Although predicting the risk of venous thrombosis (VT) in an individual from a family with inherited thrombophilia is of major importance, it is often not feasible.
Although predicting the risk of venous thrombosis (VT) in an individual from a family with inherited thrombophilia is of major importance, it is often not feasible.
To develop a simple risk assessment model that improves prediction of the risk of VT for individuals of families with inherited thrombophilia.
1201 relatives from 430 families with inherited thrombophilia (deficiencies of antithrombin, protein C or protein S, and the factor V Leiden and F2 20210A mutations) were recruited at the referral center for thrombophilia in Marseilles, France, from 1986 to 2008. One hundred and twenty-two individuals had a personal history of VT. Sixteen preselected clinical and laboratory variables were used to derive the VT risk score.
The scores based on the 16 variables and on the five most strongly associated variables performed similarly (areas under receiver operating characteristic curves of 0.85 and 0.83, respectively). For the five-variable score, named the MARNI score, derived from family history score of VT, von Willebrand factor antigen levels, age, severity of thrombophilia, and FGG rs2066865, the risk of VT ranged from 0.2% for individuals with a score of 0 (n = 186) to > 70% for individuals with a score of ≥ 7 (n = 27). The model was validated with an internal bootstrap method.
With the use of a simple scoring system, assessment of the risk of VT in subjects from families with inherited thrombophilia can be greatly improved. External validation is now needed to replicate these findings.
Currently, the most commonly tested forms of inherited thrombophilia include deficiencies of antithrombin (AT), protein C (PC), or protein S (PS), and the factor V Leiden (FVL) and F2 20210A mutations . These five defects follow an autosomal dominant inheritance model. However, genetic counseling in families with venous thrombosis (VT) based on screening for these five defects is still debated [2, 3]. Theoretically, identifying the defect in asymptomatic family members from a proband with VT would identify those at high risk, and therefore allow avoidance of high-risk situations or facilitate targeted prophylaxis at times of unavoidable high risk . However, despite harboring the same gene mutation, affected individuals, even within the same family, frequently show marked clinical heterogeneity. Moreover, in such families, individuals without the defect can also have an increased risk of VT in comparison with the general population [5, 6], reflecting selection of families with a strong thrombotic tendency in which as yet unknown thrombophilic defects have cosegregated.
These observations strongly support the hypothesis that the individual risk within these families is affected by multiple genetic and environmental factors. Thus, predicting the risk of VT in an individual belonging to a family with inherited thrombophilia solely by looking at the presence of the family defect is inadequate. Indeed, we have recently shown that ABO blood group and von Willebrand factor (VWF) plasma levels partially explains the incomplete penetrance of congenital thrombophilia within these families . In the present study, we have expanded these findings by studying other clinical and biological risk factors associated with VT in the general population. We hypothesized that stratification of individuals within these families according to their risk of first VT can be achieved on the basis of these risk factors.
In the present study, we aimed to develop a simple risk model based on these different variables that improves prediction of the risk of VT in individuals from a family affected by hereditary thrombophilia.
A detailed description of the MARseilles FAmily Study on venous Thrombosis (MARFAST) cohort has been reported previously . Briefly, the cohort included consecutive families from the Marseille area in France who attended the thrombophilia center at La Timone Hospital between September 1986 and December 2008 and who were identified as follows: probands were all screened according to the French consensus guidelines , and had at least one episode of VT, underwent screening for thrombophilia, and had at least one of the inherited defects associated with thrombophilia (AT, PC or PS deficiency, and the FVL and F2 G20210A mutations). We selected only families for whom the inheritance of the coagulation defect was demonstrated (i.e. at least two family members, including the proband, had to be carriers of the same defect). A total of 1774 relatives from 500 families with inherited thrombophilia were recruited. We obtained a detailed medical history, with particular emphasis on previous episodes of VT and family history of VT. In relatives, investigators were careful to obtain the history of VT before the results of the thrombophilia screening were known. As with VT, we made note of any documented episodes of deep vein thrombosis or pulmonary embolism. VT was considered to be established if deep vein thrombosis was confirmed by compression ultrasound or venography, and pulmonary embolism was confirmed by by ventilation and perfusion lung scanning, spiral computer tomography scanning, or pulmonary angiography, or when the patient had received full-dose heparin and a vitamin K antagonist for at least 3 months.
