Abstract
After 3 months of treatment, clinicians and patients need to decide whether to continue anticoagulant therapy long-term, which is informed by bleeding risk; to date, we have not had a reliable way to estimate that risk.
To develop a dynamic, broad, treatment-adjusted risk score for clinically significant bleeding (CSB).
UK Clinical Practice Research Datalink (2001-2020) was used to generate a retrospective cohort of patients with their first venous thromboembolism who had received 3 months of anticoagulation. Covariates were collected at baseline and as time-varying data. Hazards with 95% CIs were estimated using a Fine–Gray model. A dynamic model with an additive (logarithmic) scoring scheme was developed from sub-distribution hazard ratios. Discrimination was expressed by the C-statistic.
In 51 621 patients with venous thromboembolism, 3301 CSB events were identified in 206 298 person-years of observation. The 10-year incidence rate for CSB was 1.60 per 100 person-years. The model incorporated 18 independent predictors of CSB, assigning risk scores from 0 to ≥14. The risk of CSB decreased over 10 years, but within each year, it increased by point category, ranging from 0.60% to 18.1% in the first year (0-≥14 points). Calibration slopes were 0.81 (0.76-0.87) base model and 0.95 (0.88-1.02) extended model; absolute intercepts were <0.001 in both. Discrimination was stable, with a C-statistic of 0.78 (0.77-0.79).
Our dynamic score effectively identifies patients at risk of CSB, following 3 months of anticoagulant therapy, with excellent discrimination. It enables dynamic prediction of bleeding risk over time, accounting for anticoagulation status.
[Display omitted]