Construction and validation of a liquid neural network-based prediction model and bedside tool for spontaneous passage of common bile duct stones.

backgroundTo construct and externally validate a liquid neural network (LNN)-based risk prediction model for spontaneous passage of common bile duct stones (CBDS) and to develop a cross-platform AI bedside tool integrating real-time SHapley Additive exPlanations (SHAP) interpretability analysis.

methodsClinical data were retrospectively collected from 911 patients diagnosed with CBDS on imaging and scheduled for endoscopic retrograde cholangiopancreatography (ERCP) at two hospitals between January 2022 and February 2026. Dataset 1 (n = 759) was divided into a training set (n = 531) and a validation set (n = 228) at a 7:3 ratio, while Dataset 2 (n = 152) from a second hospital in the same city served as a regional (second-center) external test set. From 23 candidate predictors, optimal feature subsets for each model were determined through univariate screening and recursive feature addition, and synthetic minority oversampling technique (SMOTE) was applied to the training set. Five models were constructed: logistic regression (LR), decision tree (DCT), random forest (RF), extreme gradient boosting (XGBoost), and LNN. Model performance was evaluated across three dimensions: discrimination (AUROC), calibration (calibration curve), and clinical utility (decision curve analysis [DCA]). SHAP was employed for interpretability analysis. A bedside tool integrating real-time SHAP-based force plots was developed using the Streamlit framework and externally validated.

resultsOf the 911 patients, 167 (18.33%) experienced spontaneous stone passage. Stone diameter, common bile duct (CBD) dilation, and solitary CBD stone were core variables retained across all five models. In the validation set, the LNN model achieved an AUROC of 0.885 (95% CI 0.846-0.917), the highest among the five models (DCT 0.868, LR 0.865, RF 0.865, XGBoost 0.830), with overlapping confidence intervals indicating discrimination comparable to that of DCT, LR, and RF; the LNN showed an accuracy of 83.06%, sensitivity of 81.56%, specificity of 84.46%, and F1 score of 82.25%. The LNN model yielded the lowest Brier score (0.114). DCA demonstrated that the LNN model achieved higher net benefit within the threshold probability range of 0.05-0.50 and maintained positive net benefit over a broader threshold interval. In the external test set, the LNN-based bedside tool achieved an AUROC of 0.855 (95% CI 0.769-0.939), with an accuracy of 90.79%, sensitivity of 89.29%, specificity of 91.13%, and negative predictive value (NPV) of 97.41%. SHAP analysis identified stone diameter as the most important predictor, followed by CBD dilation, symptom improvement, solitary CBD stone, distal CBDS, and GGT decrease.

conclusionThe LNN-based prediction model for spontaneous passage of CBDS achieved discrimination comparable to, and calibration and clinical net benefit superior to, conventional machine learning models in this regional two-center dataset. The AI bedside tool, integrating real-time SHAP interpretability analysis, can assist clinicians in accurately identifying patients with a high probability of spontaneous stone passage prior to ERCP, demonstrating favorable potential for clinical translation.