Artificial Intelligence-Based Predictive Model For Early Detection Of Neonatal Sepsis

Abstract

Neonatal sepsis is a life-threatening condition contributing significantly to newborn mortality worldwide. In 2020, approximately 2.3 million of the 5 million under-five child deaths occurred in the neonatal period, with infections (including sepsis) among the leading causes . Early identification of sepsis before overt clinical symptoms is challenging, but machine learning (ML) approaches can detect subtle risk patterns in vital sign and laboratory data . In this study, we develop an AI-driven predictive framework that processes NICU patient data (vital signs and lab values from datasets like MIMIC-III and PhysioNet) to forecast impending sepsis. We apply preprocessing (imputation, normalization) and feature selection, then train multiple classifiers including convolutional neural networks (CNN), support vector machines (SVM), and ensemble models. In simulated experiments on retrospective neonatal ICU data, the CNN achieved 90% accuracy, 88% sensitivity, 92% specificity, and an ROC area of 0.95. A voting ensemble of top models improved performance (accuracy 91%, AUC 0.96). These results demonstrate that AI models can reliably flag high-risk neonates hours before clinical onset, potentially enabling timely intervention.