Development of Models for Predicting the 7-Month Risk of Venous Thromboembolism and Clinically Relevant Bleeding in Ambulatory Patients with Cancer: Analysis from the AVERT Trial

BACKGROUND: Patients with cancer are at elevated risk of venous thromboembolism (VTE). While primary thromboprophylaxis reduces VTE incidence, it also increases bleeding risk, necessitating accurate risk stratification. Existing tools (e.g., Khorana score) have modest predictive value, and bleeding risk models have not been validated in ambulatory cancer patients. OBJECTIVE: To develop and internally validate prediction models for VTE and clinically relevant bleeding in ambulatory cancer patients initiating chemotherapy. METHODS: We used data from 514 participants in the Apixaban for the Prevention of Venous Thromboembolism in High-risk Ambulatory Cancer Patients (AVERT) randomized trial. Outcomes included objectively confirmed VTE and clinically relevant bleeding over 7 months. For each outcome, we trained three models: logistic regression with L1 regularization and two XGBoost models (one using all variables, including biomarkers and genetic data; one using only routinely collected variables). Performance was evaluated using optimism-adjusted area under the curve (AUC), precision-recall AUC, calibration, and diagnostic statistics. RESULTS: VTE and clinically relevant bleeding occurred in 8.4% and 9.0% of patients, respectively. The full-variable XGBoost model performed best (VTE AUC: 0.92, bleeding AUC: 0.90). Key VTE predictors included Factor V Leiden, topoisomerase inhibitor use, apixaban, stomach or pancreatic cancer, hemoglobin and high-sensitivity troponin T whereas top bleeding predictors included extracellular vesicles-procoagulant activity, N-terminal pro-B-type natriuretic peptide, C-reactive protein, age, body mass index and blood counts. CONCLUSION: In this proof-of-concept study, prediction models incorporating biomarker and genetic variables accurately predicted VTE and clinically relevant bleeding in cancer patients, though caution is warranted due to the modest sample size.