Machine learning models predict risk of lower extremity deep vein thrombosis in hospitalized patients with spontaneous intracerebral hemorrhage.

Lower extremity deep vein thrombosis is one of the important complications of spontaneous intracerebral hemorrhage. We aimed to develop a risk assessment model to predict the risk of lower extremity DVT during hospitalization in patients with spontaneous cerebral hemorrhage. The retrospective study began by randomly dividing the data into a training set and a test set in a 7:3 ratio. Feature selection was performed in the training set, and Boruta and LASSO algorithms were used to screen significant predictors. Five machine learning algorithms were used to construct the prediction model and the model accuracy was evaluated by ROC curves. To validate the model, we constructed calibration curves and compared the calibration of the model using the Brier score. Finally, the clinical value of the model was assessed by Decision Clinical Curve (DCA) and the "black box" model was interpreted by SHAP. The training and test sets did not show significant differences between the individual variables. Screening by the LASSO and Boruta algorithms yielded 15 and 7 potentially relevant variables, respectively, resulting in the identification of six significant predictors associated with DVT. Subsequently, the performance of five machine learning algorithms in DVT prediction was evaluated in the test set. These results suggest that the LGBM model has significant advantages in predicting DVT after cerebral hemorrhage. We developed a model to predict the risk of lower extremity deep vein thrombosis during hospitalization in patients with spontaneous cerebral hemorrhage, and this model can accurately identify high-risk patients.