439 - Just-in-time Risk Stratification to Expedite Sepsis Evaluation and Treatment

Publication Number: 2427.439

Luke Viehl, Washington University in St. Louis School of Medicine, Saint Louis, MO, United States; Nankee K. Kumar, St. Louis Children's Hospital, St Louis, MO, United States; David B. Gootenberg, Washington University in St. Louis, St. Louis, MO, United States; Zachary A. Vesoulis, Washington University School of Medicine, St. Louis, MO, United States

Background: Sepsis evaluation in NICUs requires balancing early antibiotic administration against unnecessary treatment in low-risk infants. Current practice relies on clinical judgment alone, leading to variable responses. Objective risk stratification tools could optimize resource allocation and antibiotic stewardship.

Objective: To determine whether an AI sepsis algorithm can provide risk-stratified sepsis evaluation on top of subjective clinical concern for late onset sepsis (LOS).

Design/Methods: We retrospectively studied infants undergoing sepsis evaluation in our NICU 11/24-09/25. LOS was defined as a positive blood culture and 5 days of antibiotics or a negative culture with a sepsis-like syndrome treated with 5 days of antibiotics. All other events were classified as no sepsis. We collected demographics, time to antibiotic administration, culture results, and outcomes.

POWS-an AI-generated score derived from heart rate characteristics and oxygen saturation patterns-was calculated in the hour preceding sepsis evaluation. Using ROC analysis, we identified optimal cutoffs for sepsis prediction. We compared a two-tier stratification system against clinical suspicion alone and examined gestational age as a risk modifier.

Results: We analyzed 197 sepsis evaluations from 140 patients (mean GA 30.5±5.1 weeks, mean BW 1660±950g). Confirmed sepsis occurred in 22 cases. POWS demonstrated strong discrimination (Fig.1, AUC=0.747, p< 0.001), with optimal cutoff of 2.8 providing 73% sensitivity and 79% specificity. Two-tier stratification revealed marked risk separation: Events with a standard risk (POWS < 2.8) had a 4.1% sepsis rate-below general NICU baseline-while events with an elevated risk score (POWS≥2.8) showed a 30.8% rate, triple that of high-risk VLBW populations (Fig. 2).

Among elevated-risk infants, extremely preterm infants ( < 28 weeks GA) had 33% sepsis rate versus 27% for more mature infants, though this difference was not statistically significant (p=0.529). Median time-to-antibiotics was 61 minutes overall, with no difference between risk tiers, suggesting protocol-driven responses rather than risk-based urgency.

Conclusion(s): An AI-based physiologic scoring objectively identifies NICU infants at standard versus elevated sepsis risk compared to subjective evaluation alone. This stratification enables clinicians to prioritize rapid evaluation and antibiotic administration for elevated-risk infants (30.8% sepsis rate) while maintaining appropriate vigilance for standard-risk cases (4.1% rate).

Figure 1. POWS Score by Sepsis Outcome
Confirmed sepsis events (culture positive or clinical sepsis receiving 5 days of antibiotics had statistically higher POWS scores in the hour prior to evaluation compared to no sepsis events.

Figure 2. Sepsis Rate by Risk Category
Using the optimal cut point value of 2.8, POWS can differentiate between infants at below population average risk of true sepsis (4.1%) vs. those at markedly elevated risk of sepsis.

Tables 1 and 2
POWS performance metrics and a proposed clinical decision support tool