In this study, 16 risk factors for VT, either established or putative, were recorded, and were of two kinds:
We set up groups depending on the severity of the thrombophilia, in accordance with the recommendations [4, 7]. Three groups were thus defined: ‘high-risk thrombophilia’, which includes subjects with deficiency of AT, PC, or PS, and homozygosity for FVL or F2 G20210A, and subjects with compound heritable thrombophilia (i.e. carrying more than one defect); low-risk thrombophilia', which includes subjects heterozygous for FVL or F2 G20210A; and ‘no thrombophilia’, which includes all patients without congenital thrombophilic defects.
Probands were excluded from the analysis to avoid bias. Of the 1774 relatives included in the cohort, 1201 had no missing data regarding the 16 parameters selected for the study. No significant difference was observed between the complete database and the database with no missing data used for the score (P > 0.05 for each variable; Table S1).
The institutional review board of the Assistance Public des Hopitaux de Marseille approved the study. Participants gave their informed consent.
Continuous variables are described by mean, median, and quartiles, and categorical variables by frequencies and percentages. Crude and adjusted odds ratios (ORs) and 95% confidence intervals (CIs) were estimated by mixed-effect logistic regression to account for the clustering (family). These results were useful for checking whether the known characteristics and the risk factors of the target population were observed in our sample.
Logistic regression requires log-linearity of continuous variables. When log-linearity assumptions were not met, continuous data were categorized according to deciles. In order to estimate the most parsimonious model, modalities with similar ORs were combined.
A multivariate mixed-effect logistic model (including fixed and random effects) was performed, including the 16 factors as predictors and VT as the outcome. The random effect was added to account for the non-independence between relatives, and therefore to adjust for clustering within families. The diagnostic capacity of the linear combination of factors provided by the model was estimated from the area under the non-parametric receiver operating characteristic (ROC) curve (AUC) and the associated CI. A risk factor selection procedure was then performed to retain only informative factors and to provide an easy-to-use score: all predictors were added one by one into the score, starting with the risk factor with the highest AUC, and continuing with factors with decreasing AUCs. The AUC of the model was re-estimated after the addition of each factor, and compared with the AUC of the previous model by use of the Delong non-parametric test for paired data . If the factor contributed significantly to an increase in the AUC, the factor was retained; otherwise, it was deleted. The final model retained only the predictors that significantly improved the AUC.
We used the regression coefficients of the mixed model for each variable from the final model as weights. These coefficients were rounded to propose a simplified score. The AUC was estimated with these rounded coefficients. The loss of AUC was thus deduced, and the final score was derived. A web calculator was used and is publicly available on www.bespim.fr/formulaire/marni.php.
To correct the AUC of the model for potential overfitting, optimism was estimated with bootstrapping. The algorithm presented by Harrell et al. was used . The variability of the list of selected predictors and the number of predictors over the 300 bootstrap samples were also determined.
All models were estimated with all available complete cases with regard to covariate information. All analyses were performed with r software (www.r-project.org).
The main characteristics of the 1201 relatives from 430 families are reported in Table 1. The main characteristics of the individuals with missing data were similar to those observed for the complete cohort (data not shown), which reduced the risk of selection bias.
