Plasma Fibrinogen Level and the Risk of Major Cardiovascular Diseases and Nonvascular Mortality:  An Individual Participant Meta-analysis FREE

Many prospective epidemiological studies have reported positive associations between the risk of coronary heart disease (CHD) and plasma fibrinogen levels.^1- 37 Fibrinogen is the major coagulation protein in blood by mass, the precursor of fibrin, and an important determinant of blood viscosity and platelet aggregation.^38- 41 Because fibrinogen levels can be reduced considerably by lifestyle interventions that also affect levels of established risk factors (such as regular exercise, smoking cessation, and moderate alcohol consumption), there is interest in the possibility that measurement (or modification) of fibrinogen may help in disease prediction or prevention.^38- 40,42 A meta-analysis of published data from 18 such studies, involving about 4000 CHD cases, indicated a relative risk of 1.8 (95% confidence interval [CI], 1.6-2.0) per 1-g/L increase in plasma fibrinogen level.⁴³ However, such analyses are not able to provide detailed assessments of the nature of any independent association of fibrinogen level with CHD or with other vascular and nonvascular outcomes.^43- 45

This meta-analysis differs from previous analyses in several ways that should increase its reliability and scientific value. First, it is large and comprehensive: the data comprise 6944 first nonfatal myocardial infarction (MI) or stroke events and 13 210 deaths (cause-specific mortality information is generally available among 154 211 participants in 31 studies). Second, individual records are available for all participants, allowing detailed analyses at different plasma fibrinogen levels and in different circumstances (such as by age, in men and women, in smokers and nonsmokers, and at different plasma lipid levels and blood pressure levels). Third, harmonization of individual records by baseline variables allows a consistent approach to adjust for potential confounding factors. Fourth, individuals with known preexisting CHD and stroke are excluded, limiting any effects of clinically evident disease on plasma fibrinogen level (ie, minimizing any “reverse association”).⁴⁶ Fifth, there is correction for within-person variations in plasma fibrinogen levels (ie, “regression dilution”⁴⁷) because 27 247 participants had serial measurements taken during follow-up. The main aim of this meta-analysis is to combine primary data from all relevant prospective studies to produce reliable estimates of the associations of plasma fibrinogen with CHD (and, secondarily, with stroke, other vascular mortality, and nonvascular mortality), incorporating adjustment for confounding caused by established cardiovascular risk factors and correction for regression dilution.⁴⁸

As described previously,⁴⁸ collaboration was sought from investigators of all published prospective studies among population-based samples (ie, participants were not selected on the basis of a previous history of cardiovascular disease) in which data on plasma fibrinogen level, date of birth (or age), and sex had been recorded at a baseline screening visit, and major vascular morbidity and/or cause-specific mortality and age at event or death had been routinely sought during at least 1 year of follow-up. Individual participant data were contributed from all relevant studies that were identified by previously published meta-analyses, through updated computer searches of databases including MEDLINE and Embase, and by extensive discussions with the investigators (referred to herein as collaborators).

This study was exempt from institutional review board approval by the Cambridgeshire ethics review committee because all 31 studies included were published previously and had each previously received local institutional review board approvals and consent from participants.

If available, data also were collected on history of heart disease, stroke, or diabetes; cigarette smoking habits (including, when available, current and previous smoking habits, amount smoked, and duration); alcohol consumption; systolic and diastolic blood pressure; serum lipid levels (including triglycerides, total, low-density lipoprotein and high-density lipoprotein cholesterol); leukocyte count; serum levels of albumin and C-reactive protein; and use of hormone treatments.

