OUP user menu

Do inflammatory biomarkers add to the discrimination of cardiovascular disease after allowing for social deprivation? Results from a 10-year cohort study in Glasgow, Scotland

Mark Woodward, Paul Welsh, Ann Rumley, Hugh Tunstall-Pedoe, Gordon D.O. Lowe
DOI: http://dx.doi.org/10.1093/eurheartj/ehp115 2669-2675 First published online: 10 April 2009


Aims To assess the additional discriminative value of adding each of five inflammatory biomarkers to the ASSIGN risk score, which includes social deprivation.

Methods and results In this study, 1319 men and women aged 25–64 in the fourth Glasgow MONICA study were followed-up for cardiovascular endpoints. Baseline C-reactive protein, fibrinogen, IL-6, IL-18, and TNFα were related to risk of CVD. The discriminative value of adding each to the ASSIGN score was assessed using area under the receiver operating characteristic (AUC) and relative integrated percentage improvement in classification (RIDI). During a median of 10.5 years, 151 CVD events occurred. After adjusting for ASSIGN variables, each inflammatory marker except IL-18 had a significant (P < 0.05) association with CVD risk. The AUC using ASSIGN [0.799 (95% CI 0.790–0.809)] was improved by the inclusion of C-reactive protein and TNFα [0.805 (95% CI 0.795–0.815); P < 0.03], but not by other combinations. C-reactive protein and TNFα yielded a significant RIDI (IL-6 almost so). C-reactive protein and TNFα together improved the classification of risk by 11% (95% CI, 3–19%) when added to the ASSIGN variables.

Conclusion Some inflammatory biomarkers add moderate discriminative information to the ASSIGN CVD risk score. The clinical utility of this information, cost-effectiveness, and optimization should be assessed in future studies.

  • Cardiovascular disease
  • Risk prediction
  • Inflammation
  • C-reactive protein
  • Fibrinogen
  • Social deprivation


Meta-analyses of prospective studies have shown that plasma levels of the inflammatory variables fibrinogen,1 high-sensitivity C-reactive protein2 and interleukin (IL)-63 have independent associations with risk of coronary heart disease (CHD). Although it has been suggested that two other pro-inflammatory cytokines, interleukin-18 (IL-18) and tumour necrosis factor alpha (TNFα), may also contribute to atherosclerosis,48 the epidemiological evidence for their associations with cardiovascular disease (CVD) is currently weak. Only a limited number of prospective studies, giving the most reliable evidence of epidemiological associations,9 have been published for IL-181014 or TNFα.1518

Previously, we reported cross-sectional associations between fibrinogen, C-reactive protein, IL-6, IL-18, and TNFα and a number of demographic and lifestyle factors, including classical risk factors for CVD, in a general population study from Scotland.19 Here we use the follow-up data from this study to explore the associations between these five inflammatory variables and risk of CVD, and to explore the potential utility of these markers over and above the ASSIGN CVD risk score,20 recommended for use in Scotland,21 which includes a marker of social deprivation, as well as family history. This step is important because social deprivation and family history are also under investigation for use in risk scores in the rest of the UK,22 inflammatory markers are highly associated with social deprivation,19 and it is presently not clear if inflammatory markers would add to such risk scores.


Details of the fourth MONICA survey in north Glasgow have been described previously.19 Briefly, this was a random population sample of 1836 men and women, aged 25–64 years old, carried out in 1995. Participants completed a questionnaire, which solicited information on demographics, past medical history, and lifestyle, including tobacco use. Blood pressure, weight, and height were measured at special clinic sessions, where a 12-lead electrocardiogram was applied and a blood sample taken, anticoagulated with 0.11 M trisodium citrate (9:1 v:v), centrifuged, and aliquots stored frozen at −80°C. Fibrinogen was measured by the automated Clauss method (Organon Teknika, Cambridge, UK), C-reactive protein by an immunonephelometric method (Dade Behring, Milton Keynes, UK), and high sensitivity IL-6, IL-18, and high sensitivity TNFα by ELISA assays (R&D systems, Abingdon, UK).19

Subjects were positive for CVD at baseline if they reported a previous doctor diagnosis of angina, heart attack or stroke or previous coronary artery bypass graft (CABG), coronary angiogram or percutaneous coronary angioplasty (PTCA); if the ECG was suggestive of myocardial infarction using Minnesota codings;23 or if the national records (see in what follows) showed that they had been hospitalized for CHD or stroke or had undergone any of the three procedures listed earlier within the previous 24 years.

