Prediction of venous thromboembolism after spontaneous intracerebral hemorrhage based on machine learning

Abstract

This multicenter retrospective study aimed to develop and validate machine learning models for predicting venous thromboembolism (VTE) following spontaneous intracerebral hemorrhage (SICH). The analysis included 988 SICH patients (748 from West China Hospital for model development and 240 from Leshan People's Hospital for external validation), incorporating comprehensive clinical, radiological, and laboratory parameters. Five machine learning algorithms, including XGBoost, were evaluated using a 3:1 training-test split and external validation approach.

Results

demonstrated significantly higher VTE incidence in patients with greater anticoagulant exposure (p < 0.05), intraventricular hemorrhage (68.75 % vs 51.32 %), and infratentorial involvement (17.19 % vs 7.6 %). VTE patients exhibited larger hematoma volumes (33.5 ± 7.2 vs 25.0 ± 6.8 mL), tachycardia (88.0 ± 14.2 vs 82.0 ± 12.1 bpm), lower Glasgow Coma Scale (GCS) scores (8.0 ± 3.1 vs 13.0 ± 2.8), and elevated inflammatory markers. External validation confirmed these findings, with older age, larger hematomas, and higher D-dimer levels in VTE cases. XGBoost achieved superior predictive performance (AUC: 0.87 training, 0.81 test, 0.80 validation), with SHapley Additive exPlanations (SHAP) analysis identifying D-dimer, hematoma volume, and neutrophil count as key predictors. Conclusion:‌ XGBoost outperforms conventional methods in predicting post-SICH VTE through multidimensional data integration, providing a robust tool for personalized risk stratification and clinical prevention strategies.