Dynamic assessment of venous thromboembolism risk in patients with cancer by longitudinal D‐Dimer analysis: A prospective study

Abstract Background Venous thromboembolism (VTE) is a frequent complication of cancer. Elevated D‐dimer is associated with an increased risk of cancer‐associated VTE. Whether changes in D‐dimer over time harbor additional prognostic information that may be exploited clinically for dynamic prediction of VTE is unclear. Objectives To explore the potential role of longitudinal D‐dimer trajectories for personalized prediction of cancer‐associated VTE. Patients/Methods A total of 167 patients with active malignancy were prospectively enrolled (gastrointestinal: n = 59 [35%], lung: n = 56 [34%], brain: n = 50 [30%], others: n = 2 [1%]; metastatic disease: n = 74 [44%]). D‐dimer (median = 0.8 µg/mL [25th‐75th percentile: 0.4‐2.0]) was measured at baseline and during 602 monthly follow‐up visits. Joint models of longitudinal and time‐to‐event data were implemented to quantify the association between D‐dimer trajectories and prospective risk of VTE. Results VTE occurred in 20 patients (250‐day VTE risk = 12.1%, 95% confidence interval [CI], 7.8‐18.5). D‐dimer increased by 34%/month (0.47 µg/mL/month, 95% CI, 0.22‐0.72, P < .0001) in patients who developed VTE, but remained constant in patients who did not develop VTE (change/month = −0.06 µg/mL, 95% CI, −0.15 to 0.02, P = .121). In joint modeling, a doubling of the D‐dimer trajectory was associated with a 2.8‐fold increase in the risk of VTE (hazard ratio = 2.78, 95% CI, 1.69‐4.58, P < .0001). This finding was independent of established VTE risk factors. Highly personalized, dynamic predictions of VTE conditional on individual patients’ D‐dimer trajectories could be obtained. Conclusions D‐dimer increases before the onset of cancer‐associated VTE, but remains constant over time in patients without VTE. This study represents proof‐of‐concept that longitudinal trajectories of D‐Dimer may advance the personalized assessment of VTE risk in the oncologic setting.


| INTRODUC TI ON
Cancer and coagulation are highly linked processes. 1 Although an activated coagulation cascade contributes to tumor progression and metastasis, cancers induce a hypercoagulable state that promotes venous thromboembolism (VTE). 2 VTE is a frequent complication and a leading cause of morbidity and death in patients with cancer. 3 The overall risk of developing VTE in patients with active malignancy is approximately 5% to 10% over 2 years, 4 but strongly varies between patient subgroups according to prognostic factors such as tumor type. 5 Indeed, 2-year VTE risks can be as low as 2% in patients with prostate cancer and as high as 30% in patients with upper gastrointestinal malignancies. 6 Randomized controlled trials have shown that prophylactic anticoagulation can significantly reduce the risk of VTE in the oncologic setting, but only patients at very high risk derive a clinically meaningful magnitude of benefit from this intervention. 7 Thus, the identification of cancer patients at high risk of VTE is an important area of clinical research and necessary prerequisite for improving the therapeutic ratio of prophylactic anticoagulation in the oncologic setting. Several clinical prediction models currently exist for VTE risk assessment in cancer patients, including the Khorana score, 6 but all of these models appear to have modest prognostic performance [8][9][10] .
During the past several years, we and others have shown that elevated biomarkers of hypercoagulability, such as D-dimer, are strongly associated with a higher risk of VTE, 11,12 and can furthermore improve established clinical prediction models. This supports the concept that hemostatic biomarkers could support physicians in selecting patients with the highest VTE risk for prophylactic anticoagulation while sparing low-VTE-risk patients from unnecessary burden and bleeding complications. 13 However, coagulation and cancer are dynamic processes that may be influenced by patient-and treatment-related factors that change over time, such as disease progression and anti-neoplastic therapies.
Indeed, cancer nowadays represents an increasingly chronic disease, and hemostatic biomarker levels may thus significantly change during the patient journey. A single hemostatic biomarker measurement for VTE risk prediction may thus represent only one "snapshot" of cancer-associated hypercoagulability at a single point in time. All currently available biomarkers and clinical prediction models for cancer-associated VTE were developed as a single measurement at a baseline time point. 8 Whether repeated quantification of hemostatic biomarkers over time may represent a clinically superior approach for VTE risk stratification in cancer patients is unclear.
Thus, we hypothesize that longitudinal trajectories of the hemostatic biomarker D-dimer may harbor important "dynamic" prognostic information on cancer-associated VTE risk beyond a single D-dimer measurement in time, and may improve the clinical assessment of venous thromboembolic risk in patients with cancer. So-called joint models of longitudinal and time-to-event data have been developed for examining this hypothesis. [14][15][16] In this prospective cohort study, we used joint models to define the role of the longitudinal D-dimer trajectory for dynamic prediction of VTE in patients with cancer, with the aim of answering whether such a concept could be a clinically meaningful strategy for improving VTE risk assessment in oncology.

