(884) Risk Factors for Necrotizing Enterocolitis in Congenital Heart Disease Using Machine Learning

Introduction: Necrotizing enterocolitis (NEC) is a severe complication in neonates with congenital heart disease (CHD) with mortality approaching 25%. While many risk factors for NEC have been evaluated, studies are marred by small numbers of patients with rare outcome. We sought to apply machine learning and artificial intelligence methodology to a dataset from the multi-institutional Pediatric Cardiac Critical Care Consortium (PC4) to identify risk factors for post-operative NEC.

Methods: We analyzed perioperative PC4 data from 2/1/19 to 2/1/24 for neonates with CHD and index operation at ≤ 30 days of life. The 360 patients with NEC stage II or higher and 1080 controls were randomly divided into training (75%) and testing (25%) sets. The machine learning BorutaShap package was applied for selection of predictors out of 61 features. A XGBoost classifier was then trained with Bayesian hyperparameter tuning to optimize model performance and predict the risk of post-operative NEC.

Results: Univariate analysis revealed 29 significant variables associated with NEC. BorutaShap identified 15 variables that significantly contributed to the model: race (white or black), STAT category, antenatal diagnosis, gestational age at birth, birth weight, age at surgery in days, weight at surgery, cardiopulmonary bypass time, cross clamp time, total cerebral perfusion time, circulatory arrest time, low cardiac output syndrome, and any episodes of sepsis. In the testing set, the prediction model had an area under the receiver operating curve (AUC) of 0.80, with good sensitivity (0.69) and specificity (0.73) and negative predictive value (0.88). Calibration slope was 1.36 and observed/expected ratio was 1.006 indicating moderate to good model calibration.

Conclusions: This is an analysis of the largest cohort of postoperative NEC in neonates following congenital heart surgery. Our model can help predict the occurrence of post-operative NEC based upon 15 risk factors. Most features were non-modifiable, but future studies could include monitoring and risk category based on these factors. Further analysis of this dataset aims to identify risk factors for pre-operative NEC and perform neural network analysis to identify modifiable risk factors to reduce the incidence of NEC.