Twitter Facebook LinkedIn
Center for Global Research Data

Assessing bleeding risk in patients with cancer-associated thrombosis

Lead Investigator: Mathilde Nijkeuter, University Medical Center Utrecht
Title of Proposal Research: Assessing bleeding risk in patients with cancer-associated thrombosis
Vivli Data Request: 5095
Funding Source: Yes; University Medical Center Utrecht, Utrecht, the Netherlands
Potential Conflicts of Interest: None

Summary of the Proposed Research:

Venous thromboembolism (VTE) comprises deep vein thrombosis (DVT) and pulmonary embolism (PE). In DVT, a blood clot blocks a vein in the leg, whereas in PE, a vessel in the lung is blocked. VTE is a major complication among cancer patients, occurring in 4-20% of cancer patients. Anticoagulation is the cornerstone of treatment for cancer-associated thrombosis (CAT). CAT is associated with higher risks of both recurrent VTE and bleeding complications than usual, and risks are highly variable between patients. Therefore, finding the optimum treatment for individual patients with CAT is challenging. Guidelines advise to treat patients with active cancer and VTE for at least 3-6 months to prevent thrombus enlargement, embolization and early recurrence. Continuing anticoagulant treatment minimizes the risk of late recurrent VTE, but comes at the cost of a risk of major bleeding. These major bleedings comprise, among others, intracranial bleeds and bleeding of an internal organ. Therefore, guidelines recommend to consider extended treatment only in those with persisting risk factors for recurrent VTE and a low risk of bleeding.

Daily injections of low molecular weight heparin (LMWH) have been standard of care in the past 15 years. Nowadays, direct oral anticoagulants (DOACs) edoxaban and rivaroxaban, which can be given as tablets, are increasingly prescribed. Their use is advised in patients with a low risk of bleeding and no drug-drug interactions with other treatment[8]. Thus, estimated risk of bleeding plays an important role in deciding on treatment for CAT. Several risk scores for bleeding during anticoagulation have been developed or validated in patients with VTE. Cancer is a predictor in many of these models. This suggests that the models also apply to the CAT population. However, as baseline risk of bleeding is higher and both cancer and treatment may have a profound impact on bleeding risk, it is unclear whether these models may indeed be generalizable to patients with CAT. Up till present, none of these models have been validated in patients with CAT, nor do specific cancer-related bleeding risk scores exist. How to assess risk of bleeding in patients with CAT across different stages of the cancer journey is currently unclear. This complicates not only clinical decision-making, but also patient education.

Insight into risk factors for bleeding during anticoagulant treatment for VTE in cancer patients is of vital importance. To provide directions for further research and guidance for clinical practice, a head-to-head comparison of all published prognostic models for bleeding in one dataset is warranted. In addition, insight in risk factors for bleeding in different stages of the cancer journey is needed to allow for more individualized decision-making. A new bleeding risk score may be required to adequately assess bleeding risk in patients with CAT.

Statistical Analysis Plan

Sample size calculation

Part A: For studies validating prognostic models, a minimum sample size of 100 patients with the event and 100 patients without events has been suggested. The Hokusai VTE Cancer trial comprises 141 first bleeding events on anticoagulation in a study population of 1048 patients. Part B: With the average requirement of at least 10 outcome events per variable for a stable model, this model could involve a maximum of 18 predictors or interaction terms.

Data preparation and descriptive analyses

  • Baseline characteristics will be presented as proportions and percentage (categorical variables), mean value and standard deviation (continuous variables with normal distribution) or median and interquartile range (non-normal distribution).
  • Patterns of missing data will be examined. Missing data will be handled by single imputation.
  • Bleeding risk across different stages of the cancer journey will be calculated:
    • (moderately) extensive cancer (TNM III or IV) or metastatic cancer
    • Recently diagnosed (<3 months) cancer
    • Cancer receiving chemotherapy
  • Off-treatment bleeding rates will be calculated as well, to provide more insight into bleeding in patients with CAT.
  • Cumulative event rate for bleeding will be illustrated with a Kaplan-Meier curve for patients at different stages of the cancer journey and for the total group of patients.

Differences between Hokusai VTE Cancer and other studies are only relevant to Part A: external validation of existing bleeding risk scores. From an extensive literature review and the eCRF of the Hokusai VTE Cancer trial, we can conclude that all predictors included in the existing bleeding risk scores that we aim to externally validate are available with similar or largely comparable definitions or units, with the exception of one: frequent falls. Frequent falls is only included in one risk score, which also contains 16 other items. Regarding outcome definition, as all studies on other bleeding risk scores used ISTH definitions for bleeding or comparable definitions, no differences are expected.

