Comparative Analysis of Cancer-Induced Venous Thromboembolism Risk Assessment Models: A Systematic Review

Introduction Cancer-associated thromboembolism (CAT) is a major cause of morbidity and mortality in cancer patients. To categorize patients by VTE risk, Risk Assessment Models such as Khorana, Vienna CATS, and COMPASS-CAT have been developed, but their predictive accuracy varies across cancer types and populations [ 1 ]. Given the increasing incidence of CAT across different cancer populations, the evaluation of clinical and biomarker-driven risk assessment models (RAMs) is critical for guiding thromboprophylaxis [ 2 ]. This systematic review aims to identify, compare, and evaluate the efficacy, predictability, validity, and clinical value of VTE risk assessment models in cancer patients. The aim is to determine the most reliable clinical or biomarker-derived tool to guide thromboprophylaxis, while also identifying predictors that consistently demonstrate strong associations with VTE risk across cancer types. Method Searches were conducted in PubMed, EMBASE, and Scopus, with supplementary searches via Google Scholar. The review followed PRISMA 2020 guidelines and was registered with PROSPERO (CRD420250645949). Primary studies in English that developed and validated new RAMs for cancer patients at risk of VTE up to May 2025 were included. Three independent authors screened, selected, extracted, and qualitatively synthesized the data. In total, 28 studies published between 2017 and 2025 were included, covering more than 15,000 cancer patients across China, the U.S., Austria, and Europe. Both inpatient and outpatient settings were represented. Most studies were retrospective cohorts, though recent years have demonstrated an increase in prospective validation. Results A total of 2,022 articles were initially identified, and 28 studies met the inclusion criteria. Logistic regression-based nomograms demonstrated strong performance, especially in Chinese cohorts (AUC 0.854–0.929). ML models, like XGBoost and random forest, achieved the highest AUC values (0.745–0.990), consistently outperforming Khorana (AUC 0.756–0.875) and Caprini (0.589–0.759). D-dimer, present in over 80% of models, was the most reliable and consistent predictor (OR=24.3 in colorectal cancer; OR=5.58 in lung cancer; HR=1.14 in ovarian cancer). Other predictors included central venous catheter (OR≤7.7), advanced cancer stage (OR=2.7), BMI (OR approximately 2.0), and chemotherapy (OR=1.66). Additional tumor-specific predictors, such as Ki67 in ovarian cancer and CEA in lung cancer, were also identified. Validation practices differed across regions, with 85.7% of European studies performing external validation compared to 42.9% in Asia. Conclusion Models combining clinical predictors with biomarkers and validated externally showed the greatest clinical potential. D-dimer was the strongest and most consistent predictor across cancer types. Machine learning models provided superior discrimination but require further external validation to ensure real-world applicability. Publication History Article published online: 17 February 2026 © 2026. Thieme. All rights reserved. Georg Thieme Verlag KG Oswald-Hesse-Straße 50, 70469 Stuttgart, Germany