Associated factors and predictive modeling of deep vein thrombosis in stroke patients undergoing rehabilitation

Abstract

This study aimed to identify factors associated with deep vein thrombosis (DVT) and to develop predictive models to facilitate individualized risk assessment in this population. We retrospectively analyzed 240 patients with stroke admitted to a rehabilitation department, including 63 patients with DVT and 177 without DVT. Baseline demographic, clinical, and laboratory variables were collected. Univariate analysis was performed to screen variables potentially associated with DVT, followed by multivariable logistic regression to identify variables independently associated with DVT. In parallel, six machine-learning models were developed using all baseline variables. The full cohort was evaluated across 100 stratified 8:2 train-test splits, with repeated five-fold cross-validation performed within each training subset and corresponding held-out subsets used for testing. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), and SHapley Additive exPlanations (SHAP) were used to enhance interpretability. Multivariable analysis identified lower Barthel Index [odds ratio (OR) = 0.979, p  = 0.008], diabetes (OR = 0.434, p  = 0.032), and malignant tumor (OR = 2.883, p  = 0.014) as variables independently associated with DVT. The final logistic regression model achieved an AUC of 0.765. Among machine learning models, the random forest model demonstrated the highest mean cross-validated AUC (0.781 ± 0.022) and the highest mean held-out AUC across 100 stratified train-test splits (0.797 ± 0.062). SHAP analysis indicated that age, D-dimer level, Barthel Index, activated partial thromboplastin time (APTT), and disease course were the five most influential contributors to Random Forest predictions. In conclusion, our findings support the use of integrated predictive approaches for DVT risk assessment and targeted screening in stroke rehabilitation settings.