A Multicenter Study on Formulaic Machine Learning Models for Prediction in Neonatal Necrotizing Enterocolitis: Application and Comparison of Kolmogorov–Arnold Networks

Objective To explore the application of formulaic machine learning models in the diagnosis and severity classification of neonatal necrotizing enterocolitis (NEC). Methods Clinical data of NEC infants admitted to NICUs of four hospitals in China from January 2021 to December 2024 were retrospectively analyzed. Three different NEC-related outcomes were studied: (1) conservative treatment vs. surgical intervention, (2) severe NEC, and (3) severe adverse outcomes. After identifying key variables for each outcome, we used the Kolmogorov–Arnold Network (KAN) model as the core and compared its performance and practicality with linear regression, logistic regression, and linear support vector machine (SVM) models. Results A total of 469 cases were included for outcomes (1) and (2), with 34 and 33 variables selected, respectively; 261 cases were included for outcome (3), with 33 variables selected. The top three most significant variables were as follows: 1. For conservative treatment vs. surgery: gastrointestinal perforation ( p  = 5.42654E-11), shock ( p  = 3.75733E-08), and duration of antibiotic use ( p  = 4.39668E-08); 2. For severe NEC: gastrointestinal perforation ( p  = 3.25699E-12), shock ( p  = 8.80686E-07), and duration of antibiotic use ( p  = 5.20998E-06); 3. For severe adverse outcomes: bronchopulmonary dysplasia (BPD) (p  = 1.74208E-15), duration of antibiotic use ( p  = 4.90394E-12), and gestational age ( p  = 1.19659E-11). The KAN model fitting formula performed the best in predictingthe treatment outcome(AUC: 0.917, AP༚0.687), the sevNEC outcome༈AUC༚0.851, AP༚0.537༉ and the severe adverse outcomes༈AUC༚0.823, AP༚0.8588༉. Conclusion As an emerging neural network model, the KAN demonstrates strong capabilities in handling nonlinear and complex data distributions. It offers high interpretability and facilitates NEC diagnosis and outcome stratification.