Introduction

Venous thromboembolism (VTE)—comprising deep vein thrombosis (DVT) and pulmonary embolism (PE)—is associated with increased long-term morbidity, functional disability, and mortality [1]. Despite advances in identification, prophylaxis, and treatment, the annual event rate of VTE has increased over time [2–5].

VTE is a disease with multiple contributory risk factors, which tends to recur [1, 6], especially during the 3 years after an acute episode [7–9]. Limited data from observational studies have suggested that the cumulative recurrence rate after an acute event is approximately 8 % at 90 days, 11–13 % at 1 year, and 20 % at 3 years [10, 11]. In current practice, anticoagulation is recommended for at least 3 months for nearly all patients [6]. However, despite the proven benefits of anticoagulation, a systematic review has indicated that the case-fatality rate of recurrent VTE is >10 % within 3 months of acute treatment [12]. Therefore, understanding who is at risk for a recurrence during the acute treatment phase may help clinicians determine the optimal frequency of subsequent clinical surveillance and the appropriate type of outpatient treatment. Furthermore, the decision to continue anticoagulation beyond 3 months continues to be individually tailored [6], yet risk-assessment tools for predicting the long-term risk of VTE recurrence at the individual patient-level remain limited [13]. Inasmuch, identifying short-term and long-term risk factors associated with recurrent events after an acute VTE episode may lead to improved strategies for secondary prevention. A limited number of studies have attempted to quantify the risk of, and factors associated with, VTE recurrence. These studies either focused on a subset of VTE patients or used data from randomized trials, administrative databases, or outdated observational cohorts that may limit their value [10, 11, 14–19].

We used population-based surveillance methods to monitor residents of central Massachusetts diagnosed with an acute first-time episode of PE and/or lower-extremity DVT on a biennial basis between 1999 and 2009. We followed these individuals for 3 years to quantify the magnitude of recurrent events and identify predictors of short- and long-term recurrence after the index episode. Our risk score calculator uses characteristics assessed during the index encounter to predict VTE recurrence during the initial 3-month acute treatment period.

Materials and methods

The Worcester VTE study employed population-based surveillance methods to monitor trends in annual event rates of acute episodes of PE and/or DVT, management strategies, case-fatality rates, and recurrences after the index event among all residents of the Worcester metropolitan statistical area (WMSA) (n = 478,000 per 2000 census data) [5, 20–22]. Computer printouts of all WMSA residents with health-care system encounters in which any International Classification of Diseases, Ninth Revision (ICD-9) diagnostic codes consistent with VTE [5] had been listed during 1999–2009 on a biennial basis from all 12 hospitals serving residents of the WMSA were used to screen index events. Data queries encompassed all inpatient, outpatient, emergency department, radiology department, and diagnostic laboratory encounters. Medical records related to the index and follow-up events were reviewed retrospectively by trained abstractors and validated by clinicians; follow-up was up to 3 years for all independently validated events. National and statewide death registries were reviewed to ascertain survival status of all patients.

The institutional review committee at each participating hospital approved this study.

Definition of index episode of VTE

Patients were classified as first-time VTE or previously diagnosed (recurrent) VTE at the time of their index visit based on whether they had a history of VTE noted in their medical records. Three etiologic categories of VTE were defined [20]: (1) cancer-associated VTE (occurring in the presence of active malignancy); (2) provoked VTE (occurring within 3 months of surgery, pregnancy, trauma, fracture, or hospitalization, but not in the presence of active malignancy); and (3) unprovoked (idiopathic) VTE (occurring in the absence of provoking factors and active malignancy). Whether to classify hormone replacement therapy (HRT) or oral contraceptives (OC) as a provoking factor for VTE is a controversial topic [23–26]. Therefore, we performed a sensitivity analysis by reclassifying HRT/OC as a provoking factor.

In this analysis, only patients with a first-time (incident) episode of VTE were included. Patients diagnosed with upper-extremity DVT alone were excluded due to differences in the natural history of upper- and lower-extremity DVT [27, 28].

Recurrence after index VTE

Through the retrospective review of medical records, a recurrent episode of VTE after the patient’s index event was defined as a first occurrence of thrombosis in a previously uninvolved lower/upper extremity venous (recurrent DVT) or pulmonary (recurrent PE) segment.

