Early Diagnosis of Neonatal Sepsis Through Predictive Analytics and Feature Selection Techniques

Abstract

Goal 3.2 of the Sustainable Development Agenda aims to reduce the infant death rate by the year 2030. The leading causes of mortality in neonates are preterm and birth asphyxia, followed by neonatal infections. It is more probable for new-borns to get late-onset neonatal sepsis (LOS) from their surroundings than from their mothers. This kind of sepsis often manifests between 3 and 28 days of age. A difficult aspect of early LOS diagnosis is the lack of obvious clinical signs during the early stages of infection. Predicting LOS before obvious clinical signs is possible using physiological factors, according to studies. These metrics may be used as warning indications by clinicians to keep a careful eye on infants and act quickly to avoid problems and provide them good treatment.This research examines several machine learning algorithms that can forecast when new-born sepsis will start by analysing the MIMIC III dataset, which includes vital signs, laboratory results, and observations taken during the first 24 hours of arrival. Out of all the algorithms tested using 10-fold stratified cross-validation, the ones with the highest area under the receiver operating characteristic (AUROC) values were adaptive boosting (0.9248), light gradient boosting (0.9245), and random forest with Synthetic Minority Oversampling Technique (0.9238). With an AUROC of 0.9266, an accuracy of 0.8553, F1 score of 0.7829, and Matthew’s correlation score of 0.6995, the soft voting classifier trained on an ensemble of the most effective three models identified the beginning of newborn sepsis.