 2019. I show that adding a variety of predictors based on existing hourly recorded variables significantly improves the prediction results. I also show how tuning the number of epochs and observing the learning curves can lead to better outcomes for sepsis’s early prediction. Finally, I show that using data standardization destroys the diversity and variety in the data and weakens the early prediction outcomes. My best model achieves a utility score of 0.421, an F-score of 0.128, and an area of 0.129 under the precision-recall curve on the publicly available data. The obtained standardized utility score is slightly higher than those reported in some published papers, which use Long Short-Term Memory networks. However, for clinical application, improved approaches would be needed to produce more reliable results.