Potential prognostic factors

Potential prognostic factors included patient demographics at baseline, medical history within 3 months before the index event, index VTE characteristics [PE with or without (±) DVT vs. DVT alone; community presenting vs. hospital-acquired VTE], antithrombotic medication at presentation, type of acute therapy received in a healthcare facility, and treatment setting (in-hospital, at discharge, ambulatory). For patients admitted to hospital, length of stay and international normalized ratio assessed during hospitalization were assessed.

Statistical analysis

The cumulative incidence rate (CIR) of VTE recurrence within 3 years of the index event was estimated using the Kaplan–Meier method. Data were censored at the time of death or last medical contact (in survivors) up to 3 years following the index event. The log-rank test was used to compare CIRs of VTE recurrence among cancer-associated, provoked, and unprovoked episodes of VTE, as prior publications suggest that the recurrence rates differ among these groups [6, 13]. Hazard ratios (HRs) and 95 % confidence intervals (CIs) generated by the unadjusted Cox proportional hazards regression models were used to describe the relationship of potential prognostic factors to time-to-recurrent VTE.

Since each potential prognostic factor was assessed during the index VTE encounter, and possible unmeasured time-dependent risk factors (e.g. duration of treatment) may have impacted our results, separate prognostic models for predicting short-term (3-month) and long-term (3-year) risks of recurrence were developed. In addition, due to differences between VTE patients with and without active cancer in the risk of recurrence and in patient management practices during the acute treatment period, sensitivity analyses were conducted. The methods used to develop the prognostic models were as follows: the full model included all potential prognostic factors identified by unadjusted Cox proportional hazard regression analyses with p values ≤0.1; multivariable Cox regression with backward selection was then used to select the final independent predictors (p < 0.05). Proportional hazards assumptions were assessed by a test of the interaction between log (time metric) and each predictor (no violations found). Non-linearity of age at the index encounter was assessed by a fractional polynomial technique (no non-linearity found) [29, 30].

The existence of influential outliers was examined by plotting the Scaled Score Residuals versus each predictor to identify subjects who may have influenced the value of a single coefficient and by plotting the likelihood displacement versus the Martingale Residual to identify subjects who may have influenced the vector of coefficients (no subjects were excluded) [31]. We also assessed two-way interactions and co-linearity among the final predictors (none found). Model discrimination was assessed using the Harrell macro for Cox regression (the c-index) [32], while goodness-of-fit (calibration) was assessed by the May–Hosmer method [33]. In addition, 100 replications of bootstrapping using the unrestricted random sampling technique were used to validate the best fitting model internally.

To help clinicians improve management of their patients during the acute treatment phase, a final prognostic model for 3-month VTE recurrence among all patients was used to develop the 3-month risk score calculator based on previous methods [34–38]. A patient risk score was developed as follows: the factor with the smallest logarithmic HR (natural log of HR) was assigned 1 point, with other factors assigned points based on the proportional size of their estimates relative to the smallest logarithmic HR. Individual factor points were summed to provide a total risk score (on a 0–100 scale) for each patient. Calibration was further assessed by comparing predicted to observed risk (using the Kaplan–Meier method) over selected risk score groups, as well as May–Hosmer test.

All analyses were performed using SAS 9.2 (SAS Institute Inc., Cary, NC) and statistical significance level was pre-specified as α = 0.05 (two-sided).

Results

Over the 10-year study, 2989 WMSA residents were diagnosed with a first episode of acute PE ± DVT (42 %) or lower-extremity DVT alone (58 %). These patients were followed for 5836 person-years [mean ± standard deviation (SD) follow-up 23.4 ± 14.4 (median 30) months]. Mean ± SD age was 64.3 ± 18.0 years, 44 % were men, and 94 % were white (Table 1). The proportions of cancer-associated, provoked, and unprovoked episodes of VTE were 17, 43, and 40 %, respectively.

