Machine Learning Approaches to Predict Late-Onset Neonatal Sepsis in the Neonatal Intensive Care Unit

Background and aims: Neonatal sepsis is a major cause of neonatal morbidity and mortality andrefers to an invasive infection, usually bacterial, involving the bloodstream. Late-onset neonatalsepsis occurs in infants between ۷ and ۲۸ days old. The primary challenge in neonatal sepsis isits evasive signs and symptoms which makes the diagnosis and prognosis burdensome. The onlyunrivaled quick fix is blood culture confirmation which takes virtually two days to generate results.On the other hand, in order to diminish neonatal mortality, early medical treatment must beinitiated promptly. Thus, early and accurate prediction of sepsis ensures correct and prompt antibiotictreatment and minimizes diagnostic uncertainty. This research aimed to implement machinelearning models that can efficiently predict neonatal sepsis based on the features.Method: The target population was ۴۵۵ cases of neonates aged between ۷ and ۲۸ days admittedto the Neonatal Intensive Care Unit of the Maternal, Fetal, and Neonatal Research Center, Vali-Asr Hospital, affiliated to Tehran University of Medical Sciences during the period from ۲۰۱۶to ۲۰۱۹ due to their suspicion of late-onset sepsis. Each neonate record included ۲۱۳ demographicand laboratory features. Machine learning models including Support Vector Machines, DecisionTrees, Random Forest, K Nearest Neighbor, Gradient Boosting, and XGBoost were employed onthe provided dataset to predict late-onset sepsis. Parameters were selected based on grid searchand results were evaluated based on the area under the receiver operating curve and the best F۱score metric on the validation set. Statistical analysis was performed using Python ۳.۶.Results: In order to predict late-onset sepsis, ۴۲ features out of ۲۱۳ available features were selectedusing the decision tree classifier. Some features such as neonatal age (day) of full oralnutrition starting time, anemia, diagnosis of sepsis by the physician, fresh frozen plasma, meancorpuscular volume, respiratory distress syndrome, patent ductus arteriosus, surgical intervention,blood type, blood phosphorus, sodium, magnesium, and potassium levels, intravenous immunoglobulinintervention, intervention after resuscitation, type of nutrition, and neonatal age (day) ofphototherapy starting time can be mentioned in the headline. The results revealed that RandomForest and XGBoost performed highly at the prediction of late-onset sepsis with the area underthe receiver operating curve of ۹۴.۹% and ۹۳.۶%, respectively. These algorithms also had thehighest F۱ scores of ۰.۸۷ and ۰.۸۴, respectively.Conclusion: Outcomes obtained from this research showed the promising potential of RandomForest and XGBoost in predicting late-onset sepsis based on demographic and laboratory datawhen blood culture results are not still available or without a result. These results can benefit theclinicians in appropriate management of neonatal sepsis so as not only to reduce sepsis-relatedmortality in sick neonates but also to prevent the misuse of antibiotics in healthy newborns. However,this research can be extended by collecting patient vital signs data to feed more data to themachine learning models.

Keywords:

machine learning , neonatal mortality , sepsis