Development and validation of a screening model for early diagnosis of biliary atresia in neonates with cholestasis.

backgroundBiliary atresia (BA) is a progressive neonatal cholestatic liver disease that requires timely diagnosis and intervention. Differentiating BA from other causes of neonatal cholestasis remains a significant clinical challenge.

methodsIn this study, we retrospectively analyzed the clinical and biochemical data of 243 cholestatic neonates, comprising 61 with BA and 182 with non-BA. We utilized five supervised machine learning algorithms-logistic regression (LRM), decision tree (DET), multilayer perceptron (MLP), support vector machine (SVC), and random forest (RF)-to construct diagnostic models for BA. The performance of each model was evaluated based on its accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC). We then developed an online diagnostic tool based on the best-performing model.

resultsThe BA and non-BA groups showed significant differences across multiple biochemical markers. All five models demonstrated good diagnostic performance, with the random forest (RF) model achieving the best results (AUC = 0.93, sensitivity = 88.5%, specificity = 85.2%). The combination of multiple biochemical parameters substantially improved diagnostic accuracy compared to using single indicators. The web-based tool provides an intuitive and user-friendly interface to support early BA screening in clinical practice.

conclusionMachine learning-based models, particularly the RF model, show great potential for the early diagnosis of BA in cholestatic neonates. The implementation of a dedicated online platform may facilitate timely identification and assist clinicians in decision-making.