Table 1 Characteristics assessed during index encounter among all patients with first-time VTE
Full size table

During follow-up, 329 patients developed a recurrence [85 (26 %) PE and 244 (74 %) DVT]. Among 244 recurrent DVT cases, 161 (66 %) were proximal DVT, 11 (4.5 %) were isolated calf DVT, 17 (7.0 %) were upper-extremity DVT alone, and in 55 (23 %) the location of DVT was undocumented. The CIRs of VTE recurrence were 5.1 % within 3 months and 15 % within 3 years after the index event among all patients (Fig. 1a). The CIRs of VTE recurrence among patients with active cancer, provoked, and unprovoked VTE were 8.7, 5.2, and 3.8 % within 3 months and 25, 13, and 13 % within 3 years, respectively (Fig. 1b). After we had re-classified 117 patients with HRT/OC (but without other provoking factors or active cancer) from the unprovoked to the provoked group, the proportion of provoked VTE increased to 1408 (47 %) and the proportion of unprovoked VTE decreased to 1084 (36 %). However, the CIRs of recurrent VTE within 3 months or 3 years after an index event were essentially unchanged (Fig. 1c). Furthermore, we stratified the type of index event into PE ± DVT [1251 (42 %)], proximal DVT alone [1525 (51 %)], and isolated calf DVT [213 (7.1 %)]. The CIRs of VTE recurrence within 3 years among the three groups were 13.5, 15.8, and 11.9 %, respectively, but the differences were not statistically significant (log-rank test p = 0.24).

Fig. 1
figure 1figure 1

Kaplan–Meier estimates of cumulative recurrence of VTE among WMSA residents with a first-time VTE diagnosed from 1999 through 2009 in a all patients, b stratified by type of index event, and c stratified by type of index event (HRT/OC as a provoking factor). CI confidence interval, d day, HRT hormone replacement therapy, m month, OC oral contraceptives, VTE venous thromboembolism, WMSA Worcester Metropolitan Statistical Area, y year

Full size image

Predicting recurrence within 3 years: all patients

Factors associated with a risk of recurrence during 3-year follow-up are shown in Table 2. Five independent predictors of recurrence were identified (c-index 0.62) (Fig. 2a): active cancer (with or without chemotherapy), superficial thrombophlebitis, varicose vein stripping, inferior vena cava (IVC) filter placement; while previous surgery predicted a lower risk of recurrence.

Table 2 Characteristics associated with risk of recurrence after index VTE event among all patients with first-time VTE, and within 3 months after index VTE event stratified by active cancer (unadjusted Cox proportional hazards model)
Full size table
Fig. 2
figure 2

Independent predictors of VTE recurrence within a 3 years, b 3 months, and c 3 months in patients without active cancer, among WMSA residents with a first-time VTE diagnosed from 1999 through 2009. *Reference group without active cancer. chemo chemotherapy, CI confidence interval, HR hazard ratio, IVC inferior vena cava, VTE venous thromboembolism, WMSA Worcester Metropolitan Statistical Area

Full size image

Predicting recurrence within 3 months: all patients

Factors associated with recurrence within 3 months are shown in Table 2. Five independent predictors of recurrence were identified (c-index 0.64) (Fig. 2b): active cancer, previous major trauma, varicose vein stripping, anticoagulant therapy at admission, and IVC filter placement. Separating the independent predictor “active cancer” into with or without chemotherapy categories did not improve the model performance; the HRs of active cancer with and without chemotherapy were 1.70 (95 % CI 1.01–2.85) and 1.58 (95 % CI 0.96–2.63) versus no-cancer, respectively. Thus, “active cancer (with/without chemotherapy)” was used as a covariate in the final predictive model (HR 1.63, 95 % CI 1.11–2.41).

Predicting recurrence within 3 months: patients without active cancer

Among 2492 patients without active cancer (106 recurrent VTE episodes at 3 months), risk factors associated with 3-month recurrence were similar to those for all patients, with the exceptions that stroke was associated with an increased risk of recurrence and warfarin treatment was no longer associated with a decreased short-term risk (Table 2). Four independent predictors of recurrence were identified (c-index 0.64) (Fig. 2c): previous major trauma, varicose vein stripping, anticoagulants at admission, and IVC filter placement.

Predicting recurrence within 3 months: patients with active cancer

Among 497 patients with active cancer (35 recurrent episodes of VTE at 3 months), family history of VTE and compression stockings were associated with an increased 3-month risk of VTE recurrence (Table 2). However, only the single factor of family history of VTE remained an independent predictor of recurrence in multivariable regression.