|Female sex||730 (60.8)|
|Age at inclusion (years)||34.26 (17.9)|
|Age of relative* (years)||33.23 (17.1)|
|Body mass index (kg m−2)||23.1 (4.6)|
|ABO blood group|
|Severity of thrombophilia|
|No thrombophilia||465 (38.7)|
|Low-risk thrombophilia||528 (44.0)|
|High-risk thrombophilia||208 (17.3)|
|Individuals with VT||122 (10.2)|
|Age at first VT||41.41 (16.0)|
|Number of VTs|
|> 1||50 (41.0)|
|Type of first VT|
|DVT + PE||14 (11.5)|
|Unprovoked first event||56 (45.9)|
|Fibrinogen plasma level (g L−1)||3.17 (0.70)|
|FVIII plasma level (IU dL−1)||103.8 (38.9)|
|von Willebrand antigen (IU dL−1)||114.8 (47.7)|
In univariate analysis (Table 2), all of the preselected risk factors except gender, number of VT events in the proband and the circumstance of the first VT event of the proband were significantly associated with the risk of VT in relatives. It is of note that the F11 rs2036914-C and rs2289252-T alleles were both associated with an increased risk of VT in a dominant model (OR 2.21 [95% CI 1.28–3.88], and OR 1.57 [95% CI 1.00–2.49], respectively). Following an additive (or codominant) model, the FGG rs2066865-T allele was associated with an increased risk for VT of 1.76 (95% CI 1.22–2.57).
|Predictors||VT+ (n)||VT− (n)||Crude OR (95% CI)||Adjusted OR (95% CI)||P-value|
|Age of relative (years)*|
|20–57||93||635||7.2 (3.3–15.7)||6.7 (3.0–15.4)||< 0.00001|
|57–85||22||99||11.0 (4.5–26.4)||6.2 (2.4–16.4)||0.0002|
|Number of triggering circumstances|
|> 1||41||228||2.2 (1.4–3.5)|
|Number of VTs of proband|
|1 (n = 231)||51||513||1||NS|
|> 1 (n = 199)||71||566||1.3 (0.9–1.9)|
|Age at first VT of proband (years)|
|0–38 (n = 247)||86||657||1.6 (1.0–2.5)||1.8 (1.1–2.9)||0.0112|
|> 38 (n = 183)||36||422||1||1|
|Circumstance of first VT of proband|
|Unprovoked (n = 253)||47||425||1.0 (0.7–1.6)||NS|
|Provoked (n = 177)||75||654||1|
|Family history score (%)|
|0.1–0.3||69||376||3.8 (2.5–6.0)||3.5 (2.2–5.7)||< 0.00001|
|> 0.3||22||56||8.2 (4.5–15.1)||7.4 (3.7–14.8)||< 0.00001|
|Fibrinogen plasma levels (g L−1)|
|> 3.06||76||517||1.8 (1.2–2.6)|
|von Willebrand factor antigen (IU dL−1)|
|116–178||50||313||2.9 (1.8–4.5)||1.9 (1.1–3.2)||0.0254|
|> 178||33||76||7.9 (4.7–13.5)||3.8 (1.8–8.2)||0.0006|
|FVIII plasma levels (IU dL−1)|
|98–153||50||410||1.9 (1.2–3.1)||1.3 (0.8–2.3)||0.2971|
|> 153||35||91||6.5 (3.8–11.1)||2.4 (1.1–5.2)||0.0231|
|Severity of thrombophilia|
|Low risk||57||471||1.9 (1.2–3.1)||2.0 (1.2–3.4)||0.0087|
|High risk||37||171||3.5 (2.0–6.0)||3.7 (2.0–6.8)||< 0.00001|
|ABO blood group|
|A, B||79||660||1.6 (1.0–2.5)|
|Number of O or A2 allele|
|CT||61||489||2.1 (1.2–3.7)||1.9 (1.0–3.5)||0.0502|
|CC||45||320||2.4 (1.3–4.3)||2.1 (1.1–4.1)||0.0238|
|CT||52||398||1.5 (1.0–2.3)||1.7 (1.1–2.7)||0.0190|
|TT||19||78||2.9 (1.6–5.1)||3.0 (1.6–6.0)||0.0011|
In multivariate analysis (Table 2), a higher age of relatives was significantly associated with an increased risk of VT (OR 6.7 [95% CI 3.0–15.4] and OR 6.2 [95% CI 2.4–16.4] for 20–57 years and 57–85 years, respectively, as compared with age < 20 years). The age at first VT of the proband was also inversely associated with the risk of VT in relatives (OR 1.8 when the first VT occurred before the age of 38 years [95% CI 1.1–2.9]). A higher family history score was also significantly associated with the risk of VT in relatives (OR 3.5 [95% CI 2.2–5.7] and OR 7.4 [95% CI 