# Machine Learning-Based Prediction of Surgical Intervention in Preterm Infants with Necrotizing Enterocolitis: A Retrospective Cohort Study

## TL;DR

This study uses machine learning to predict which preterm infants with NEC will need surgery, improving early identification and management.

## Contribution

A novel machine learning framework integrating clinical, lab, and imaging data to predict surgical intervention in preterm NEC infants.

## Key findings

- Neural Network achieved highest accuracy (0.875) in predicting surgical need in NEC infants.
- CRP, peritoneal irritation signs, and gestational age were top predictors identified via SHAP analysis.
- ML models offer objective decision support for early surgical intervention in preterm NEC patients.

## Abstract

Background: Necrotizing enterocolitis (NEC) is a life-threatening gastrointestinal disorder in neonates, particularly preterm infants. Early identification of infants requiring surgical intervention remains challenging due to nonspecific clinical manifestations and rapid disease progression. Methods: We conducted a retrospective cohort study of 320 preterm infants with NEC (gestational age <37 weeks) who were admitted to the NICU of the Capital Center for Children’s Health, Capital Medical University, Beijing, China, between June 2017 and December 2024. Forty-three clinical, laboratory, and imaging variables were collected. Feature selection was performed using LASSO regression and the Boruta algorithm. Four machine learning (ML) models—LightGBM, XGBoost, Random Forest, and Neural Network—were constructed. Model performance was evaluated using ROC-AUC, PR-AUC, accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and SHAP-based interpretability. Results: Among 320 infants, 119 underwent surgery and 201 received non-operative management. Thirteen consensus features were selected for modeling, including gestational age, CRP, lactic acid, peritoneal irritation signs, pneumatosis intestinalis, and hematologic parameters. The Neural Network achieved the highest overall classification performance (accuracy 0.875, sensitivity 0.824, specificity 0.903, balanced accuracy 0.863); Random Forest achieved the highest ROC-AUC (0.922), and XGBoost showed the highest PR-AUC (0.867). SHAP analysis identified CRP, peritoneal irritation signs, and gestational age as the most influential predictors. Conclusions: ML models integrating clinical, laboratory, and imaging variables can accurately predict the need for surgical intervention in preterm NEC patients. These models provide objective decision-support tools to improve early identification and optimize surgical management.