Risk score calculator for predicting 3-month VTE recurrence

A risk score calculator was developed based on the five independent predictors of VTE recurrence during the first 3 months after the index event among all patients (Fig. 3).

Fig. 3
figure 3

a Risk score calculator for b predicted risk of, and c observed versus predicted rate of, VTE recurrence during the first 3 months after the index event among all patients. *May–Hosmer goodness-of-fit test. chemo chemotherapy, IVC inferior vena cava, VTE venous thromboembolism

Full size image

Discussion

We assessed the cumulative risk of VTE recurrence over 3 years among residents of central MA diagnosed with a first-time PE and/or lower-extremity DVT between 1999 and 2009 on a biennial basis. Despite advances in treatment, the 3-year CIR of VTE recurrence remained high in our population-based surveillance study, particularly among patients with active cancer. We systematically evaluated a large number of patient characteristics assessed during the patient’s index encounter as potential risk predictors, and identified several independent predictors of VTE recurrence during the acute (3-month) treatment phase and the long-term (3-year) follow-up window among patients diagnosed with a first confirmed episode of VTE.

Cumulative risk of VTE recurrence

The CIRs of VTE recurrence in our study were higher than have been observed in randomized controlled trials (RCTs) [12]. These differences are likely related, in part, to the inclusion criteria employed in RCTs, resulting in a more narrowly defined “healthier” population. In addition, anticoagulant therapy is monitored more carefully in RCTs than in an uncontrolled community practice setting. Indeed, compared with our findings, published observational studies have reported higher rates of recurrent VTE after an acute episode of VTE: 8 % at 3 months and 20 % at 3 years [10, 11].

After stratifying the patient’s index event into cancer-associated, provoked, and unprovoked VTE, the CIR of VTE recurrence was highest among patients with active cancer. Consistent with our findings, a prospective analysis of >800 patients with VTE, of whom 181 had known cancer at study entry, revealed that the 1-year cumulative incidence of recurrent VTE was threefold higher in patients with cancer versus those without [17]. Historically, recurrence rates after an acute episode of VTE among cancer patients has been higher than among patients without cancer [6, 13, 30]. Our estimated 3-year CIR of recurrent VTE among individuals with a provoked VTE was similar to the 3-year CIR generated from a population-based cohort study including all non-cancer-associated first-time VTE patients identified through the United Kingdom (UK) primary care database between 2001 and 2011 [16]. In this UK study, the 3-year CIR of recurrent VTE was approximately 5 % higher among patients with an unprovoked index VTE versus those with a provoked episode [16]. However, in our study, we found no suggestion of a difference in the recurrence rate between these two groups. Despite re-classifying HRT/OC as a provoking factor, the results remained unchanged. Further investigations on whether HRT/OC is a provoking factor for VTE are needed, as published findings provide conflicting results [25, 39]. While we recognize that differences between the UK study and our study may be related to the use of an administrative database in the UK study and their inability to robustly differentiate index and recurrent events, this may be at least partly explained by the different practices in the US and UK in managing patients with unprovoked VTE. Although we did not collect data on the duration of long-term treatment in the current study cohort, based on our prior publication that focused on a subset of elderly VTE patients, we found that the median duration of warfarin treatment was >1 year for patients with unprovoked VTE and around 6 months for patients with provoked VTE [40]. This suggests that patients in central MA with unprovoked VTE were protected from recurrence for a much longer period than those with provoked VTE. This prolonged treatment pattern may explain why no differences were found in risk of recurrence within 3 years between our provoked and unprovoked groups. In addition, other studies have reported that approximately 4 % of patients with unprovoked VTE develop a recurrence within 6 months [41, 42], similar to our findings. Furthermore, recently published data from a US population-based case-cohort study demonstrated a similar time-to-event pattern as ours: the CIR of recurrent VTE was higher among provoked than unprovoked cases of VTE within the first 2 years after index event but was lower during the later follow-up period [43]. However, unmeasured time-dependent variables may change the strength of association between baseline factors and outcomes, as the literature has shown that interim hospitalization, active cancer, pregnancy, central venous catheter, and respiratory infection are associated with increased hazards of recurrence, whereas warfarin and aspirin are associated with reduced hazards [43].

