JMIR Publications

ABSTRACT

Background:

Neonatal sepsis (NS) continues to be a predominant cause of neonatal mortality in low- and middle-income countries (LMICs), accounting for over 550,000 deaths per year. Challenges in early diagnosis characterized by unspecific symptoms and delays in standard testing, highlight the necessity for innovative solutions. This study evaluates machine learning models for predicting newborn sepsis in Jordan, with the objective of addressing diagnostic gaps in resource-limited setting.

Objective:

This study evaluates machine learning models for predicting newborn sepsis in Jordan, with the objective of addressing diagnostic gaps in resource-limited setting.

Methods:

A retrospective cohort study investigated structured electronic health records (EHR) of 3,274 neonates admitted to a tertiary neonatal intensive care unit (NICU) in Jordan from 2018 to 2024. Three machine learning models—XGBoost (eXtreme Gradient Boosting), Decision Trees, and Neural Networks—were trained utilizing clinical, laboratory, and demographic data. Class imbalance was addressed via Synthetic Minority Oversampling (SMOTE), while overfitting was managed by regularization and cross validation. The evaluation of model performance was conducted by accuracy, sensitivity, specificity, and AUC-ROC (Area under the Receiver Operating Characteristic Curve).

Results:

XGBoost outperformed other models, achieving 94% accuracy, 0.98 AUC-ROC, and 98% sensitivity. Key predictors included C-reactive protein (CRP), platelet count, and gestational age. SMOTE improved sepsis case detection (recall: 0.98), while Decision Trees provided interpretable decision pathways. Neural Networks showed moderate performance (AUC: 0.81), limited by dataset size and hyper parameter sensitivity.

Conclusions:

XGBoost provides a strong, usable framework for predicting newborn sepsis in low- and middle-income countries, with the potential for enhancing early identification and antibiotic management. Future research should include multicenter validation and the incorporation of antibiotic resistance patterns to enhance therapeutic utility. This research develops a reproducible machine learning methodology for neonatal intensive care units in regions with limited resources, addressing major gaps in newborn healthcare.