3.7–14.8] for scores of 0.1–0.3 and > 0.3, respectively, as compared with a score of < 0.1). Among laboratory variables, the severity of thrombophilia remained associated with VT risk in relatives with ORs of 2.0 (95% CI 1.2–3.4) and 3.7 (95% CI 2.0–6.8) for the low-risk and high-risk thrombophilia groups, respectively, as compared with the group without any defect. An increased risk of VT was observed with increasing plasma levels of VWF and FVIII. For VWF, ORs for levels between 116 IU dL−1 (60th percentile) and 178 IU dL−1 (90th percentile) vs. < 116 IU dL−1 and > 178 vs. < 116 IU dL−1 were 1.9 (95% CI 1.1–3.2) and 3.8 (95% CI 1.8–8.2), respectively. ORs for FVIII levels between 98 IU dL−1 (median) and 153 IU dL−1 (90th percentile) vs. < 98 IU dL−1 and > 153 vs. < 98 IU dL−1 were 1.3 (95% CI 0.8–2.3) and 2.4 (95% CI 1.1–5.2), respectively. In this multivariate analysis, both F11 rs2036914 (P = 0.02) and FGG rs2066865 (P = 0.001) remained significantly associated with the risk of VT.
The AUC was independently estimated for each of the 16 variables (Table 3). The AUC ranged from 0.50 (95% CI 0.46–0.55) for unprovoked first VT of the proband to 0.69 (95% CI 0.64–0.73) for the family history score.
|1||Family history score (%)||0.6895||0.6446–0.7344|
|2||von Willebrand factor antigen (IU dL−1)||0.6847||0.6364–0.7333|
|3||FVIII plasma levels (IU dL−1)||0.6534||0.6031–0.7037|
|5||Severity of thrombophilia||0.6169||0.5682–0.6657|
|6||Number of triggering circumstances||0.5914||0.5413–0.6415|
|7||Number of O or A2 alleles||0.5909||0.5422–0.6396|
|9||ABO blood group||0.5788||0.5327–0.6249|
|10||Fibrinogen plasma levels (g L−1)||0.5719||0.5262–0.6176|
|12||Age at first VT of proband (years)||0.5480||0.5048–0.5912|
|15||Number of VTs of proband||0.5293||0.4827–0.5759|
|16||Unprovoked first VT of proband||0.5043||0.4586–0.5501|
The AUC derived from the clinical score alone, which is based on family history score, age, gender, number of triggering circumstances, and clinical characteristics of the proband, was 0.78 (95% CI 0.74–0.82) (Fig. 1). When the severity of thrombophilia was added to this clinical score, the AUC marginally (P = 0.07) increased to 0.80 (95% CI 0.76–0.84) (Fig. 1). When the other biological variables were further added, the AUC significantly (P < 0.0001) increased to 0.85 (95% CI 0.82–0.89) (Fig. 1).
Figure 2 shows the variables ordered by ascending diagnostic capacity and the corresponding AUC of the model when risk factors were added one-by-one. The P-values shown correspond to the comparison of successive models (for the first risk factor, the AUC is compared with 0.5). The discriminative accuracy of the models improved rapidly with the addition of each variable until five were included in the model (Fig. 2). These five variables, which form our proposed MARNI score, were family history score, VWF level, age of relatives, severity of thrombophilia, and FGG rs2066865. The AUC for this five-variable model was 0.83 (95% CI 0.80–0.87). The MARNI score model performed significantly better in terms of AUC than the model derived from the clinical score and the severity of thrombophilia (0.83 vs. 0.80, P = 0.002). The MARNI score also performed better than a model based on unmodifiable and easily available variables, including clinical score, thrombophilia, and ABO blood group (0.83 vs. 0.80, P = 0.01).
