Abstract
BACKGROUND & AIMS: Neonatal necrotizing enterocolitis (NEC) remains a leading cause of morbidity and mortality in preterm infants. Current diagnostic methods, relying on clinical signs and radiography, often lack sensitivity for early detection. This study aimed to develop and validate a machine learning (ML) model integrating ultrasound and serological markers to improve NEC prediction in neonates. METHODS: This retrospective, case-control study included 191 neonates (cases with Bell's stage ≥ II NEC and matched controls) admitted to a tertiary NICU. Data were extracted from electronic medical records, including demographics, clinical variables, ultrasound findings (bowel wall thickness, edema, gas location, peristalsis, seroperitoneum), and serological markers (WBC, neutrophil count, CRP, ALP, albumin, procalcitonin, platelet count, INR, hemoglobin). Twelve ML algorithms were evaluated using 10-fold cross-validation on a training set (70%). The optimal model was selected based on AUC-ROC and further optimized via hyperparameter tuning. Model performance was assessed on an independent validation set (30%). Explainable AI (XAI) using SHAP values was employed to identify key predictive features. RESULTS: XGBoost demonstrated the highest performance (AUC = 0.97, 95% CI: 0.92-0.99) during cross-validation. The optimized XGBoost fusion model-Ultrasound combined Serological Predict NEC (USPN) achieved an AUC of 0.88 (95% CI: 0.76-0.99) in the validation set, with a sensitivity of 0.73 and specificity of 1.00. The USPN model outperformed models based solely on ultrasound (AUC = 0.73) or serological markers (AUC = 0.79). SHAP analysis identified bowel peristalsis, C-reactive protein, albumin, bowel thickness, and procalcitonin as the most influential predictors. Decision curve analysis demonstrated a positive relative net benefit of the USPN model compared to the US and serological models in the validation set. CONCLUSION: A machine learning model integrating ultrasound and serological markers significantly improves the prediction of NEC in neonates compared to single-modality approaches. This multimodal approach has the potential to facilitate earlier diagnosis and intervention, potentially improving outcomes in this high-risk population.