MACHINE LEARNING MODELS FOR EARLY PREDICTION OF NEONATAL SEPSIS: PERFORMANCE ANALYSIS AND CLINICAL UTILITY ASSESSMENT

Abstract

Background: Neonatal sepsis remains a leading cause of neonatal mortality, with early recognition critically impacting survival outcomes. Objective: To evaluate the diagnostic and predictive performance of machine learning (ML) models for early neonatal sepsis identification compared to conventional clinical screening tools. Methods: A prospective observational study was conducted in a tertiary neonatal intensive care unit (NICU) over 24 months. ML models including Random Forest, XGBoost, and Long Short-Term Memory (LSTM) networks were trained on electronic health record data. Results: Among 2,847 neonates, 412 (14.5%) developed confirmed sepsis. The LSTM model achieved the highest predictive performance (AUC-ROC 0.94, sensitivity 88.3%, specificity 93.7%), outperforming conventional sepsis screening scores (AUC-ROC 0.71). Time to diagnosis was reduced by a mean of 6.2 hours. Conclusion: ML-based early warning systems demonstrate superior performance to conventional screening tools for neonatal sepsis prediction, with clinically significant reductions in time to diagnosis.