= 78, 2020-2024) were assembled from three tertiary centers. Thirty-one admission variables underwent recursive feature elimination and LASSO regression, yielding nine key predictors. Five machine learning algorithms-support vector machine, logistic regression, random forest, extreme gradient boosting, and light gradient boosting machine (LightGBM)-were trained and optimized using nested five-fold cross-validation, with Borderline-Synthetic Minority Oversampling Technique applied to address class imbalance. Model performance was further assessed in the prospective cohort, and bootstrap resampling (2,000 iterations) was used to estimate confidence intervals. Among the evaluated models, LightGBM showed the best predictive performance, with an internal mean area under the receiver operating characteristic (ROC) curve of 0.91 ± 0.03 and an external validation area under the curve (AUC) of 0.86 (95% CI, 0.76-0.94), together with a sensitivity of 0.74, specificity of 0.86, and F1 score of 0.81. SHapley Additive exPlanations analysis was used to improve interpretability and quantify the contribution of individual predictors. In addition, an online risk calculator was developed to facilitate individualized risk estimation. These findings suggest that an explainable machine learning framework may provide useful support for early prognostic stratification in tCVST and may assist clinical decision-making in this complex neurotrauma population.
Li et al. (Mon,) studied this question.