A High-Performance Transfer Learning Model for Early Screening of Biliary Atresia in the Neonatal Period

Background Biliary atresia (BA) is a life-threatening cause of neonatal cholestasis, and early recognition is essential to enable timely Kasai portoenterostomy. However, clinical diagnosis during the neonatal period remains challenging due to the overlap of symptoms with other cholestatic disorders. Methods We retrospectively analyzed 268 cholestatic neonates (≤ 30 days old) evaluated at Beijing Children’s Hospital between 2017 and 2025, including 86 BA and 182 non-BA infants. Clinical features and routine biochemical indicators were collected. Traditional machine learning (ML) models and a transfer learning (TL) model were developed; model performance was assessed by receiver operating characteristic curves and standard classification metrics. Model interpretability was evaluated using SHAP. A graphical user interface (GUI) was created for clinical deployment. Results Multivariate analysis identified acholic stools, gamma-glutamyl transferase (GGT), total bilirubin (TBIL), and total bile acids (TBA) as independent predictors of BA. LASSO selected six key variables (sex, acholic stools, ALB, GGT, TBIL, TBA) for ML model construction. Among traditional algorithms, Random Forest showed the best performance. The TL model demonstrated improved discrimination, achieving an AUC of 0.987 in the validation cohort. SHAP analysis confirmed that acholic stools, GGT, and TBIL contributed most to BA prediction. A GUI prototype was developed to facilitate convenient clinical use. Conclusion This study suggests that a TL-based model integrating common clinical features and routine biochemical markers may assist in early identification of BA among neonates. Although the internal performance was promising, external validation is needed to assess its broader applicability and clinical utility.