Subtle changes in vital signs and their interactions occur in preterm infants prior to overt deterioration from late-onset septicemia (LOS) or necrotizing enterocolitis (NEC). Optimizing predictive algorithms may lead to earlier treatment.
For 1065 very low birth weight (VLBW) infants in two NICUs, mean, SD, and cross-correlation of respiratory rate, heart rate (HR), and oxygen saturation (SpO 2) were analyzed hourly (131 infant-years’ data). Cross-correlation (co-trending) between two vital signs was measured allowing a lag of +/− 30 seconds. Cases of LOS and NEC were identified retrospectively (n=186) and vital sign models were evaluated for ability to predict illness diagnosed in the ensuing 24h.
The best single illness predictor within and between institutions was cross-correlation of HR-SpO 2. The best combined model (mean SpO 2, SD HR, and cross correlation of HR-SpO 2,) trained at one site with ROC area 0.695 had external ROC area of 0.754 at the other site, and provided additive value to an established HR characteristics index for illness prediction (Net Reclassification Improvement 0.25, 95% CI 0.113, 0.328).