245 Inflammatory Mediators as Predictors of Neonatal Sepsis

Despite the frequency and severity of sepsis in neonatal intensive care units, diagnosing infections remains difficult and time consuming and lacks sensitivity. Evaluating clinical data in conjunction with serial inflammatory mediators may improve detection, ideally before clinical signs and symptoms are present. We designed a pilot study to determine whether a complex computational model evaluating changes in inflammatory mediators combined with clinical data could assist in early detection of neonatal infection. Eligible infants were < 72 hours old and were ≥ 37 weeks EGA (term) or ≤ 32 weeks EGA (preterm). Infants were enrolled for 28 days or until discharge. The plasma from spun hematocrit tubes was evaluated for 165 inflammatory mediators (including IL 1-18, MMPs, CRP, ENA-78, RANTES, TNF, G-CSF, GM-CSF, ICAM-1, TIMP-1, and others) using a chemiluminescent bead assay. Based on clinical and traditional laboratory data, infants were classified daily as having no infection, suspected infection, clinical infection without positive cultures, and culture-proven infection. Changes in inflammatory mediators were correlated with this clinical classification by creating a Bayesian network classifier over the clinical data, using a leave-one-out cross-validation methodology. Fifty-six infants were enrolled, 17 term (39 ± 1.2 wks, 3,242 ± 509 g) and 39 preterm (29 ± 2.2 wks, 1,267 ± 389 g). Nine infants were excluded from analysis (2 due to incomplete sample data and 7 due to incomplete clinical data). Nineteen episodes of positive blood, urine, or wound cultures in conjunction with clinical signs of infection were diagnosed in 9 infants (2 term). Thirteen others had clinical evidence of infection but negative cultures (7 term). Results based on a single application of the initial algorithm using a combination of clinical variables (RR, BP, temperature) and mediators (ENA-78, CRP) identified profiles capable of separating infected from noninfected cases with an overall accuracy of 94%. These preliminary results suggest that, using clinical data and a broad protein array, complex computational analyses may play a role in distinguishing early severe infections in newborns. Supervised machine learning techniques previously developed for microarray analysis are being employed to build and evaluate a classifier to distinguish normal samples from samples related to infection. Further evaluations using temporal computational analysis are ongoing.