The AUC of the MARNI score with the original dataset was 0.83. The bootstrap-corrected performance of the original stepwise model was 0.82. Table 4 shows the frequency of appearance of each predictor over the 300 bootstrap samples. Three predictors contributed significantly to the c-index (AUC) in 100% of the bootstrap samples: family history score, VWF, and age. Thrombophilia severity and FGG rs2066865 were significant in several bootstrap samples (~ 40%). The other predictors were less frequently represented (< 20%). The AUCs from the bootstrap procedure were averaged over iterations with three, four and five significant factors. The mean AUCs were 0.80, 0.82 and 0.83 when three, four and five factors, respectively, were significantly associated with the risk of VT.
|Family history score (%)||300||100.00|
|von Willebrand factor antigen (UI dL−1)||300||100.00|
|Age of relative (years)||298||99.33|
|Severity of thrombophilia||128||42.67|
|Fibrinogen plasma levels (g L−1)||41||13.67|
|FVIII plasma levels (UI dL−1)||40||13.33|
|Age at first VT of proband||39||13.00|
|Number of triggering circumstances||27||9.00|
|Number of O or A2 alleles||23||7.67|
|Unprovoked first VT of proband||19||6.33|
|ABO blood group||14||4.67|
|Number of VTs of proband||13||4.33|
The observed and predicted risks of VT are shown in Fig. 3, according to the MARNI score. Up to a score of 5, the predicted and observed risk were very close, indicating that the MARNI score correctly identified patients with a low to moderate risk of VT. However, beyond a score of 5, the MARNI score tended to overestimate the risk of VT. This could be explained by the low number of relatives with a high score (see Fig. S1).
All of these indicators led us to validate our proposed MARNI score. Components of this score are reported in Table 5. This score ranged from 0 to 8. When subjects had scores between 4 and 6, their risk for VT was almost 20-fold (95% CI 9–55) higher than that of individuals with a score between 0 and 2 (see Fig. S1).
|Family history score* ≤ 10%||0|
|Family history score* 10–30%||1|
|Family history score > 30%||2|
|von Willebrand factor antigen (IU dL−1) ≤ 116||0|
|von Willebrand factor antigen (IU dL−1) 116–178||1|
|von Willebrand factor antigen (IU dL−1) > 178||2|
|Age (years) ≤ 20||0|
|Age (years) > 20||2|
|High-risk and low-risk thrombophilia||1|
|FGG rs2066865 (CC)||0|
|FGG rs2066865 (CT, TT)||1|
We generated a risk score based on 16 clinical and biological risk factors, which appeared to accurately differentiate between individuals with and without VT in families with inherited thrombophilia. A five-variable score appeared to discriminate as well as the initial score based on 16 variables. These five risk factors were the family history score, age, VWF level, severity of thrombophilia, and FGG rs2066865.
Predicting the risk of VT in families with inherited thrombophilia is often not feasible. Indeed, in a family with inherited thrombophilia, although those individuals who present with the defect have an increased risk of VT as compared with those without the defect, the presence or absence of the thrombophilic defect in an individual does not perfectly identify those at risk of VT. We thus investigated the extent to which a risk score, based on clinical and biological factors, can improve the accuracy of VT assessment by the use of ROC curves within these families. The major predictor of the risk of VT in our cohort of relatives was the family history score, with an AUC of 0.69. It has already been shown that family history of VT remains an important risk factor for the first VT event after adjustment for known variants . It was also recently reported that a higher number of affected relatives increased the chance of having a VT [13, 14]. This is certainly attributable to the fact that a family history represents the integration of risk within a family resulting from shared genetic susceptibilities and family clustering of environmental exposures, lifestyles, and behaviors . Consistent with previous studies [16-18], age was also an important clinical predictor of VT. This confirms that the risk of VT increased with age in families with inherited thrombophilia in both relatives with the thrombophilic defect and those without the thrombophilic defect. The validated risk score based on the five variables performed better than a clinical score including family history assessment and age, suggesting that risk assessment based only on clinical variables, as is often suggested, is not accurate enough to differentiate individuals with or without VT within these families. Thrombophilia screening within these families should not be dismissed, as knowing the severity of thrombophilia adds information to the clinical variables. The validated score based on the five variables performed better than the score including clinical variables and the severity of thrombophilia. Besides the family history score, age, and severity of thrombophilia, this score also includes VWF levels and FGG rs2066865. The importance of VWF level assessment within this score is striking. An increased level of VWF is a known risk factor for VT . The role of increasing VWF plasma levels in VT has been recently highlighted by the fact that variations in genes associated with VWF levels were found to be significantly associated with incident VT . VWF plasma level could be considered as an integrative parameter, as it is associated with other factors linked with VT, such as ABO blood group and FVIII level , and is also a marker of endothelial dysfunction and inflammation .