Subjects who gave written permission were followed-up for deaths through national death registration and record linkage databases. The latter was also used to follow-up, and trace back, hospitalizations. Records were obtained for the period from 1981 to 2005.20 The protocol complies with the Declaration of Helsinki.

Statistical methods

The primary outcome was pre-defined to be any cardiovascular event during follow-up, over a median of 10.5 years. Two people in the database were censored early due to being lost to follow-up. This outcome encompassed death due to CVD [International Classification of Diseases (ICD) 9 codes 390–459, ICD 10 codes I00–I99], any hospital discharge diagnosis post-recruitment (up to six diagnoses were recorded) which included CHD (ICD version 9, 410–414 or version 10, I20–I25, or cerebrovascular disease (ICD-9 430–438 or ICD-10 G45, I60–I69),24,25 and surgical codes for CABG or PTCA. Two secondary outcomes were defined: incident CVD (i.e. analyses as earlier, but excluding those with baseline CVD) and any CHD event during follow-up: all coronary events within the list earlier.

Only those subjects with complete information on all variables were included in any analyses. Cox models were used to find hazard ratios (HRs), adjusted for age and sex, comparing the highest and middle thirds to the lowest, for each inflammatory variable. The proportional hazards assumption was tested using log cumulative hazards plots.9 Since the HRs suggested, as expected from previous data, a log-linear increase in the risk of CVD with increasing values of each biomarker, further analyses took each biomarker as a continuous variable.9 For a one inter-quartile range (IQR) increase, HRs for the primary outcome were obtained: adjusted for age and sex alone; further adjusted for all the variables used in the ASSIGN cardiovascular risk score; and then further adjusted for all the other inflammatory markers. The measure of social deprivation used in ASSIGN is the Scottish Index of Multiple Deprivation, which ranges from 0.54 to 87.6, with a higher score indicating relatively more deprivation.20 As a sensitivity analysis, the steps above were repeated using variables from the Framingham risk score,26 rather than ASSIGN. Hazard ratios for the secondary outcomes were calculated after adjustment for ASSIGN variables only.

Areas under receiver operating characteristic curves (AUCs)9 were estimated, using 500 bootstrap samples, for eight pre-defined models, each of which included all the ASSIGN variables: that with no inflammatory biomarkers; those with a unique single inflammatory biomarker; that with all the significant (P < 0.05) inflammatory variables in the model for the primary outcome that included all five biomarkers; and that with all five inflammatory biomarkers. In all models, quantitative variables were treated as continuous. To avoid bias otherwise caused by testing discrimination on the same data from which the prediction score was derived, each bootstrap sample was used to determine the prediction score for the particular model which was then tested for discrimination on the observed data. Wald tests,9 computed from the individual bootstrap results, were used to compare AUCs for both the model with all significant inflammatory biomarkers and that with all the inflammatory biomarkers against the model with no inflammatory biomarkers—these being the only pre-defined tests. The ability of each inflammatory marker, the combination of ‘significant’ markers, and all five markers to correctly reclassify those with and without CVD events, compared with using ASSIGN alone, was quantified through the relative integrated discrimination improvement (RIDI).27 The RIDI measures the percentage improvement in classification, averaged over all potential thresholds for defining who is at sufficiently high risk for intervention. Confidence intervals and tests were constructed from 500 bootstrap results.

The expected clinical utility of adding the ‘best’ combination of inflammatory markers (as determined by the above analyses) to those already used in ASSIGN was estimated for subjects who are eligible for application of the ASSIGN risk score—those aged 30+ years, free of CVD. The classification of subjects by ASSIGN into four CVD risk categories (<10, 10 to <15, 15 to <20, and 20% + 10 year risk) was compared with the same classification using ASSIGN variables plus the best inflammatory biomarkers.

All statistical tests were two-sided except for the one-sided tests of the RIDI values, which investigate whether an improvement in classification has been achieved. Tests were considered significant if P < 0.05.


Of the 1836 study participants sampled and evaluated, 1319 had valid measures for all variables in ASSIGN and all five inflammatory biomarkers, and gave full consent for follow-up. Table 1 shows summary statistics, at baseline, for all relevant variables.

