## Haemostatic/inflammatory markers predict 10-year risk of IHD at least as well as lipids: the Caerphilly collaborative studies

## Abstract

**Aims** We compare the predictive values of plasma lipids (total and HDL-cholesterol, triglycerides) and three haemostatic/inflammatory risk markers for subsequent ischaemic heart disease (IHD).

**Methods and results** Two UK populations totalling 4860 men were screened for evidence of IHD between 1979 and 1983. Men were followed over 10 years and validated coronary events were recorded. Risk estimates were made using relative odds, receiver operating characteristic (ROC) curves and deciles of risk. Regression dilution effects were also examined. By 10 years, 525 men had a coronary event (fatal, non-fatal or silent myocardial infarction, MI). Two alternative multivariate models were compared – a lipid model (total, HDL-cholesterol, triglyceride) and a haemostatic/inflammatory model (fibrinogen, viscosity and white cell count). `Correction' for regression dilution increased relative odds for most risk factors. In the distribution of predicted risk, using established risk factors in conjunction with either lipid or haemostatic/inflammatory factors, the deciles of risk analysis showed that the observed 10-year risk of IHD was 34–35% in men in the top tenth, compared to 2–3% in the lowest tenth of the distribution.

**Conclusion** At the 10 years' follow-up, major, haemostatic/inflammatory risk factors showed a graded relationship to incident IHD that was at least as strong as that given by plasma lipids. Haemostatic/inflammatory factors provide possible additional targets for intervention.

- Inflammation
- Haemostasis
- IHD
- Lipids
- Risk models
- Epidemiology

## Introduction

Numerous epidemiological studies have identified the major classical^{1,2} and many putative^{3,4} risk factors for ischaemic heart disease (IHD). However, on an individual basis, the prediction of subsequent IHD risk from levels of lipids, blood pressure and smoking habit is poor.^{5} This may contribute to the strong scientific interest in the discovery of new risk predictors. Inflammatory and haemostatic processes are of intrinsic importance in coronary thrombosis.^{6,7} Many haemostatic/inflammatory factors have been investigated in prospective epidemiological studies, but meta-analyses suggest that levels of fibrinogen,^{8} white cell count^{9} and C-reactive protein^{9} are well established as risk factors.

We previously showed that fibrinogen, viscosity and white cell count were major risk factors for subsequent IHD at 5 years of follow-up.^{10} Some haemostatic/inflammatory factors seem to act only in the short term^{11} but we have shown that fibrinogen,^{12} viscosity^{13} and white cell count^{14} also predict in the longer term. In this present report we examine the relative strength of the predictive values of plasma lipids (total and HDL cholesterol, triglycerides) and these three haemostatic/inflammatory risk markers in the Caerphilly collaborative studies at 10 years of follow-up during which 525 new IHD events occurred. The effect of biological variation and regression dilution on the predictive ability of each of the models is also examined.

## Methods

### Study populations

In Caerphilly, a 100% sample of men was selected from within a defined area. They were aged between 45 and 59 years when first examined. A total of 2512 men were seen – 89% of the 2818 found to be eligible. In Speedwell, a 100% sample of men was selected from the age-sex registers of 16 general practitioners working from two neighbouring health centres. A total of 2348 men were seen during the following three years, 92% of those eligible. The men were aged 45–59 years when chosen, immediately before the study began in 1978; thus they were aged between 45 and 63 years when first examined. The combined cohort thus numbers 4860 men.

### Survey methods and follow-up procedure

The two studies had a common core protocol and common procedures, which have been described in detail elsewhere.^{10,15} The studies were each approved by the appropriate ethical committee and each subject gave informed consent. Briefly, at recruitment the men attended an afternoon or evening clinic during which a standard medical and smoking history was obtained. The London School of Hygiene and Tropical Medicine Chest Pain questionnaire was administered, height, weight and blood pressure were measured, and a 12-lead electrocardiogram (ECG) was recorded. They then returned after an overnight fast to an early morning clinic, where a blood sample was taken with minimal venous stasis. Fasting samples were obtained from 4641 men.

