Early Prediction of Insufficient Response to Ursodeoxycholic Acid: A Multimodal Model Based on Radiomics and Clinical Indicators.

RATIONALE AND

OBJECTIVES

This study aimed to integrate two-dimensional ultrasound (2D-US) radiomics and clinical indicators to develop a prediction model for early identification of primary biliary cholangitis (PBC) patients with insufficient response to ursodeoxycholic acid (UDCA) at 6 months.

MATERIALS AND

METHODS

136 PBC patients were enrolled. Least absolute shrinkage and selection operator regression was applied to select key features and construct a radiomics model. Six machine learning algorithms were compared using 5-fold nested cross-validation. Univariate and multivariate logistic regression were performed to identify independent clinical predictors and construct a clinical model. By combining the radiomics score and significant clinical factors using multivariate logistic regression analysis, a combined model was subsequently developed. Model performance was evaluated by AUC, calibration curves, and decision curve analysis.

RESULTS

The support vector machine (SVM) algorithm achieved the highest mean AUC (0.696 ± 0.028) and the lowest coefficient of variation (4.055%) in nested cross-validation. Alkaline phosphatase (ALP) and platelet count (PLT) were independent predictors of insufficient response to UDCA. The combined model achieved an AUC of 0.865 in the test set, higher than those of the clinical model (0.778) and the radiomics model (0.833). The response status at 6 months was highly consistent with that at 12 months (kappa = 0.809). The model showed generally consistent predictive performance across different UDCA response criteria and histological fibrosis stages.

CONCLUSIONS

A noninvasive multimodal model based on pretreatment 2D-US radiomics features, ALP, and PLT can accurately predict the insufficient response to UDCA in PBC patients at an early stage.