Information on cause-specific mortality was provided by collaborators in the greatest detail available. International Classification of Diseases coding to at least 3 digits was used and was based on death certificates (and all but one²³ of the contributing studies reported that medical records, autopsy findings, or other supplementary sources further helped to classify deaths).⁴⁸ Codes from versions 8, 9, and 10 of the International Classification of Diseases were accepted along with cohort-specific codes (details are available at: http://www.phpc.cam.ac.uk/FSC/docs/JAMA_2005_Webtab1.pdf). All studies reported use of standard definitions of MI (eg, clinical features, characteristic electrocardiogram changes, and marked elevations in blood levels of cardiac enzymes) and were typically based on criteria from the Monitoring Trends and Determinants in Cardiovascular Disease (MONICA) study or from the World Health Organization. In studies that distinguished cases of “possible,” “probable,” “suspect,” or “silent” MI from “confirmed” MI, any cases in the nonconfirmed categories were excluded from analyses. The large majority of studies reported diagnosing strokes on the basis of typical clinical features and characteristic changes on computed tomographic and/or magnetic resonance imaging brain scans, but only about half provided attribution of stroke subtype (such as cerebral ischemia, cerebral hemorrhage, or subarachnoid hemorrhage).⁴⁸ Prior to the conversion of the data from each study to a standard format for incorporation into a central database, the data were checked for consistency by the members of the secretariat (see list at end of this article) and any queries were referred back to the collaborators prior to the final harmonization of the data.

Because there are substantial differences in the mean (SD) values of fibrinogen across the contributing studies (partly due to differences in assay methods [involving clotting rate, clot weight, and various immunologic assays] and in fibrinogen standards [details are available at: http://www.phpc.cam.ac.uk/FSC/docs/JAMA_2005_Webfig1.pdf]), the present analyses involved comparisons of cases with noncases within each cohort. The prespecified unit of comparison in this meta-analysis was a 1-g/L increase in fibrinogen, with parallel analyses conducted for standardized increases in plasma fibrinogen (ie, per SD of fibrinogen values for each study) to allow for variations in fibrinogen levels across studies. Because the 2 approaches yielded similar patterns of findings, only the former are shown herein (the parallel analyses are available on request from the secretariat).

All analyses used the Cox proportional hazards model, stratified for study, sex, and, in the 2 studies^15,37 that randomized participants to different interventions, for trial group. The assumptions of the proportionality of hazards for fibrinogen level were satisfied in each of the 31 studies, which were assessed by plots of Schoenfeld residuals (data available on request). Each participant contributed only either the first nonfatal end point or death recorded at age 40 years or older (ie, deaths preceded by nonfatal CHD or stroke were not included in the analyses). These analyses were used in 2 ways to describe the dose-response relationships.⁴⁹ First, to investigate the shape of the association, the hazard ratios (HRs) for groups defined by fifths of baseline values of fibrinogen within each study were calculated and the 95% CIs were estimated from the floated variances that reflect the amount of information underlying each group (including the reference group).⁵⁰ A proportional or log scale was used for plotting disease rates. Second, if these analyses were consistent with a log-linear association, regression coefficients were calculated for the percentage change in risk associated with a 1-g/L difference in fibrinogen level. Heterogeneity of these associations was investigated with respect to mean age at baseline, other prespecified personal characteristics (eg, sex, smoking status, body mass index, blood lipid levels, and blood pressure level), and study characteristics (eg, geographical setting, sampling framework, and fibrinogen assay method). A random effect analysis also was used to take further account of between-study variation.

Analyses were adjusted for the effects of age at screening, smoking status, measured values of systolic blood pressure, total cholesterol level, and body mass index. More comprehensive adjustment for possible confounders was made in the subset of studies that had recorded details on alcohol consumption, subtypes of serum lipid levels, C-reactive protein, and history of diabetes. Statistical analyses were performed using STATA software version 8 (STATA Corp, College Station, Tex) and SAS software version 8 (SAS Institute Inc, Cary, NC). Analyses for the 2 nested case-control studies^3,22 (comprising a total of 439 CHD cases and 1554 controls) yielded essentially similar findings to those for the 31 cohort studies, albeit with much wider 95% CIs (data not shown). For results from individual studies, 99% CIs were used to make some allowance for the increased scope for random error in multiple comparisons and 95% CIs were used for combined analyses. Unless stated otherwise, all analyses have excluded the 1616 individuals with fibrinogen levels greater than 5.62 g/L (ie, the highest 1%) due to the potential distortions arising from assay imprecision at such levels⁵¹ and from acute-phase reactions.⁵²