| Study design
In this longitudinal substudy of the prospective and ongoing Vienna Cancer and Thrombosis Study (CATS), we enrolled patients with solid cancers who were treated at Vienna General Hospital from January 2011 to July 2014. Detailed in-and exclusion criteria were reported previously. 17 • We modelled the association between D-Dimer trajectories and risk of cancer-associated VTE.
• D-Dimer increased before the onset of cancer-associated VTE.
• D-Dimer remained constant over time in patients without VTE.
• Highly personalized dynamic predictions of VTE based on D-Dimer trajectories could be obtained.
250-day observation period. All events were adjudicated by an independent panel (n = 3 experts in vascular medicine, radiology, and nuclear medicine). So-called incidental PE was counted as an event if the panel deemed it to be of clinical significance requiring anticoagulation. Fatal PE was defined as (1) PE as the cause of death on autopsy record or (2) assessment of PE as the immediate cause of death by the adjudication committee.

| Laboratory analysis
Citrated venous blood samples (3.2% trisodium citrate tube, VACUETTE®, Greiner-Bio One) were obtained at each visit by antecubital venipuncture or from central venous catheters. D-dimer was assessed within our hospital's routine laboratory with immunoturbidimetry using the STalia D-DI assay (Diagnostica Stago). 19

| Statistical methods
All statistical analyses were performed with Stata 15.1 (Stata Corp.). Cumulative VTE incidence was estimated with competing risk estimators, treating death from any cause other than fatal VTE as the competing event. [20][21][22][23] The association between baseline D-dimer and VTE was modeled with uni-and multivariable Weibull proportional hazards regression. 24 The primary analysis quantity of this study was the association parameter α (i.e., the association between the longitudinal D-dimer biomarker trajectory and the hazard of VTE expressed as a hazard ratio). 16  Moreover, we adjusted α for whether patients had metastatic cancer at baseline, and investigated a "first derivative" specification of α (i.e., the "slope" or rate of change in D-dimer). 26 All joint models were fitted with the user-contributed Stata routine stjm, 27 freely available on the Boston College Statistical Software Components archive. 28 Two specifications of D-dimer were studied in the joint models, namely D-dimer on a continuous original scale (µg/ dL) and on a log2-transformed scale (i.e., per doubling). The fit of these two specifications was compared using the AIC. Predictions of VTE risk for individual patients conditional on their D-dimer trajectories were obtained using a Stata routine (stjmcsurv, currently in development but freely available online) 29 based on the dynamic prediction approach of Rizopoulos. 30

A Strengthening
The Reporting of Observational studies in Epidemiology checklist (see supporting information). The full analysis code is available on requested from F.P.

| Evolution of D-dimer levels over time in patients who did and did not develop VTE
After baseline, patients returned for 602 follow-up visits, for a total number of 769 visits included in the analysis (median number of visits per patient = 5, 25th-75th percentile: 3-7, range: 1-7). Measurements of D-dimer were available for 761 visits (1% missing, Table S2). In univariable joint modeling of D-dimer and time-to-thrombosis, mean D-dimer at baseline was 1.84 µg/mL and remained constant during follow-up in the overall study population (change = −0.03µg/mL/ month, P = .573, Table 2). Notably, the change in D-dimer over time was different in patients who did and did not develop VTE during follow-up ( Figure 1). In detail, D-dimer remained stable over time in patients who did not develop VTE, but increased by 0.47 µg/mL/ month in patients who developed VTE (P < .0001, Table 2). This result could be confirmed on a relative scale, where D-dimer decreased by 2.6%/month in patients who did not develop VTE and increased by 34%/month in patients who developed VTE, respectively (P < .0001, Table 2).   b Patients with primary brain tumors were assigned to the "local" stage group.

