Development and validation of an interpretable machine learning model for predicting portal vein system thrombosis following splenectomy in Wilson's disease.

Post-splenectomy portal vein system thrombosis (PVST) is a serious complication that adversely affects patient prognosis. The aim of this study was to develop and validate a machine learning-based model for early risk prediction of PVST after splenectomy. A total of 594 patients who underwent splenectomy were included in this study. The data were randomly divided into a training set (n = 417) and a validation set (n = 177) in a 7:3 ratio. Clinical characteristics associated with PVST risk were screened using LASSO regression and univariate analysis. Seven machine learning algorithms were applied for model training, and the models’ performance on the validation set was evaluated using the area under the curve (AUC), sensitivity, specificity, and F1 score. SHAP interpretability analyses and web-based deployment were used to derive clinical insights. In the training set, the key variables identified included splenomegaly grade, splenic thickness, portal vein flow velocity (PVV), portal vein diameter (PVD), portal vein flow (PVF), splenic vein diameter (SVD), postoperative platelet count (POD PLT), and postoperative D-dimer level (POD D-dimer). The XGBoost model outperformed the logistic regression (LR) and random forest (RF) models, achieving an AUC of 0.971 (95% CI: 0.950–0.987), accuracy of 92.8%, sensitivity of 91.2%, and specificity of 94.6%. In the validation set, the XGBoost model exhibited stable performance, with an AUC of 0.970, accuracy of 91.5%, precision of 89.9%, sensitivity of 88.6%, and specificity of 93.5%. the SHapley Additive exPlanations (SHAP) summary analysis indicated that PVV, PVF, and POD D-dimer were the top three predictors of PVST risk (SHAP Value = 0.076, 0.074 and 0.064). SHAP dependency plots and force plots provided explanations of the model’s predictions at the feature and individual levels, respectively. The XGBoost model demonstrated excellent discrimination and calibration for predicting PVST risk after splenectomy. It can serve as an early warning tool to identify high-risk patients, enabling timely interventions and improved clinical outcomes.