View this table:
Table 1

Summary statistics, all subjects (n = 1319)

Continuous variables
 Age (years)43.611.543345319
 Systolic blood pressure (mmHg)126.718.712411313724
 Serum total cholesterol (mmol/L)5.881.155.825.056.621.58
 Serum HDL-cholesterol (mmol/L)1.330.391.291.051.530.48
 SIMD score44.224.145.623.365.442.2
 Fibrinogen (g/L)3.110.922.902.463.551.09
 C-reactive protein (mg/L)2.744.481.300.643.222.58
 Interleukin-6 (pg/mL)2.352.181.711.172.641.47
 TNFα (pg/mL)1.770.971.591.282.030.75
 Interleukin-18 (pg/mL)247139221169291122
Binary variablesNumber%
 Female sex68952
 Family history of CHD44534
 Cigarette smoker53140
  • aCigarette smokers only.

  • SMID, Scottish index of multiple deprivation (higher score means more deprived); SD, standard deviation; Qt, quartile; IQR, inter-quartile range; CHD, coronary heart disease.

One hundred and fifty-one CVD events occurred during follow-up. Each of the inflammatory variables showed a monotonic trend in CVD risk across their thirds after age and sex adjustment (Figure 1), with the HR comparing the extreme thirds for C-reactive protein (2.46) being much higher than the other four biomarkers, which were all well below two. All five HRs comparing extreme thirds were significant. However, after age and sex adjustment, fibrinogen had the highest HR for a linear (one IQR) increase, while the linear effect of IL-18 was not significant (Table 2). Further adjustment for the ASSIGN risk set, followed by cross-adjustment for the other inflammatory biomarkers, either attenuated the HRs progressively or left them much as at the previous level of adjustment, except in the case of TNFα, for which the HR increased after cross-adjustment. As after age/sex adjustment, all but IL-18 were significant after adjustment for ASSIGN variables, but only C-reactive protein and TNFα were significant when further adjusted for the other inflammatory biomarkers. The HRs for the two secondary outcomes were similar to those for the primary outcome. Furthermore, the conclusions were exactly the same when the ASSIGN variables were replaced by those used in the Framingham CVD risk score.

Figure 1

Hazard ratios for cardiovascular disease (with 95% confidence intervals) by thirds (first third = reference) for each inflammatory variable, after adjustment for age and sex. The vertical axis uses a logarithmic scale and hazard ratios are plotted against the median of each third. The tertiles were 2.58 and 3.30 g/L for fibrinogen; 0.82 and 2.31 mg/L for C-reactive protein; 1.32 and 2.28 pg/mL for interleukin-6; 186 and 262 pg/mL for interleukin-18; and 1.37 and 1.83 pg/mL for TNFα.

View this table:
Table 2

Hazard ratios (95% confidence intervals) for a one inter-quartile range increase in each inflammatory variable

CVD (all subjects, n = 1319)Incident CVD (n = 1187)CHD (all subjects, n = 1319)
Age/sex adjustedMultiple adjustedaCross-adjusted bMultiple adjustedaMultiple adjusteda
Fibrinogen1.33 (1.14–1.55)1.21 (1.02–1.43)1.07 (0.88–1.31)1.29 (1.04–1.60)1.14 (0.94–1.37)
C-reactive protein1.12 (1.07–1.18)1.12 (1.06–1.18)1.04 (1.01–1.07)1.14 (1.06–1.23)1.13 (1.07–1.20)
IL-61.11 (1.03–1.20)1.12 (1.03–1.22)1.01 (0.93–1.09)1.12 (1.00–1.26)1.14 (1.05–1.25)
TNFα1.11 (1.03–1.20)1.11 (1.03–1.21)1.16 (1.03–1.30)1.12 (1.00–1.25)1.09 (0.99–1.21)
IL-181.09 (0.97–1.24)1.02 (0.88–1.18)1.000 (0.999–1.001)0.98 (0.82–1.19)1.04 (0.89–1.21)
  • aAdjusted for age, sex, systolic blood pressure, serum total and HDL cholesterol, cigarettes/day, SIMD score, diabetes, and family history of coronary heart disease.

  • bAdjusted as for multiple adjustment plus all other variables in this table.

  • CVD, cardiovascular disease; CHD, coronary heart disease.