| Longitudinal D-dimer trajectories for prediction of VTE risk
c Tumor site categories were defined as in the original publication by Khorana et al (i.e., colorectal cancer included in the "low/moderate VTE risk" group), 6 with brain tumors being assigned to the "very high VTE risk" group according to Ay et al. 37 1 and 2 in Table 3). This prognostic association applied similarly to patients with and without metastatic cancers (Table S3)

| Personalized prediction of VTE according to D-dimer trajectories
The joint model (model 1 in Table 3) could be used to obtain highly person- Cancer is becoming an increasingly chronic disease. 31,32 Consequently, patient-related, tumor-related, and treatment-related risk factors can change over the patient journey and dynamically modify an individual patient's prognosis with regard to VTE and survival. 5 The dynamic reassessment of prognosis according to changes in clinical and laboratory parameters is a highly intuitive concept for clinicians, who have used it in everyday clinical practice for decision-making since ancient times. 33 In contrast, appropriate statistical methods for this purpose have until recently been limited. 34 The advent of so-called joint models of longitudinal and time-to-event data greatly facilitates a systematic analysis of the prognostic relationship between a longitudinal risk factor trajectory and a clinical outcome. 14 trajectories, predictions of 6-month VTE incidence ranging from below 10% to above 20% were obtained. In the future, such pre-  Note: Hazard ratios per doubling were obtained by using log2-transformed D-dimer. Log2transformed D-dimer gave a better fit to the data than D-dimer on its original µg/mL scale (AIC: 2434 vs 3308).

TA B L E 3 Associations of longitudinal D-dimer trajectories and prospective thrombotic risk-univariable joint models
Seven limitations of our study should be discussed. First, although we analyzed one of the strongest known risk factors for cancer-associated VTE (D-Dimer), it is reasonable to assume that other static and dynamic factors such as type of chemotherapy, intercurrent infections or invasive diagnostic procedures, and cancer remission status will further modify VTE risk in a dynamic fashion. 35,36 Moreover, it can be reasonably speculated that these and potentially also other unmeasured risk factors may not only modify VTE risk but also the D-dimer trajectory. Unfortunately, these time-dependent data could not be included in the current analysis because of a lack of pertinent data. Nonetheless, joint models can easily accommodate multiple static and dynamic co- are not yet implemented in routine statistical analysis software. 26 Rather, we used the Weibull model, which is a proportional hazards model with parametric representation of the baseline hazard. Not considering competing risks by using the Weibull model implies that the patient-specific predictions of VTE risk may have been slightly overestimated (see Figure S2) [21][22][23] ; more work will need to be done in the future to incorporate competing risks in this setting. Third, because of the rarity of prospective longitudinal biomarker studies in the oncologic setting. 5    longitudinal D-dimer cutoff that would warrant the initiation of primary thromboprophylaxis. In our study, D-dimer was measured each month, but whether this is feasible within a routine oncology setting needs to be further explored. We therefore encourage others to implement and validate the joint modeling process in patients with cancer for assessing VTE risk in external cohorts.
In summary, we believe that the current study presents a first proof-of-concept rather than an immediately applicable system on how longitudinal biomarker trajectories can be used clinically for dynamically (re-)assessing VTE risk in oncology.

| CON CLUS ION
In summary, we conclude that D-dimer levels increase in patients with cancer before the onset of VTE, but remain stable in patients

CO N FLI C T S O F I NTE R E S T
The authors do not declare any conflict of interest.

P R I O R P R E S E N TAT I O N
Parts of this work were presented as an abstract at the December 2017 Annual Meeting of the American Society (ASH) in Atlanta, GA.
The abstract was awarded an ASH Young Investigator Award to F.P.