Predicting 3-month prognosis of cerebral venous thrombosis: a machine learning approach incorporating inflammatory...

This study aimed to explore a machine learning (ML) model for predicting the long-term prognosis of Cerebral Venous Thrombosis (CVT). We conducted a comprehensive retrospective analysis on CVT patients admitted to Fujian Medical University Fuzhou General Hospital and The Affiliated Zhongshan Hospital of Xiamen University from January 1, 2020 to December 31, 2023. The retrospective cohort of 350 patients was used for model development using a 5-fold nested cross-validation framework. To address class imbalance, the Borderline-SMOTE technique was applied within the internal training folds. A total of 31 clinical and laboratory variables were evaluated as potential predictors. Lasso and Elastic Net (cv-Enet) were employed for preliminary variable selection, and the intersection of these two methods yielded 11 key variables. Random Forest was then applied to further reduce dimensionality, resulting in a final set of 5 crucial variables. These variables were ranked by importance using SHapley Additive exPlanation (SHAP) analysis. In internal validation using nested cross-validation, the SVM model achieved an area under the receiver operating characteristic curve (AUROC) of 0.870, with a sensitivity of 0.784 and a specificity of 0.897. This performance was corroborated in an independent, prospectively collected external cohort of 152 patients from the First Hospital of Shanxi Medical University (January 1, 2023 to December 31, 2025) where the model demonstrated robust discrimination with an AUROC of 0.862 (95% CI: 0.829-0.884). However, calibration analysis indicated a slight overestimation of risk (calibration-in-the-large ≈ 9%), suggesting that the model currently serves best as a risk stratification adjunct rather than a definitive outcome predictor. Bootstrap resampling confirmed the stability of the discriminative performance. Exploratory subgroup and fairness analyses suggested stable model performance across key demographic subgroups. Our study focused on the development and validation of an ML-based model for long-term CVT prognosis prediction, highlighting the significant role of inflammatory markers. These findings indicate that the SVM model, with further calibration and prospective validation, could serve as a useful adjunct for risk stratification in CVT.