Fifteen of the contributing studies^23- 37 provided data on 27 247 individuals and provided 2 or more fibrinogen level measurements. Most of the serial fibrinogen level measurements were taken within the first decade of follow-up. Baseline fibrinogen levels were used to divide individuals into fifths, and the ratio of the range between the mean values of the repeat measurements in the top compared with the bottom fifth was used to estimate regression dilution ratios for various intervals^42,47,49 (and similar results were obtained using the Rosner regression method⁵³; details are available at: http://www.phpc.cam.ac.uk/FSC/docs/JAMA_2005_Webfig2.pdf). The regression dilution ratio declined moderately with increasing intervals between measurements, with a value of about 0.5 for the first 5 years from the initial fibrinogen level measurement, but data were insufficient to estimate this value reliably for more prolonged durations of follow-up. Hence, the regression coefficient and SE relating risk to usual (ie, long-term average) fibrinogen level was estimated as approximately twice (ie, 1/0.5) the regression coefficient and its SE.

This meta-analysis includes 154 211 participants in 31 prospective studies with available individual records, no known history of previous CHD (ie, MI or angina, which was defined in each study⁴⁸) or stroke at baseline, fibrinogen levels of 5.62 g/L or lower, and any nonfatal CHD and stroke events or any deaths recorded for individuals aged 40 years or older. Of these participants, 115 658 (75%) were from Europe, 27 758 (18%) were from North America, and 10 795 (7%) were from Japan. During 1.38 million person-years of follow-up (mean [SD], 8.9 [4.9] years to first event or death), there were 4681 nonfatal MIs, 2263 nonfatal strokes, and 13 210 deaths comprising 2437 deaths attributed to CHD, 512 to stroke, 992 to other vascular causes (including 175 due to aortic aneurysm, 125 to heart failure, and 88 to acute pulmonary heart disease), 8007 to nonvascular causes (of which 4856 were due to cancer), and 1262 to unknown causes (details by cohort available at: http://www.phpc.cam.ac.uk/FSC/docs/JAMA_2005_Webtab2.pdf).

There were generally strong and highly significant positive associations between plasma fibrinogen levels and values of several established risk factors at baseline, such as age, cigarette smoking, serum levels of total and low-density lipoprotein cholesterol and triglycerides, blood pressure, body mass index, and history of diabetes (and, in the subset with data available on inflammatory factors, with leukocyte count and C-reactive protein values) (Table 1). There also were highly significant inverse associations between plasma fibrinogen level and male sex, alcohol consumption, and serum levels of high-density lipoprotein cholesterol and albumin.

There was a log-linear association with usual fibrinogen (plotted by fifths of values) for the risk of nonfatal or fatal CHD, nonfatal or fatal stroke, other vascular mortality (eg, non-CHD, nonstroke), and nonvascular mortality. There was no threshold in the range of available fibrinogen levels at which lower levels were no longer associated with lower risk (Figure 1). For usual fibrinogen levels, all 5 data points within each age group representing patients who had CHD, stroke, or nonvascular mortality are reasonably well fitted by the log-linear age-specific regression lines in Figure 1, although the trend is less distinct among the small numbers of participants who had other vascular deaths. The slopes of these regression lines were used to calculate the age-specific HRs for each of the major end points associated with a 1-g/L increase in usual fibrinogen level (Figure 2). The strengths of the association with usual plasma fibrinogen level declined to some extent with increasing age at event for CHD and for stroke. At ages 60 to 69 years, such differences in usual fibrinogen level were associated with about a 2-fold increase in risk for each of these outcomes. The HRs were about 50% more extreme for CHD and for stroke at ages 40 to 59 years compared with older ages, but the association with each of the major end points remained moderately strong even at older ages.

