Recurrent Neural Networks for Early Detection of Late Onset Sepsis in Premature Infants Using Heart Rate Variability

Abstract

Early diagnosis of late onset sepsis (LOS) in premature infants can help reduce morbidity and mortality in this particularly vulnerable population. In this work, we propose a machine learning model based on recurrent neural networks for the early detection of late onset sepsis in premature infants. The model combines gated recurrent units and long short-term memory units, and uses heart rate variability features as input data. The population used for this study consisted of 259 premature infants; 193 of them were used for training the model, which was then tested in the remaining 66 infants. Thus, we obtained an area under the receiver operating characteristics curve (AUROC) of more than 80% for the 24 hours before the onset of the infection, and reaching 90.4% (95% CI [88.1%, 92.6%]) six hours before the time of the infection. The proposed method has the potential to be easily implemented as a decision support system for real-time LOS detection in neonatal intensive care units, as it uses only data which is continuously and automatically available in such settings.