Transfer learning prediction of surgical necrotizing enterocolitis in preterm infants without pneumoperitoneum on abdominal X-ray.

Background: Necrotizing enterocolitis (NEC) remains a leading cause of mortality in preterm infants, with 30-39% requiring surgical intervention. However, existing models for predicting surgical NEC lack accuracy and clinical utility, especially for infants without pneumoperitoneum on abdominal X-ray (AXR). In this study, we aimed to develop a prediction model to earlier identify NEC requiring surgical intervention. Methods: All preterm infants diagnosed with NEC (modified Bell's stage ≥ II) without pneumoperitoneum on AXR from Beijing Children's Hospital between January 2016 to December 2022 were retrospectively reviewed. Demographic, perinatal, clinical, laboratory, and imaging findings were analyzed. Six machine learning (ML) algorithms-logistic regression, decision tree, random forest, support vector machine, multilayer perceptron, and extreme gradient boosting-were trained and optimized via ten-fold cross-validation. The best-performing support vector machine model was further enhanced using transfer learning. The optimal algorithm was deployed into a web-based graphical user interface (GUI) for real-time risk stratification. Results: A total of 144 preterm infants with NEC without pneumoperitoneum on AXR were included in our study, including the surgical NEC group (n=31) and the medical NEC group (n=113). Multivariate analysis identified lower gestational age (P=0.010), pregnancy vaginitis (P=0.014), respiratory support (P=0.005), positive abdominal examinations (P<0.001), elevated C-reactive protein (P=0.003), and turbid peritoneal fluid on abdominal ultrasonography (P<0.001), as independent risk factors for surgical NEC. Then we constructed six ML models to predict surgical NEC by utilizing five variables derived from clinical, laboratory, and imaging findings in NEC-afflicted infants. Of all the models, support vector machine achieved perfect discrimination and superior reproducibility across training and validation sets. The transfer-learning model, built on the support vector machine base, achieved superior performance in the training set [area under the receiver operating characteristic curve (AUC) =0.964, 95% confidence interval (CI): 0.921-0.995] and validation set (AUC =0.937, 95% CI: 0.829-1.000). SHapley Additive exPlanations analysis highlighted positive abdominal examinations, turbid fluid on abdominal ultrasound, and bowel sounds grades as the top predictors. Furthermore, we developed a transfer-learning based GUI for the predictive model to facilitate clinical application. Conclusions: This study pioneered an interpretable ML framework integrating multimodal data to predict surgical NEC with near-perfect discrimination. Furthermore, the transfer-learning based GUI represented a transformative approach to optimizing surgical timing.