Table 3 shows the bias-corrected AUCs for the primary outcome. These ranged from 0.799 for the model with only the ASSIGN risk set to 0.805 for the model with ASSIGN variables plus the two inflammatory variables that were significant in the cross-adjusted model of Table 2 (C-reactive protein and TNFα). Although this difference is modest, it was statistically significant (P = 0.03), while 97% of the bootstrap evaluations found the larger model to have better discrimination than the smaller. The difference between the ASSIGN-only model and the full model with ASSIGN variables and all inflammatory biomarkers was about one-half of the above difference, and was not significantly different from zero (P = 0.35). The model with highest AUC among all those with ASSIGN variables plus only one inflammatory variable was that with C-reactive protein: AUC = 0.803, the same as the ‘full model’. The RIDI was significant for C-reactive protein and TNFα, and virtually so for IL-6, but not IL-18 or fibrinogen. The combination of TNFα and C-reactive protein yielded 11% (P = 0.004) improvement in classification, compared with the ASSIGN variables alone. The combination of all five inflammatory markers gave a 12% improvement (P = 0.006) in classification.

View this table:
Table 3

Areas under receiver operating characteristic curves and integrated relative percentage improvement in classification for predictive scores for cardiovascular disease in all subjects (n = 1319)

Inflammatory variablesAUCRIDI
Estimate(95% CI)Estimate(95% CI)P-value
None0.799(0.790, 0.809)
Fibrinogen0.799(0.789, 0.809)3.34(−2.02, 8.71)0.11
C-reactive protein0.803(0.793, 0.813)7.16(0.82, 13.50)0.01
IL-60.800(0.790, 0.810)5.30(−0.98, 11.59)0.05
IL-180.799(0.789, 0.809)0.09(−0.96, 1.14)0.43
TNFα0.802(0.792, 0.813)4.85(0.06, 9.65)0.02
C-reactive protein+TNFα0.805(0.795, 0.815)11.24(3.06, 19.42)0.004
All 50.803(0.791, 0.814)12.05(2.76, 21.34)0.006
  • Every score includes age, sex, systolic blood pressure, serum total and HDL cholesterol, cigarettes/day, SIMD score, diabetes, and family history of coronary heart disease.

  • P-values from one-sided test.

  • CI, confidence interval; AUC, areas under receiver operating characteristic curves; RIDI, integrated relative percentage improvement in classification.

Taking C-reactive protein and TNFα as, thus, the best combination of inflammatory biomarkers to add to ASSIGN, the effects of such an addition, among the sub-population within the MONICA study eligible for the ASSIGN risk score, are shown in Table 4. For those with CVD at 10 years, 21 are reclassified, using C-reactive protein and TNFα, into better (i.e. higher) CVD risk categories and 12 are reclassified into worse (i.e. lower) risk categories. The net result is that 9 (11%) of the 81 subjects with CVD are classified better when C-reactive protein and TNFα are taken into account. For those without CVD, 121 are classified into better (i.e. lower) risk categories and 125 into worse (i.e. higher) risk categories when C-reactive protein and TNFα are taken into account, giving a net worsening of 4/929 (0.43%). To put this in a general context, consider a population with a million people aged 30 or more who are free of CVD, and assume that, as in the MONICA population, 10% will go on to develop CVD in the next 10 years. Then, compared with using ASSIGN, an additional 11 000 of those who will go on to develop CVD, and an additional 3870 of those who will not go on to develop CVD, can be expected to be identified as being in a higher risk category when information on C-reactive protein and TNFα are taken into account, giving a net improvement, in purely numerical terms, of 7100 classifications.

View this table:
Table 4

Classification of 10 year cardiovascular disease risk using ASSIGN compared with classification using a new score which adds C-reactive protein and TNFα to ASSIGN variables among subjects without cardiovascular disease and aged 30 or more at baseline

Classification using ASSIGNClassification using ASSIGN variables plus C-reactive protein and TNFα
<10%10 to <15%15 to <20%20%+Total
All subjects (n = 1010)
 <10%530 (89)42 (7)12 (2)10 (2)594
 10 to <15%37 (27)55 (40)27 (20)17 (13)136
 15 to <20%18 (18)25 (25)19 (19)38 (38)100
 20%+12 (7)16 (9)25 (14)127 (71)180
Subjects with CVD within 10 years (n = 81)
 <10%14 (70)3 (15)2 (10)1 (5)20
 10 to <15%2 (14)6 (43)3 (21)3 (21)14
 15 to <20%2 (13)1 (7)3 (20)9 (60)15
 20%+1 (3)2 (6)4 (13)25 (78)32
Subjects without CVD at 10 years (n = 929)
 <10%516 (90)39 (7)10 (2)9 (2)574
 10 to <15%35 (29)49 (40)24 (20)14 (11)122
 15 to <20%16 (19)24 (28)16 (19)29 (34)85
 20%+11 (7)14 (9)21 (14)102 (69)148
  • Those classified better using the score with C-reactive protein and TNFα are shown in bold; those classified worse are shown in italics (these fall either side of the diagonal in the lower two sub-tables).

