A Transfer Learning-Based Diagnostic Model for Biliary Atresia: A Multicenter Study with 389 Cases

Background Early and accurate diagnosis of biliary atresia (BA) among infants presenting with cholestasis remains a challenging clinical task. This study aimed to develop and validate an interpretable machine learning (ML) model using readily available clinical and sonographic features to facilitate early discrimination of BA. Methods A total of 389 cholestatic infants were included from two medical centers. We trained and evaluated five conventional ML models—logistic regression, decision tree, support vector classifier, multilayer perceptron, and random forest—along with one transfer learning (TL) model. Model development utilized seven clinically relevant features selected through LASSO regression. Results The TL model demonstrated superior performance, achieving AUROCs of 0.966 (95% CI: 0.944–0.984) and 0.967 (95% CI: 0.935–0.991) in the cross-validation on the training set and validation set, respectively. Among conventional models, random forest performed best, with a mean AUROC of 0.961 (95% CI: 0.940–0.982) in cross-validation and 0.893 (95% CI: 0.835–0.952) in validation. SHAP interpretability analysis identified abnormal gallbladder emptying index (GEI), elevated GGT, and acholic stools as the top predictive features across all models. A graphical user interface (GUI) was developed to support clinical deployment. Conclusion The proposed TL-based model outperforms conventional ML algorithms in diagnosing BA, showing promising generalizability. With its high accuracy and interpretable predictions, the model offers a practical decision-support tool for the early identification of BA in infants with cholestasis.