Identifying the Critical Parameters of Late-Onset Neonatal Sepsis Using Statistical and Machine Learning Methods

Background: Late-onset neonatal sepsis (LONS), which occurs 72 hours after birth, remains a major cause of neonatal morbidity and mortality, particularly in resource-constrained settings. Traditional diagnostic methods are often delayed or infeasible in such environments. This study investigated statistical and machine learning approaches to enhance the early prediction of LONS and support timely clinical decision-making. Methods: Neonatal data from the first admission to the neonatal intensive care unit (NICU) within five days of delivery were retrieved from the MIMIC-III database. After preprocessing, a total of 83 variables were retained. Fifteen key features were identified based on their clinical relevance and one-way ANOVA across seven different sepsis criteria frameworks. Machine learning techniques such as extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), and random forest were evaluated with and without applying SMOTE to correct class imbalance. Ensemble classifiers were developed, and their performance was assessed via 10-fold cross-validation and an independent test dataset. Results: Gradient boosting and LightGBM achieved an area under the receiver operating characteristic curve (AUROC) of 0.9269 and 0.9258, respectively, whereas random forest with SMOTE reached an AUROC of 0.9255. The final soft voting classifier, which combines the top three models, demonstrated robust performance, with an AUROC of 0.9271, an accuracy of 0.8521, and an F1 score of 0.7698. Conclusion: A machine learning–based ensemble approach using only 15 features can effectively predict LONS from clinical data within the first 24 hours of NICU admission. This data-driven model offers a scalable, real-time decision-support tool for neonatal care.