Owing to increasing use of objective diagnostic tests in the 2000s, there was an increasing trend in the annual event rate for PE [5, 44]. As such, the proportion of patients diagnosed with first-time PE in our study was higher than in historic data [5, 22]. However, after stratifying type of index event into PE ± DVT, proximal DVT alone, and isolated calf DVT, the CIR of VTE recurrence within 3 years after index event was not statistically different among the groups. There are conflicting reports as to whether type of index event (PE vs. DVT) is a predictor of VTE recurrence, irrespective of follow-up duration [30]; these include no increased risk by type of VTE event [10, 41] and an increased risk for patients with incident PE [12, 13] or proximal DVT [11]. In our study, the index diagnosis was not an independent predictor of recurrence during long-term follow-up. These discrepancies could be related to variations in study design, study criteria, incomplete documentation, and duration of follow-up.

Independent predictors of recurrence

After systematically evaluating a large number of characteristics assessed during the index encounter, it was not surprising that active cancer was strongly associated with an increased risk of VTE recurrence over both short- and long-term follow-up. Based on relative model Chi square values (indicating the relative predictive strength of a model’s risk predictors), patients with active cancer undergoing chemotherapy were at greatest risk for recurrence during our long-term follow-up. In a population-based observational study conducted among residents of Olmsted County, Minnesota, over the period 1976–1990, patients with cancer receiving chemotherapy had a more than fourfold increased risk of VTE recurrence within 10 years after the index event compared to patients without cancer, while cancer patients not taking chemotherapy had a twofold increased risk [10]. However, during 3-month follow-up, we observed no additional increase in the risk of VTE recurrence due to receipt of chemotherapy in patients with active cancer. We hypothesize that VTE patients with active cancer who were not undergoing chemotherapy may have received as much clinical monitoring as those undergoing chemotherapy during the acute VTE treatment phase, consistent with the recommendations of contemporary guidelines [6].

With regard to the remaining independent predictors in both the short- and long-term prediction models, these predictors are consistent with the three underlying factors associated with the development of venous thrombi that were first proposed by Virchow in 1884: vascular endothelial damage (i.e. having varicose vein stripping, superficial thrombophlebitis prior to index event), stasis of blood flow, and blood hypercoagulability [45].

Based on relative model Chi square values, IVC filter placement was a strong independent predictor for VTE recurrence over both short- and long-term follow-up. This finding is supported by data in a recent expert consensus review that IVC placement may increase the risk of early VTE recurrence by as much as 50 % [30]. Patients who received an IVC filter may have been unable to receive anticoagulant treatment due to contraindications [42]. Our prior publication demonstrated that patients who received an IVC filter were older and had more comorbidities [46]. These patients require special attention, as an IVC filter alone is inadequate therapy for acute VTE [30].

The role of “transient” risk factors (associated with a reduced risk of recurrence in the long term or increased risk of recurrence in the short term) is widely acknowledged [6, 13]. We found that patients who had undergone surgery within 3 months before the index event had an approximate 30 % reduction in the risk for recurrence during long-term follow-up versus those who had not undergone surgery; this was because the risk of subsequent recurrence declines after the patient recovers from surgery, and is consistent with the literature [6, 13, 30, 47]. In addition, we found that patients with a recent major trauma had a twofold increased risk of VTE recurrence during the first 3 months. Most of these patients were likely still in recovery from their major trauma, which could extend the period of VTE risk, particularly due to prolonged immobility [45]. Taking anticoagulant therapy at admission was an independent predictor for increased risk of VTE recurrence within 3 months, which could be a proxy for other unmeasured risk factors, including comorbidities and genetic predisposition.

Not surprisingly, several predictors identified in the unadjusted Cox proportional hazards model were no longer predictors in the multivariable model (i.e. adjusted for confounders). For instance, use of graduated compression stockings was a predictor for increased hazard of recurrent VTE within 3 months after the index event among patients with active cancer, but was not an independent predictor in the multivariable model.

