Prediction model for complex venous malformation complete remission after sirolimus using clinicopathological and CEUS parameters

Abstract To develop and validate a machine learning model incorporating clinicopathological features and contrast-enhanced ultrasound (CEUS) parameters for predicting complete remission (CR) following sirolimus therapy in patients with complex venous malformations (CVM), thereby supporting individualized treatment decisions. A retrospective cohort of 260 CVM patients treated with sirolimus (January 2014-December 2024) was enrolled. Patients were randomly allocated to training (n = 182) and validation (n = 78) sets (7:3 ratio). Baseline clinicopathological (e.g., maximum lesion diameter, volume), CEUS (time-to-peak [TTP], peak intensity [PI]), and laboratory (D-dimer, hemoglobin) data were collected. Univariate analysis and multivariate logistic regression were performed to identify independent predictors. Subsequently, random forest, logistic regression, and K-nearest neighbors models were constructed. Model performance was evaluated using receiver operating characteristic curves, calibration plots, and decision curve analysis. SHAP values and a nomogram enhanced interpretability. The training set CR rate was 30.77% (56/182). Maximum lesion diameter, volume, TTP, PI, hemoglobin, and D-dimer were significantly associated with CR on univariate analysis (P < 0.05). Multivariate logistic regression identified maximum lesion diameter (OR = 0.800), lesion volume (OR = 0.987), TTP (OR = 1.189), PI (OR = 0.815), and D-dimer (per 0.1 mg/L FEU increase: OR = 0.415) as independent predictors (P < 0.05). The RF model yielded an AUC of 0.869 (95%CI: 0.809–0.929) in the training set and 0.766 (95%CI: 0.629–0.902) in the validation set, with satisfactory calibration with satisfactory calibration (calibration intercept and slope close to 0 and 1; Hosmer-Lemeshow P > 0.05), supported by bootstrap optimism correction and 10-fold cross-validation. Decision curve analysis confirmed net clinical benefit. SHAP value ranking was D-dimer > TTP > PI > maximum lesion diameter > lesion volume. The nomogram enabled individualized CR risk visualization. A random forest model and nomogram based on clinicopathological and CEUS parameters accurately predict sirolimus‑induced CR in CVM patients. Maximum lesion diameter, volume, TTP, PI, and D-dimer are independent predictors. This model offers an effective tool for candidate selection and personalized treatment planning.