  • CVD, cardiovascular disease.


In a general population sample, within a population with relatively high rates of CVD, we have compared plasma levels of three pro-inflammatory cytokines (TNFα, IL-6, and IL-18) and two ‘downstream’ inflammatory markers (C-reactive protein and fibrinogen) and risk of CVD. We report that each of these five inflammatory biomarkers has approximately a log-linear association with the risk of CVD. Associations, on a standardized scale, were stronger for fibrinogen, but we found that adjustment for confounding variables explained much of this extra strength. Only C-reactive protein and TNFα had associations with CVD that were independent of the other inflammatory variables.

Importantly, we found evidence that inflammatory biomarkers add to the discrimination of CVD risk, over and above those used in the ASSIGN score, which is recommended by national guidelines for CVD prediction in Scotland.21 In our data, when C-reactive protein and TNFα were added to the ASSIGN score risk set, the discrimination was improved slightly, but significantly, according to both the AUC and RIDI statistics. This is the first evidence that inflammatory markers may add discrimination over and above classical risk factors plus social deprivation (as well as family history), as used in the ASSIGN risk score.

Demonstration that inflammatory markers can improve discrimination of the ASSIGN score (albeit moderately) is important, given the strong correlation between inflammatory variables and social deprivation in this study19 and elsewhere.28,29 However, given the high correlation between the five inflammatory variables,19 we cannot make any firm conclusion that C-reactive protein and TNFα are more useful than the other inflammatory markers.

Although our study is the first to assess additional discrimination of inflammatory markers in any risk score which includes social deprivation, previous studies have assessed discrimination in more classical risk scores. These studies have generally reported that the AUC for such risk scores has been moderately improved by the addition of C-reactive protein or fibrinogen,3032 although some report no significant improvement in discrimination using C-reactive protein or fibrinogen.33,34 Indeed, in the ARIC study of 19 novel biomarkers, no single biomarker (including C-reactive protein or IL-6) added much discrimination to a basic risk assessment model.35 Our findings are broadly in line with these studies.

Strengths of this study include: the general population source, the prospective design, with 10+ years follow-up, and the high quality of survey procedures used within MONICA.36 We have used both the traditional AUC method of discrimination and two recent approaches based on reclassification.27 Limitations include the lack of repeat measures over time and the moderate sample size. The former means we cannot adjust for regression dilution bias; however, previous meta-analyses for fibrinogen, C-reactive protein, and IL-6 suggest that the attenuation coefficients for these variables are between 0.35 and 0.55.13 These are the divisors for the slope of the line associating the log of the HR and each risk factor which corrects for regression dilution bias.9 No such meta-analyses have been performed for IL-18 or TNFα, to date. The second limitation has led us to use all the available data in each analysis, including taking all subjects, rather than just those without CVD at baseline. Our secondary analysis of incident CVD suggests that this has not caused any bias. Another limitation is that the mean age of those missing from the entire fourth MONICA study is 5 years higher than those included here, with consequent differences in many risk factors and a 5% higher CVD event rate. This age difference seems to be due to a greater chance of incomplete blood samples, rather than lack of blood samples entirely, with increasing age, since there was little difference in ages between those with and without cholesterol measures (the first assays done). We know of no systematic mechanism that may have led to this; as far as we are aware this is a chance finding. In any case, relative risks and AUCs are often estimated from non-random samples, assuming lack of bias. We have no reason to suppose any bias arises from missingness in the estimates and tests presented here, even if we cannot rule it out.

In conclusion, we report for the first time that the addition of two inflammatory markers (TNFα as well as C-reactive protein) may improve discrimination in the ASSIGN clinical risk score. The clinical utility of this information, cost-effectiveness, and optimization of the signal using combinations of inflammatory markers, should be assessed in further studies.


We thank Helen Mosson for secretarial assistance in preparing the manuscript.

Conflict of interest: none declared.


View Abstract