Identifying other modifying risk factors in these families will allow a more accurate definition of the individual risk within these families. We thus evaluated the impact of weak and common SNPs on the risk of VT within these families. We have recently demonstrated the contribution of ABO blood group in this cohort . In the present study, two SNPs located in the F11 gene and one in the FGG gene were also found to be associated with the risk of VT. F11 gene polymorphisms are associated with changes in FXI plasma levels. The SNP located in the FGG gene influences the level of the fibrinogen γ′-isoform, which has already found to be consistently associated with the risk of VT in unrelated VT cohorts . It must be stressed that the risk that these SNPs confer is of similar magnitude to that observed in populations of unrelated individuals, with no specific interaction with the type of thrombophilia (data not shown). However, of these three SNPs, only the one located in the FGG gene was an independent predictor of the risk of VT and significantly added to the present score.
In the field of VT, two attempts have already been made to develop a score that accurately predicts the occurrence of a first event. These recent studies showed that algorithms based on clinical risk factors  or clinical and genetic risk factors  were able to discriminate between patients and control subjects, with AUCs of 0.75 and 0.82, respectively. However, these algorithms have been tested in cohorts different from ours, as they consisted of unrelated individuals and not families, and did not take into account rare thrombophilias such as AT, PC and PS deficiencies.
Some limitations of the present study need to be mentioned. We used cross-sectional data, and the use of prospective data would have been more relevant for risk estimates. The score provided here was specifically designed to predict the risk of VT in families with known thrombophilia, and not in the general population. It would be valuable to assess the general applicability of the proposed MARNI score in families selected according to screening procedures that differ from the French consensus guideline . Our model had not undergone an external validation process based on replication in an independent cohort. However, it has undergone an extensive internal validation process similar to the one used for the Vienna prediction model for recurrence of VT . To validate the risk model, we used an internal validation procedure based on bootstrapping. A bootstrap sample of the same size as the original dataset was analyzed with a stepwise model using the same stopping rule as was used in the original score. The AUC was calculated, and the performance of this reduced model with the original dataset was estimated. The procedure was repeated 300 times, and the final AUC was penalized for model optimism. We can also not exclude the possibility that adding new SNPs to the model may have increased the AUC. However, our model included the five SNPs that were found to perform similarly in the prediction of VT in unselected individuals as a score based on 31 candidate SNPs selected from the literature . We have thus demonstrated that our model has included most of the information provided by common SNPs that are known to be associated with VT (AUC 0.83).
In conclusion, we have demonstrated that a risk score based on five parameters strongly predicts the risk of VT in families with inherited thrombophilia. The model should undergo external validation before it is applied in clinical practice.
W. Cohen, C. Castelli, P. Suchon, S. Bouvet, M. F. Aillaud, D. Brunet and M. C. Barthet collected, analyzed and interpreted data. W. Cohen, C. Castelli, M. C. Alessi, D. A. Trégouët and P. E. Morange designed the research study and wrote the manuscript.
We thank M. Billerey and F. Menaa for their excellent technical assistance. This study was funded by the Programme Hospitalier de Recherche Clinique (PHRC 2003 and AORC 2012) for the collection of the data.
The authors state that they have no conflict of interest.