The 10-year follow-up in Caerphilly was at a nearly constant median interval of 119 months (IQ range 118–122) and was the second follow-up of the cohort. In Speedwell, results related to the third follow-up and the median interval was 112 months (IQ range 111–116).

The chest pain questionnaire was again administered at each follow-up and a further ECG recorded. All ECGs were coded using the Minnesota scheme by two experienced coders. The chest pain questionnaire was extended to include questions about hospitalisation for severe chest pain. These, together with the Hospital Activity Analysis notifications of admissions, coded as 410–414 – IHD in the ninth revision of the International Classification of Diseases (ICD), were used as the basis for a search of hospital notes. This was then checked by an epidemiologist for events which satisfied the World Health Organization (WHO) criteria for definite acute myocardial infarction (MI). For men who had died before the end of the follow-up, a copy of the death certificate was automatically received from the National Health Service Central Register. From this information, three categories of incident IHD events were defined: IHD death, clinical non-fatal (definite acute) MI and electrocardiographic MI, as previously described.^{10,15} A major IHD event was defined as one or more of the three possible outcomes described above.

### Laboratory methods

Haemostatic/inflammatory factors were measured in the same laboratories for both study areas. Due to the heavy workload separate laboratories had to be used for the lipid analyses of the two areas, but adjustment of the results was made on the basis of a comparability study using split samples assayed in both laboratories.^{15} Plasma samples were transported by rail to the laboratories on the day of venepuncture. Fibrinogen, viscosity and white blood cell count (WCC) were measured as previously described.^{10}

Lipid estimations were made using enzymatic methods; quality control and standardisation methods have been described in detail previously.^{15} In brief, the co-efficients of variation for duplicate samples measured blindly for the lipids (total cholesterol, HDL cholesterol and triglycerides) were 4%, 13% and 5%, respectively, and the haemostatic factors fibrinogen, viscosity and WCC were 10%, 2% and 3%, respectively. All presented results are based on plasma samples obtained after an overnight fast.

### Statistical methods

Mean differences in lipids and haemostatic/inflammatory risk factors between various groups (Table 1) were adjusted for age and area by analysis of co-variance. Total triglyceride and WCC were both logarithmically transformed prior to analysis. Multiple logistic regressions were performed with the occurrence or not of a major IHD event as the dependent variable. The initial model included the established non-lipid risk factors (age, pre-existent IHD, smoking, diastolic blood pressure and body mass index). Three further models were then obtained by adding to the initial model, irrespective of the significance of their contribution: (1) all lipid risk factors, (2) all haemostatic risk factors, and (3) all lipid and haemostatic risk factors.

The aim of the analysis was to determine whether the lipid, or the haemostatic, risk factors contributed more to other established risk factors. The slight difference in the duration of follow-up between the two studies was taken into account in the logistic regression analysis by the inclusion of study area as a co-variate in the model. Survival analysis was not used because no time-to-event was available for electrocardiographically defined myocardial infarctions which comprised 12% of the total incident IHD events. The odds of IHD for fifths of the distributions of lipid and haemostatic predictors relative to the fifth at lowest risk were calculated (Fig. 1). Likelihood ratio tests were used to check for deviations from linearity in these relationships. Thereafter, logistic regression was performed on the original values of the continuous predictors. Results are presented in the form of relative odds standardised to a one standard deviation change in predictors (Table 2).

Hosmer and Lemeshow goodness of fit tests were used to assess the fit of the logistic regression models by comparing the numbers of IHD events observed with the numbers predicted in sub-groups formed by deciles of predicted risk^{16} (Fig. 2).