Historically, increasing age has been considered to be associated with higher incidence rates of VTE in the general population [5, 22, 45, 48]. However, there are conflicting findings as to whether age is an independent predictor of VTE recurrence in the published literature. While some studies indicate that advancing age is associated with an increased risk of VTE recurrence [18, 30], others indicate that increasing age is associated with a decreasing risk of recurrence [19, 49]. Moreover, some studies did not identify age as an independent predictor of VTE recurrence after an acute episode [14, 15, 50], similar to the findings in our study. Similarly, evidence on male sex as an independent predictor of recurrent VTE is conflicting [14, 15, 18, 19]. While VIENNA and DASH indicated that male sex was an independent predictor of increased risk of VTE recurrence and the Louzada’s model indicated a decreasing risk, the Rodger’s model did not identify sex as a predictor for VTE recurrence as what we have found [14, 15, 18, 19]. Again, these discrepancies could be related to variations in study design, selection criteria, duration of follow-up, and incomplete documentation of the physician’s examination for clinical VTE risk factors. Further studies are needed to clarify the reasons for these observed discrepancies [51].

Risk score calculator for predicting VTE recurrence within 3 months of the index event

Realizing that time-dependent characteristics may change over the course of long-term follow-up, we developed a risk score calculator based solely on our 3-month prediction model. Our intention was to develop a risk-assessment tool for use at the index encounter to enable clinicians to tailor individual patient management practices during acute VTE treatment. Our risk calculator included variables that are readily available to clinicians and uses a simple point system to estimate the risk of VTE recurrence.

Although the results of any prediction models are only generalizable to populations similar to the one from which they were derived, we believe that our models are robust, based on our population-based study design, the comprehensive list of patient and clinical characteristics assessed during the index encounter, and internal validation. We hope that our calculator may serve as a tool to assist clinicians at the time of the index encounter to better determine the optimum frequency of subsequent clinical surveillance and the appropriate anticoagulant agents after initial treatment at clinical facilities. Our risk calculator differs from currently available calculators that are focused on information assessed at the end of the acute treatment phase and are only applicable to a subset of VTE cases (e.g. DASH, VIENNA, and Rodger et al’s model for unprovoked VTE, and Louzada et al’s model for cancer-associated VTE) [14, 15, 18, 19]. Owing to differences in the study designs and the inclusion of candidate predictors, not surprisingly, the final sets of predictors differ. Therefore, external validations of these tools are needed.

Study strengths and limitations

This study employed population-based surveillance methods to systematically monitor the clinical epidemiology of VTE among residents of central MA. We conducted broad screening for cases of VTE using multiple databases, validated each potential case of VTE, and performed regular chart audits; nonetheless, we may have missed some cases of asymptomatic VTE. Owing to low autopsy rates in the WMSA and the limited validity of death-certificate data [2, 7], only clinically recognized cases of acute VTE were described and some cases of fatal PE could have been missed. Consistent with other observational studies, unmeasured or inadequately measured variables (e.g. duration of anticoagulant treatment after initial treatment in the clinical facility, changes in dose/type of medication subsequent to initial treatment prescriptions, and changes in medical conditions) may have impacted our findings, despite the inclusion of >50 potential prognostic factors assessed during the index encounter. Further refinement of our prediction model may require the inclusion of additional time-dependent risk factors and perhaps biomarkers to increase precision; and of anticoagulation strategy and adherence to it at 3 months. Moreover, when the sample size is adequate, different prediction models may be developed by stratifying the endpoint (recurrent VTE). Although we conducted internal validation, we recognize that no study can effectively validate itself [52], and external validations are needed to fully assess the model performance. Nevertheless, we believe that our risk score discrimination allows separation of patients into broad, clinically meaningful categories and provides guidance for improving decisions on patient management from the index encounter through the patient’s acute treatment.

Conclusions

This population-based study has quantified the risk of developing a recurrent episode of VTE over a 3-year follow-up among patients with a first episode of VTE. We have identified independent predictors of recurrence that will be useful in the design of future studies focused on estimating the true risks and benefits associated with VTE treatment at the individual patient level. Our risk score calculator can be used during the initial treatment phase for predicting recurrence during the acute treatment phase. This may help clinicians to determine the optimal frequency of subsequent clinical surveillance and the appropriate outpatient treatment of VTE.