An LSTM-based Recurrent Neural Network for Neonatal Sepsis Detection in Preterm Infants

Abstract

Early and accurate neonatal sepsis detection (NSD) can help reduce mortality, morbidity and antibiotic consumption in premature infants. NSD models are often designed and evaluated in case control setups and using data derived from patient electrocardiogram (ECG) only. In this article, we evaluate our models in a more realistic retrospective cohort study setup. We use data from different modalities, including ECG, chest impedance, pulse oximetry, demographics factors and repetitive measurements of body weights. We study both the vanilla and long-short-term-memory (LSTM) Recurrent Neural Networks (RNN) architectures in a sequence to sequence mapping framework for NSD. We compare the performances of the models with logistic regression (LR) on a variety of classification metrics in a leave-one-out cross validation framework. The population we used contains 118 very low birth weight infants, among which 10 experienced sepsis. We showed that LSTM-based RNNs are both (1) more conservative and (2) more precise than LR or vanilla RNN, with a true negative rate at least +26% higher and a precision score of 0.16 compared to 0.06 for LR. This indicates that LSTM-based RNNs have the potential to reduce the false alarm rate of existing linear models, while providing a reliable diagnostic aid for neonatal sepsis.