Abstract
BACKGROUND: Lower extremity deep venous thrombosis (LEDVT) is a frequent and serious complication after aneurysmal subarachnoid hemorrhage (aSAH). Existing risk scores poorly discriminate LEDVT risk in this population. OBJECTIVE: To develop and externally validate machine learning (ML) models for early prediction of LEDVT in aSAH patients treated with endovascular therapy. METHODS: We performed a retrospective multicenter study including an internal cohort (n = 593) for model development and internal validation and an external cohort (n = 142) for external validation. Thirty-seven clinical and laboratory variables were considered. Variable selection used LASSO followed by multivariable logistic regression. Seven ML algorithms (XGBoost, LightGBM, random forest, logistic regression, SVM, KNN, MLP) were trained with 5 × 5-fold cross-validation; AUC was the primary metric. Model interpretability used SHAP. An online risk calculator was implemented. RESULTS: Six predictors were selected (age, albumin, D-dimer, GCS, AISI, and MCA aneurysm). XGBoost achieved the best discrimination (internal AUC 0.88; external AUC 0.80). Decision curve analysis showed clinical net benefit across relevant thresholds. SHAP analysis highlighted D-dimer, albumin, and GCS as key contributors. A web-based calculator was deployed to facilitate clinical use. CONCLUSIONS: An XGBoost-based model incorporating six routinely available variables accurately predicts LEDVT risk after aSAH and generalized to an external cohort. The web tool may help target preventive strategies for high-risk patients.