An interpretable machine learning model for outpatient referral risk assessment in pediatric biliary atresia after Kasai portoenterostomy.

OBJECTIVES: To develop an interpretable model using routine outpatient data to assist risk stratification and referral assessment in children with biliary atresia (BA) after Kasai portoenterostomy (KPE). METHODS: This multicenter retrospective study included 512 children with type III BA who underwent KPE across eight pediatric centers. Clinical, laboratory, and ultrasonographic variables collected during outpatient follow-up were analyzed. Feature selection was performed using least absolute shrinkage and selection operator, and eight machine learning algorithms were evaluated. Model performance was assessed using discrimination, calibration, and decision curve analysis. The optimal model was further interpreted using Shapley Additive Explanations analysis and implemented as an online calculator. RESULTS: Ten variables were selected for model development. Among the evaluated algorithms, the K-nearest neighbors model showed the best overall performance in the test set (area under the receiver operating characteristic curve 0.954) with good calibration and clinical utility. Key contributors included albumin, cholestasis index, the variceal prediction rule, and total bile acids. CONCLUSIONS: This model provides an accessible tool for outpatient assessment after KPE and may assist clinicians in identifying high-risk patients and improving risk stratification during follow-up.