Associations of fibrinogen level with different vascular and nonvascular outcomes are shown in Figure 3 and Table 2. For any CHD, the HR per 1-g/L increase in usual fibrinogen level was 2.42 (95% CI, 2.24-2.60); there was a possibly higher HR of 2.68 (95% CI, 2.36-3.03) for fatal CHD compared with a HR of 2.30 (95% CI, 2.10-2.52) for nonfatal MI (χ²₁ for heterogeneity = 3.8; P = .05). For all types of stroke, the corresponding HR was 2.06 (95% CI, 1.83-2.33). For ischemic stroke, the HR was 2.08 (95% CI, 1.74-2.48); hemorrhagic stroke, 1.44 (95% CI, 1.05-1.96); and stroke attributed to unclassified causes, 2.36 (95% CI, 1.94-2.86) (χ²₂ for heterogeneity = 7.0; P = .03). The HR for other vascular mortality (eg, non-CHD, nonstroke) was 2.76 (95% CI, 2.28-3.35); nonvascular mortality, 2.03 (95% CI, 1.90-2.18); and unclassified deaths, 1.95 (95% CI, 1.64-2.31). The combined HR for all vascular deaths and events (eg, a combination of all CHD, all stroke, and other vascular deaths) was 2.35 (95% CI, 2.21-2.49), which was similar to that for all nonvascular mortality. There were too few cases of particular types of cancer to enable reliable analyses by type of cancer. Apart from a possibly stronger association with smoking-related cancer, subdivision of the aggregate of nonvascular deaths (most of which were cancer) generally did not suggest much more specific associations, although the association with deaths due to external causes was of marginal statistical significance. The HRs for cardiovascular end points were generally similar in analyses restricted to never smokers or to participants with first events (or deaths) recorded after the first 5 years of follow-up compared with the corresponding HRs involving all participants (Table 2).

Figure 4 and Figure 5 show the impact of a 1-g/L increase in usual fibrinogen level on CHD and stroke, respectively, by various study and individual-level characteristics after adjustment for age, sex, cohort, smoking, blood pressure, total cholesterol, and body mass index. For CHD (Figure 4), there was substantial heterogeneity among the 31 separate cohorts (χ²₃₀ = 84.5; P<.001; details by cohort available at http://www.phpc.cam.ac.uk/FSC/docs/JAMA_2005_Webfig3.pdf). Only some of the heterogeneity is explained by recorded characteristics. There was little evidence that the magnitude of the association between fibrinogen levels and CHD was importantly influenced by sex, blood pressure, or serum cholesterol level, a history of smoking or diabetes, or assay method. There was some evidence of heterogeneity among the findings of the 30 separate studies contributing stroke end points (χ²₂₉ = 48.7; P = .01; details by cohort available at: http://www.phpc.cam.ac.uk/FSC/docs/JAMA_2005_Webfig4.pdf), but there were no substantial variations in the associations by characteristics recorded (Figure 5). When the few studies that contributed most to the heterogeneity for each end point were excluded (ie, so that among the remaining studies P>.05 for heterogeneity for each end point), the combined HRs were little changed, suggesting the absence of important biases from the studies with more extreme results. Furthermore, use of a method that incorporates heterogeneity into the uncertainty around HRs calculated for each end point yielded only slightly wider 95% CIs than the fixed-effects model.⁵⁴

Table 3 shows the impact of progressive adjustment for baseline values of potential confounding factors on HRs. For CHD and stroke, the age-, sex-, and cohort-adjusted HRs decreased to 1.93 (95% CI, 1.79-2.08) and 1.75 (95% CI, 1.55-1.98), respectively, after adjustment for smoking, total cholesterol, blood pressure, and body mass index, and were 1.82 (95% CI, 1.60-2.06) and 1.82 (95% CI, 1.54-2.16), respectively, after further adjustment for alcohol consumption, concentrations of lipid subfractions, and history of diabetes in the studies with information on these variables. The substantial decreases in the χ² statistic for the association with fibrinogen level after these adjustments (from 246 to 87 for CHD and from 97 to 48 for stroke) suggests that a considerable part of the association was due to confounding. Moreover, because these confounding factors may have been measured with some error, substantial residual confounding may remain.⁵⁵ For nonvascular mortality (and, to some extent, for other vascular mortality), the effect of adjustment for potential confounding factors on HRs and on corresponding χ² statistic values was less marked (Table 3).

The Pearson partial correlation coefficient (adjusted for age) between fibrinogen and log C-reactive protein was 0.47 (95% CI, 0.45-0.49) in the 7011 participants with available C-reactive protein measurements. Among these 7011 participants, 263 CHD cases were recorded and the HR for CHD per 1-g/L increase in usual fibrinogen level was reduced from 2.23 (95% CI, 1.59-3.14) to 1.78 (95% CI, 1.19-2.66) after adjustment for age, sex, several established coronary risk factors, and C-reactive protein.The numbers of other vascular and nonvascular disease outcomes were too few in this subset to enable reliable analyses adjusted for C-reactive protein. Additional figures are available at http://www.phpc.cam.ac.uk/FSC/FSC_index.htm.

