Using Vital Signs for the Early Prediction of Necrotizing Enterocolitis in Preterm Neonates with Machine Learning.

Background/objectives: Necrotizing enterocolitis (NEC), a devastating neonatal gastrointestinal disease mostly seen in preterm infants, lacks accurate prediction despite known risk factors. This hinders the possibility of applying targeted preventive therapies. This study explores the use of vital signs, including cerebral and splanchnic oxygenation, measured with near-infrared spectroscopy in early NEC prediction.

Methods: Several machine learning algorithms were trained on data from very preterm patients (<30 weeks gestational age). Time Series FeatuRe Extraction on the basis of scalable hypothesis tests (TSFRESH) extracted significant features from the vital signs of the first 5 postnatal days. We present the F1-scores and area under the precision-recall curve (AUC-PR) of the models. The contribution of separate vital signs to the selected TSFRESH features was also determined.

Results: Among 267 patients, 32 developed NEC Bell's stage > 1. Using a 1:4 NEC:control ratio, support vector machine and logistic regression predicted NEC better than extreme gradient boosting regarding the F1-score (0.82, 0.82, 0.76, resp., = 0.001) and AUC-PR (0.82, 0.83, 0.77, resp., < 0.001). Splanchnic and cerebral oxygenation contributed most to the prediction (40.1% and 24.8%, resp.).

Conclusions: Using vital signs, we predicted NEC in the first 5 postnatal days with an F1-score up to 0.82. Splanchnic and cerebral oxygenation were the most contributing vital predictors. This pioneering effort in early NEC prediction using vital signs underscores the potential for targeted preventive measures and also emphasizes the need for additional data in future studies.