Part A: validation of existing bleeding risk scores

  • A systematic search of literature has been conducted to identify published prognostic models for bleeding during anticoagulation for VTE, as part of a broader systematic review. Search terms included were ‘venous thromboembolism’, ‘deep vein thrombosis’, ‘pulmonary embolism’, ‘prediction’, ‘risk’, ‘prognosis’, ‘bleeding’ and relevant synonyms. Prognostic models for bleeding were selected if they were developed or validated within a population of patients (≥ 18 years) with VTE. This led to the identification of 7 prognostic models for bleeding.
  • All prognostic models of which the full prediction rule including intercept are available will be applied to the Hokusai VTE Cancer dataset in their original form.
  • Subsequently, recalibration will be performed for each of the models. This will result in updated coefficients and intercept.
  • Predictive abilities in terms of discrimination (Harrell’s C statistic) and calibration (calibration plot, Hosmer-Lemeshow goodness-of-fit test) of both original and recalibrated versions of all prognostic models for bleeding will be assessed and compared.
  • Model performance will also be assessed separately by severity of bleeding (CRNMB, major bleeding, fatal bleeding).
  • HRs and associated 95% CIs will be calculated for the difference between those at high risk of bleeding versus those at low risk of bleeding according to all published scores

Part B: developing a new bleeding risk score

  • Continuous variables will be centered to improve interpretability of coefficients. To minimize the influence of outliers, continuous variables will be truncated at the 1st and 99th percentile.
  • Restricted cubic splines will be used to assess log-linearity of the relationship between continuous variables and outcomes. Transformations will be applied when strong non-linear effects will be found.
  • A Fine & Gray subdistribution hazards model with left truncation and right censoring will be used to derive a new prediction model for bleeding in anticoagulated patients with CAT within six months of anticoagulation. Candidate predictors will be chosen beforehand based on literature and expert opinion. Combining all potentially relevant variables in one model may lead to a large and complicated model which would be of limited use. To improve interpretability and reduce overfitting, Lasso (least absolute shrinkage & selection operator) will be used. Lasso estimates coefficients by finding optimal values for the shrinkage factor based on Akaike’s Information Criterion (AIC). By shrinking coefficients to zero, it simultaneously performs model selection in a way that is justifiable in a small dataset.

Hazard ratios for the first six months will be extracted and combined with model coefficients from a model developed in the entire on-treatment period. This will be done to maximize the number of outcomes in the study population while assigning similar weights to patients whom received treatment for a shorter or longer duration.

Competing risks are cessation of anticoagulation and non-bleeding related mortality.

  • Proportional hazards assumptions will be assessed graphically based on Schoenfeld residuals for each covariate. If violated, an interaction term of the covariate with time will be added.
  • The newly derived model will be used to assess differences in risk factors across cancer stages. This will be done by including cancer stage as predictor in the model. Interaction terms will be added to assess the association between cancer stage and candidate predictors.
  • Internal validation will be assessed in terms of discrimination (Harrell’s C statistic) and calibration (Hosmer-Lemeshow test and calibration plots). To assess whether the model performs equally across different stages of the cancer journey, the dataset will be stratified accordingly and discrimination and goodness-of-fit will be assessed in all groups separately.
  • Predictive accuracy of our newly derived model will be compared to previously published bleeding risk scores mentioned under part A.
  • HRs and associated 95% CIs will be calculated for the difference between those at high risk of bleeding versus those at low risk of bleeding according to our score.
  • Cumulative event rate for bleeding will be illustrated with a Kaplan-Meier curve for patients at low risk and at high risk of bleeding according to the newly derived model.
  • External validation of the model will be planned. To ensure generalizability to clinical practice, validation in real world data is preferred.

Software:

All analyses will be performed in R-Statistic Programming. The following add-on packages will be used: glmnet, mvtnorm, Hmisc, editrules, VIM, lme4, mice, foreign, survival, survAUC, pec, rms, mstate, ggplot2, logistf, ipw, mnormt, dplyr, riskRegression, cmprsk, crrstep, timereg, crskdiag, crrp, survminer, xlsx, car, nephron, data.table, tableone, plyr.

Requested Studies:

A Phase 3b, Prospective, Randomized, Open-label, Blind Evaluator (PROBE) Study Evaluating the Efficacy and Safety of (LMW) Heparin/Edoxaban Versus Dalteparin in Venous Thromboembolism Associated With Cancer
Sponsor: Daiichi Sankyo, Inc.
Study ID: NCT02073682
Sponsor ID: DU176b-D-U311

Public Disclosure:

de Winter MA, Dorresteijn JAN, Ageno W, Ay C, Beyer Westendorf J, Coppens M, Klok FA, Moustafa F, Riva N, C Ruiz Artacho P, Vanassche T, Nijkeuter M. Estimating Bleeding Risk in Patients with Cancer-associated Thrombosis: External Validation of Existing Risk Scores and Development of a New Risk Score [abstract]. Res Pract Thromb Haemost. 2021; 5 (Suppl 1). https://abstracts.isth.org/abstract/estimating-bleeding-risk-in-patients-with-cancer-associated-thrombosis-external-validation-of-existing-risk-scores-and-development-of-a-new-risk-score/. Accessed July 23, 2021.

de Winter MA, Dorresteijn JAN, Ageno W, Ay C, Beyer-Westendorf J, Coppens M, Klok FA, Moustafa F, Riva N, Ruiz Artacho PC, Vanassche T, Nijkeuter M. Estimating Bleeding Risk in Patients with Cancer-Associated Thrombosis: Evaluation of Existing Risk Scores and Development of a New Risk Score. Thromb Haemost. 2021 Sep 20. doi: 10.1055/s-0041-1735251.