DeepLOS: Deep learning for late-onset sepsis prediction in preterm infants using heart rate variability

Abstract

Late onset sepsis (LOS) is one of the main causes of death in preterm infants in a neonatal intensive care unit (NICU). LOS can be better treated with early detection, reducing its morbidity and mortality. In this study, an end-to-end deep learning model called DeepLOS was developed to predict LOS in preterm infants in a NICU. The model is based on a residual convolutional neural network (ResNet) with feature map (or channel) attention and uses RR intervals (i.e., interbeat intervals) as input. The model was trained and tested on a dataset composed of 128 preterm infants (60 blood-culture-proven LOS patients and 68 control patients). To minimize the possible age effect on modeling, we also considered an age-matched dataset including 32 LOS and 32 control patients from the full dataset. Prediction was done with a one-hour (non-overlapping) sliding window from 24 h before LOS to onset of LOS. We used 5-fold patient-independent cross validation and F-score to evaluate the model performance. The DeepLOS achieves an F-score of 0.72 for the full dataset and 0.73 for the matched dataset in LOS prediction for all one-hour segments, outperforming the baseline ResNet model without channel attention. F-score is generally higher (>0.75) when coming closer to the onset of LOS. Our study demonstrates the feasibility of deep learning for end-to-end LOS prediction in preterm infants. Furthermore, the model uses readily available RR intervals as input only; and is therefore vendor-processing independent and has the potential to be easily deployed in different NICUs.