A machine learning system enables just-in-time risk-stratified sepsis evaluations in the neonatal intensive care unit

NICU sepsis evaluation balances rapid antibiotic administration for mortality reduction against unnecessary treatment. Objective risk stratification optimizes resource allocation.

We retrospectively studied 191 sepsis evaluations from 136 NICU patients. Sepsis was defined as positive blood culture or clinical sepsis with ≥5 days of antibiotics. A machine learning score (POWS) used cardiorespiratory data 1 h prior to evaluation to quantify sepsis risk. We performed ROC analysis, identified optimal thresholds by Youden Index, and assessed postmenstrual age (PMA) and gestational age (GA) as effect modifiers.

POWS was higher in sepsis versus no sepsis (4.53 vs. 2.14, p < 0.005; AUC 0.744 [95% CI: 0.608-0.870]). A POWS threshold of 2.9 provided 71% sensitivity and 81% specificity for detection, stratifying into groups with a sepsis incidence of 4.2% vs. 31.9%. PMA, but not GA, modulated POWS (-0.054 points/week, p < 0.0001).

POWS discriminates sepsis risk at the time of evaluation.

A machine learning-developed score (POWS) provides real-time risk stratification during sepsis evaluations, distinguishing a truly low-risk group (4.2% sepsis rate) from a high-risk group (31.9% sepsis rate), among infants with clinical concern for infection. Postmenstrual age at evaluation significantly modulates POWS's performance but birth gestational age does not, suggesting that current developmental maturity is more important than historical prematurity when interpreting and utilizing POWS. Prospective implementation with real-time POWS display may enable risk-based triage thereby prioritizing resources to pursue urgent workup and early antibiotic administration in high-risk infants.