Medical Decision Support using Machine Learning for Early-Onset Neonatal Sepsis Prediction

Background and aims: The early neonatal period, which extends from birth to the seventh day oflife, is the most dangerous period for a neonate, with an increased risk of morbidity and mortalityfrom early-onset sepsis mainly due to bacteria acquired before and during delivery. Forasmuchas these are newborns, the procedure of diagnosis and treatment is extremely difficult. Given thatsepsis management is highly time-sensitive, early prediction of sepsis before its onset in newbornpatients is crucial in preventing mortality as it gives clinicians additional lead time to plan andexecute treatment plans. Thus, this research aimed to investigate the predictive value of machinelearning models of demographic and laboratory data to detect early-onset sepsis.Method: Datasets were collected from the medical records of ۴۵۹ neonates up to ۷ days of lifehospitalized in the Neonatal Intensive Care Unit of the Maternal, Fetal, and Neonatal ResearchCenter, Vali-Asr Hospital, affiliated to Tehran University of Medical Sciences during the periodfrom ۲۰۱۶ to ۲۰۱۹ due to their suspicion of early-onset sepsis, in which patient names were eliminatedin order to have an anonymized dataset. The neonates were evaluated for ۲۱۳ demographicand laboratory features. The performance of the machine learning models including SupportVector Machines, Decision Trees, Random Forest, K Nearest Neighbor, Gradient Boosting, andXGBoost were investigated for predicting early-onset sepsis. Parameters were selected based ongrid search and results were evaluated based on the area under the receiver operating curve andthe best F۱ score metric on the validation set. Statistical analysis was performed using Python ۳.۶.Results: Based on the decision tree classifier, ۵۷ features out of ۲۱۳ were selected as the mostessential features in the diagnosis and prediction of early-onset sepsis. Some of these features arecontinuous positive airway pressure therapy, neonatal age (day) of oral nutrition starting time,premature rupture of membranes, corticosteroid administration in pregnancy, low birth weight,hematocrit test, hemoglobin concentration, umbilical vein catheterization, neonatal jaundice, absolutemonocytes and lymphocytes count, blood calcium and potassium levels, red blood celldistribution width, mean platelet volume, respiratory distress syndrome, and Apgar score at fiveminutes. As per findings, Support Vector Machines and Random Forest gave the best results ascompared to the other models to predict early-onset sepsis with the area under the receiver operatingcurve of ۹۸.۴% and ۹۷.۶%, respectively. Both algorithms had F۱ scores of ۰.۹۴.Conclusion: The results demonstrated that Support Vector Machines and Random Forest canhelp identify early-onset sepsis hours prior to clinical recognition regarding demographic andlaboratory data while screening a large portion of negative cases and may therefore be valuableas a medical decision support tool. Thus, these models could be implemented in the hospital inorder to significantly reduce the in-hospital mortality rate, unnecessary hospital stay, and cost oftreatment. Further prospective research is warranted to assess machine learning models usingvital signs data to improve the accuracy of antibiotic use in the management of neonatal sepsis.

Keywords:

decision tree classifier , early prediction , sepsis