To allow for the overall biological and laboratory variation in individual plasma risk factors, mean values for all Caerphilly subjects re-screened at 5-year follow-up were obtained. Speedwell samples during follow-up were taken non-fasting and could not therefore be used therefore for this purpose. Pearson product-moment correlation coefficients between the initial and 5-year samples were 0.68, 0.53 and 0.59 for the lipids (total cholesterol, HDL cholesterol and triglycerides), and for fibrinogen, viscosity and WCC were 0.39, 0.55 and 0.69; for diastolic blood pressure and body mass index the correlation coefficients were 0.53 and 0.91. Subjects were asked to fast overnight prior to venesection (as in the baseline examination 5 years previously) and blood samples were assayed in the same laboratory. Subjects who had died or did not attend for follow-up could not be included. These repeat measurements were used to calculate the regression dilution effects by the parametric method of Rosner et al.^{17}

Multivariate risk scores were calculated for participants from their characteristics using the co-efficients given by logistic regression. Sensitivity (true positive rate) and 1-specificity (false-positive rate) were calculated for all possible cut-off values across the risk score range. Sensitivity was then plotted against 1-specificity to produce a receiver operating characteristic (ROC) curve. The area under the ROC curve measured the predictive accuracy and may be interpreted as the probability that the score of a randomly chosen man who had an event during the 10-year follow-up exceeded that of a randomly chosen man who had not. An area of 0.5 is indicative of complete lack of discriminating power; an area of 1 indicates a risk score providing perfect discrimination. Areas were compared between models, taking into account their derivation from the same subjects^{18} (Table 3).

Significance tests were performed at the conventional 5% level (two-sided) with no correction for multiple comparisons.

## Results

All analyses were based on 4325 men who provided a fasting blood sample at recruitment and who had a complete set of data for all variables used in the analysis. Missing data were due to non-fasting status or unavailability of plasma or serum sample. These data were considered to be `missing at random' and therefore their omission should not produce any bias. By 10 years 525 subjects (12.1%) had experienced a major IHD event (fatal, non-fatal or silent MI).

The levels of lipids and haemostatic/inflammatory variables in subjects who experienced a major IHD event, or not, by 10 years are shown in Table 1. Age-adjusted mean levels of total cholesterol, triglyceride, fibrinogen, viscosity and WCC were significantly elevated among men who developed IHD over the 10-year follow-up period, while levels of HDL cholesterol were significantly lower.

Fig. 1 shows the relative odds of major incident IHD by fifths of the distribution of the lipid and haemostatic/inflammatory variables adjusted for both age and area. The odds rise steadily as fibrinogen increases so that in the top 20% the odds on an event are 3.19 (95% CI 2.32, 4.39) times the odds in the bottom 20%. For viscosity the corresponding relative odds are 3.30 (95% CI 2.40, 4.54) and for WCC 2.79 (95% CI 2.06, 3.77). Corresponding values for the lipids are: total cholesterol 2.07 (95% CI 1.55, 2.78), triglycerides 2.72 (95% CI 1.98, 3.72) and for the lowest fifth for HDL cholesterol 2.34 (95% CI 1.72, 3.19) relative to the highest fifth. For all variables, the associations with incident IHD showed no significant departures from linearity. The broken line indicates results for subjects with no evidence of IHD at baseline. The trends for the fifths of the distribution are very similar to those for the total cohort.

Multivariate analyses were performed to assess the contribution of putative risk factors adjusted for established risk factors and vice versa. In order that the relative strengths of the odds in these risk factors could be compared these have been standardised to represent the increase in odds associated with one standard deviation change in each of the continuously distributed variables. Categorical variables are presented as relative odds in the usual way (i.e., for each group compared to a baseline group). Table 2 shows these results in which lipids adjusted for standard non-lipid risk factors are compared to those for the haemostatic risk factors similarly adjusted.

All three lipids remain independently associated with risk of major IHD by 10 years after adjustment for non-lipid risk factors. In the haemostatic/ inflammatory model plasma viscosity and WCC remain independently associated with risk of subsequent IHD after full adjustment, whilst fibrinogen achieves borderline significance. The contribution of smoking is diminished in the haemostatic model compared to the lipid model.

