Artificial Intelligence for Neonatal Sepsis Prediction: A Review of Advances, Clinical Integration, and Future Prospects.

内容摘要

Neonatal sepsis (NS) remains a cardinal cause of morbidity and mortality worldwide, disproportionately affecting neonates in low- and middle-income countries (LMICs). Blood culture, the current diagnostic reference standard, is burdened by prolonged turnaround times and the pervasive phenomenon of culture-negative sepsis, a limitation compounded by antenatal antibiotic exposure in mothers rather than representing an intrinsically poor tool. Conventional biomarkers, including C-reactive protein (CRP), procalcitonin, and lactate, are time-dependent in their rise and require integration with clinical context and professional experience to yield meaningful diagnostic value. These limitations create a persistent window of diagnostic uncertainty during which adverse outcomes may compound, and it is precisely in this gap that artificial intelligence (AI) offers supplementary clinical value. This structured narrative review synthesizes published literature on AI, machine learning (ML), and deep learning (DL) applications for NS prediction. Evidence was drawn from peer-reviewed studies identified through a structured search of the PubMed, Google Scholar, and IEEE Xplore databases, retrieving approximately 430 records in total, of which 44 studies met inclusion criteria following title and abstract screening and full-text review. The review incorporated both early-onset sepsis (EOS) and late-onset sepsis (LOS) prediction models, with particular attention to algorithm design, validation methodology, and clinical integration. AI-based models demonstrate substantial discriminative performance when evaluated alongside but not in place of existing clinical tools. Ensemble tree-based classifiers for EOS, including CatBoost (Categorical Boosting) and XGBoost (Extreme Gradient Boosting), report area under the receiver operating characteristic curves (AUROCs) exceeding 0.95 in retrospective cohorts (95% confidence intervals not reported in source studies). Continuous physiological monitoring models for LOS achieve AUROC values of 0.81 to 0.90, with early warning signals emerging 6 to 12 hours before clinical deterioration. The HeRO (Heart Rate Observation) monitor, a validated heart rate characteristics system, produced a statistically significant 22% relative reduction in mortality in a prospective randomized controlled trial. Nevertheless, most studies remain retrospective, single-center, and deficient in external validation, limiting confidence in real-world generalizability. AI holds transformative potential as a clinical decision-support adjunct for NS prediction, offering early risk stratification that complements rather than supplants the clinician's judgment and existing diagnostic tools. Realizing this potential demands prospective multicentric validation, context-sensitive model development for LMICs, advances in explainable AI, and robust regulatory and ethical frameworks. The goal is an equitable, globally applicable partnership between AI and the clinician.