The present meta-analysis involves individual participant data from 31 prospective studies of major cardiovascular diseases and nonvascular mortality among 154 211 individuals without known cardiovascular disease at baseline. It provides the first reliable demonstration that fibrinogen is associated with the age-specific incidence rates of CHD, stroke (especially nonhemorrhagic stroke), other (eg, non-CHD, nonstroke) vascular mortality, and, interestingly, of the aggregate of all nonvascular causes (mainly comprising cancer). These data demonstrate that throughout the range of fibrinogen levels recorded in Western populations, the proportional differences in risk of each of these end points associated with a given absolute difference in usual fibrinogen are generally similar at all fibrinogen levels (ie, the relationships are approximately log linear). The present meta-analysis indicates that the associations of fibrinogen level with CHD or with stroke do not differ substantially by baseline levels of established risk factors such as sex, smoking, blood pressure, serum lipid levels, or several features of study design.

The associations between fibrinogen level and the incidence of a broad range of different chronic diseases are particularly striking, including not just major ischemic cardiovascular diseases but also the aggregate of nonvascular mortality.^6,13 The magnitude of the associations persisted largely unchanged in analyses restricted to never smokers and to disease cases recorded several years after the baseline examination, reducing the likelihood that they were mainly due to cigarette smoking habits and/or early cardiovascular disease. It has been suggested that such associations reflect a response to cumulative environmental stressors (as indicated by circulating levels of fibrinogen and other inflammatory factors), which may modify the risk, progression, and outcomes of various chronic diseases.⁵⁶ Although the lack of specificity in the associations of fibrinogen level with different vascular and nonvascular outcomes does not necessarily exclude a causative role for fibrinogen in ischemic cardiovascular diseases, there remains scope in these estimates for biases due to residual (or unmeasured) confounding by other factors.

Fibrinogen levels were correlated with several established risk factors (such as blood pressure and serum cholesterol levels). After adjustment for baseline values of such risk factors, the HRs for CHD and stroke each reduced to about 1.8 per 1-g/L increase in usual fibrinogen level. These confounding factors probably have been measured with some error so substantial residual confounding may remain.⁵⁵ For example, the within-person correlations on measurements of blood cholesterol levels and blood pressure values taken a few years apart are only about two thirds for each.⁴⁷ Hence, adjustment of the relationship between fibrinogen level and CHD or stroke for the measured values of smoking, blood pressure, and serum cholesterol level (and other confounders) may have only reduced the associations by approximately two thirds of the amount that the adjustment for usual fibrinogen levels would have done.Moreover, more complete adjustment may have been achieved by correction for variables that more directly reflect the relevant confounding factors and/or preexisting disease (such as apolipoproteins B and A1 rather than direct measurement of cholesterol subfractions,⁵⁷ or indices of subclinical atherosclerosis [eg, the ankle-brachial pressure index⁵⁸ or carotid intimamedia thickness⁵⁹]), but such considerations are difficult to quantify. After adjustment for established risk factors, the association between fibrinogen level and risk of CHD was essentially unchanged following additional adjustment for C-reactive protein in the 7011 participants with available C-reactive protein measurements. Further investigation of this association in larger numbers of participants with both fibrinogen levels and C-reactive protein values will be required to assess this more reliably.

Assessment of the likelihood of any causal relationships of fibrinogen level with disease will require research strategies that can minimize potential residual biases, such as investigation of the genetic determinants of fibrinogen level in large-scale observational studies (ie, “Mendelian randomization”^60- 61 and B. Keavney, MD, J. Danesh, DPhil, S. Parish, DPhil, et al, unpublished data, 2005) or randomized trials of selective fibrinogen-lowering agents. This meta-analysis provides the most comprehensive and detailed description, to date, of the observational epidemiological associations between plasma fibrinogen level and a range of different chronic disease outcomes and demonstrates moderately strong associations in a wide range of circumstances.