Also shown in Table 2 are the relative odds after `correction' for regression dilution bias. All analyses are based on 4325 subjects but the `correction factors' were derived only from the repeated measurements in the Caerphilly cohort after 5 years of follow-up. Compared to the uncorrected figures, the relative odds increased for most of the variables that had repeated measurements available: total cholesterol, HDL cholesterol, total triglycerides, fibrinogen, viscosity, WCC, diastolic blood pressure and body mass index. Although the corrected odds ratios for fibrinogen and viscosity showed particularly large increases, these were accompanied by even larger increases in the standard errors of the associated regression co-efficients so that the two corrected odds ratios no longer differed significantly from one. Decreases in the relative odds of some of the other variables in the model were observed, particularly for the relative odds for smoking in the haemostatic model.

The predictive value of the models was compared formally by generating ROC curves for the two models and comparing the areas under these curves. The lipid model predicts subsequent IHD marginally less well (although not significantly so) than the haemostatic model (areas 0.724 and 0.728, respectively). The improvement in prediction obtained by the combined lipid and haemostatic model is small (area 0.737) but is significantly different from both the lipid model and the haemostatic model . Correction for regression dilution bias resulted in models with rather lower area values (0.711, 0.713 and 0.719 for the lipid, haemostatic and combined models, respectively). Table 3 also shows results for subjects without evidence of IHD at baseline; these show that the areas were again similar in the lipid and haemostatic models.

Relative odds provide little idea of the model's overall predictive power or ability to discriminate between low and high-risk subjects. We estimated the absolute risk for each model based on 10 sub-groups of men defined by deciles of predicted risk for the lipid and haemostatic risk factor models, incorporating into each model other risk factors (age, smoking, previous IHD, DBP and BMI). These results are summarised in Fig. 2. These show that both models produce similar values for absolute risk. The decile of men judged from their lipid values to be at highest risk experienced a 34.3% chance of developing a coronary event by 10 years (35.5% predicted) compared with 3.2% in the tenth at lowest risk (3.1% predicted). For the haemostatic model, the results were very similar. Correction for measurement error had the effect of weakening the overall prediction provided by both models. For example, the top decile of risk for the combined model decreased from a risk of 37.0% to 31.7% following correction for measurement error.

## Discussion

We have previously shown that fibrinogen, viscosity^{13} and WCC^{14} are powerful, long-term predictors of subsequent IHD. In this present report, we compare the predictive abilities of the lipid and haemostatic/inflammatory risk factors and demonstrate that the latter provide a powerful alternative model to the conventional lipid model. Others^{19,20} have also found that haemostatic factors show a similar predictive power to that shown by lipids. One study^{19} had a small number of outcome events and a different selection of haemostatic factors to that of the present report, whilst the other concluded that fibrinogen, WCC, factor VIII and von Willebrand factor were risk factors for coronary disease, but that their measurement in healthy adults added little to the prediction of coronary events beyond that of more established risk factors.^{20} However, a recent report from the PRIME study showed that fibrinogen rather than lipid risk factors accounted for much of the difference in risk of coronary disease between France and Northern Ireland.^{21} The present study supports the view that the haemostatic/inflammatory risk factors are of similar importance to the lipid factors, which are given major prominence in primary and secondary prevention strategies both in the United States^{22} and in Europe.^{23} These present results suggest that a potentially important component in the pathogenesis of IHD is currently under-examined and under-treated.

Although we adjusted for any possible effects of pre-existing IHD in these analyses in order to ensure that there were no residual effects due to treated or symptomatic IHD we re-ran the analysis on the group of men who had no clinical or historical (questionnaire) evidence of IHD at baseline examination. The pattern of results was very similar to that for the whole cohort of men although the areas under the ROC curves were smaller in value (Table 3). As in the analysis for the whole cohort both the lipid and haemostatic models predict significantly less well than the combined model.

We have previously shown the long-term consequences of smoking on these haemostatic/inflammatory risk factors, which did not appear to be entirely reversed after 10 years of stopping smoking.^{24} We examined in turn the influence of each of the haemostatic variables on the reduction in the relative odds of smoking associated with the haemostatic model in comparison to the lipid model. The addition of WCC reduced the relative odds of IHD for current cigarette smoking from 2.27 (95% CI, 1.63, 3.15) to 1.63 (95% CI, 1.15, 2.31) whilst the further addition of fibrinogen and viscosity did not further reduce the relative odds, suggesting that the main contribution to this effect was the WCC. This is consistent with the suggestion that, in susceptible individuals, inflammatory markers such as WCC may be elevated by smoking and may continue to act independently as long-term risk markers. This would not preclude the possibility that other stimulants to inflammation such as infectious agents^{25} may also contribute to this tendency in susceptible individuals. Furthermore, the complex interaction between inflammatory markers and thrombosis has previously been noted by our group^{26} and others.^{19} It has also recently been shown that treatment with statins lowers plasma viscosity, which may be one mechanism for their early beneficial effect on IHD risk.^{27} While statins probably lower viscosity by reducing plasma lipoprotein levels, a meta-analysis has shown that viscosity appears to be a predictor of IHD risk, independent of classical risk predictors including lipids.^{28}

The results obtained by applying a method for correcting for regression dilution bias were varied. Although most odds ratios showed the increases in magnitude that might have been expected after correction for measurement error, there seemed to be some undesirable interaction between the fibrinogen and viscosity corrections leading to the relative odds both increasing and becoming insignificant after correction. Dropping either variable from the correction process eliminated this phenomenon suggesting that it may be attributable to the high correlation between these two variables in our dataset .

When areas under the curve were calculated for the models corrected for regression dilution bias, the results were generally lower than for the uncorrected models. Likewise the deciles of risk analysis suggested that the corrected model performed less well than the uncorrected model in discriminating between men at low and high risk.

There are a number of possible reasons why the model that had been corrected for regression dilution bias failed to out-perform the standard uncorrected model. These could include the 5-year delay before repetition of the measurement, the fact that repeated measurements were only available in the Caerphilly cohort and repeated measurements were only available for those who survived at least 5 years. Of course, it is possible that the corrected model would have performed better relative to the uncorrected model had it been evaluated in a new dataset or applied to error-free data rather than to real data with its inherent measurement error.

Several practical issues need to be addressed when considering the addition of WCC, fibrinogen or viscosity to risk prediction scores. The first is additional cost, although each of these assays is relatively inexpensive, and the increased cost may be offset by more efficient targeting of costly risk modifications (e.g., statins). Second, C-reactive protein was not included in our model; however it was not a significant predictor of IHD in the Caerphilly cohort, after adjustment for classical risk factors and fibrinogen.^{29} Third, other haemostatic factors which may predict IHD^{19,20} were not included in our model; however, unlike fibrinogen, viscosity and WCC, they are not routine, robust, or inexpensive.

In this report, we also examined the effect of measurement error, i.e., overall biological and laboratory variation for the blood parameters tested. Measurement errors were similar for lipid and haemostatic/inflammatory variables. Furthermore the inclusion of corrections for measurement error in the multivariate models had the effect of weakening the overall predictive ability of the model which has not, to our knowledge, been reported elsewhere. Whilst acknowledging that measurement error should always be assessed and kept to a minimum, statistical methods for correction may not always result in better predictions of risk.

In conclusion, at 10 years of follow-up major haemostatic/inflammatory risk markers show a strong, graded relationship to incident IHD which is at least as strong as that shown by plasma lipids. Although these effects are independent of smoking habit, early cellular damage by smoking, perhaps acting in combination with other factors, may determine the level of haemostatic/inflammatory risk. These results suggest that fibrinogen.^{30} viscosity and WCC merit further assessment in risk prediction for IHD, and could also provide targets for future therapeutic interventions.

## Acknowledgments

This work was supported by a grant from the British Heart Foundation and the Medical Research Council. The latter has established the database and steering group for the former MRC Epidemiology Unit (South Wales) at the Department of Social Medicine, University of Bristol.

- The European Society of Cardiology