The mediation of coronary calcification in the association between risk scores and cardiac troponin T elevation in healthy adults: Is atherosclerosis a good prognostic precursor of coronary disease?

BACKGROUND: Conventional cardiac risk scores may not be completely accurate in predicting acute events because they only include factors associated with atherosclerosis, considered as the fundamental precursor of cardiovascular disease. In UK in 2006-2008 (Whitehall II study) we tested the ability of several risk scores to identify individuals with cardiac cell damage and assessed to what extent their estimates were mediated by the presence of atherosclerosis.
METHODS: 430 disease-free, low-risk participants were tested for high-sensitivity cardiac troponin-T (HS-CTnT) and for coronary calcification using electron-beam, dual-source, computed tomography (CAC). We analysed the data cross-sectionally using ROC curves and mediation tests.
RESULTS: When the risk scores were ranked according to the magnitude of ROC areas for HS-CTnT prediction, a score based only on age and gender came first (ROC area=0.79), followed by Q-Risk2 (0.76), Framingham (0.70), Joint-British-Societies (0.69) and Assign (0.68). However, when the scores were ranked according to the extent of mediation by CAC (proportion of association mediated), their order was essentially reversed (age&gender=6.8%, Q-Risk2=9.7%, Framingham=16.9%, JBS=17.8%, Assign=17.7%). Therefore, the more accurate a score is in predicting detectable HS-CTnT, the less it is mediated by CAC; i.e. the more able a score is in capturing atherosclerosis the less it is able to predict cardiac damage. The P for trend was 0.009.
CONCLUSIONS: The dynamics through which cardiac cell damage is caused cannot be explained by 'classic' heart disease risk factors alone. Further research is needed to identify precursors of heart disease other than atherosclerosis.

Background

The pathophysiology of heart disease has been studied for centuries (Nabel & Braunwald, 2012) and our knowledge of the disease has undergone a remarkable evolution during the past decade (Libby, 2013). Coronary atherosclerosis, indicated as the fundamental precursor of coronary artery disease, was previously considered as a passive cholesterol storage disease whereas we currently view it as an inflammatory disorder and modern views of the dynamics underlying acute cardiac events also highlight the role of inflammation. Therefore our understanding of phenomena such as lesion formation, arterial remodelling, plaque rupture, thrombosis, and others will likely continue to evolve (Libby, 2013).

Risk assessment is an important part of routine clinical practice, and tools for the prediction of coronary artery disease (CAD) events in healthy subjects and the correlated administration of preventive cures have a long history (About the Framingham Heart Study). However, in spite of our constant progress in the comprehension of CAD pathophysiology, standard risk algorithms are based on the idea that cardiac events arise through the accumulation of coronary atherosclerosis and widely-used tools like the Framingham, the Joint British Societies & British National Formulary, the Assign, and the Q-Risk scores are still based on variables such as age, gender, blood lipids, blood pressure, and other ‘classic’ risk factors associated with atherosclerosis, as it was traditionally conceived (D'Agostino et al., 2008; British Cardiac Society et al., 2005; Woodward et al., 2007; Hippisley-Cox et al., 2008). The standard risk algorithms may not therefore be completely accurate in predicting cardiac events, rather than coronary stenosis, because they only include factors associated with ‘passive’ atherosclerosis, considered as the main mediator for CAD events.

Cardiac Troponin T (CTnT) is a plasma protein routinely tested for the diagnosis of acute myocardial infarction (AMI), since it is a marker of myocardial cell damage (Thygesen et al., 2007). In clinical settings, CTnT is measured using standard assays that have a lower detection limit of 10 ng/L (Wallace et al., 2006) and a diagnostic threshold of 35 ng/L (Thygesen et al., 2007; Wallace et al., 2006). However, high-sensitivity assays have recently been developed (HS-CTnT) with a lower detection limit of 3 ng/L (Giannitsis et al., 2010; Collinson, 2011; Collinson et al., 2012). In healthy people not fulfilling any diagnostic criterion for AMI, greater HS-CTnT is associated with greater incidence of AMI, other structural and functional heart diseases, cardiovascular mortality, and all-cause mortality, and can be therefore considered the most proximal sentinel marker of heart disease (De Lemos et al., 2010; deFilippi et al., 2010).

The mechanisms and clinical relevance of HS-CTnT in apparently low risk participants remains poorly understood, and further research is needed to understand if HS-CTnT should be included in screening tools as part of routine clinical practice. However, identifying groups of people with low risk profile, absent coronary calcification, but with detectable HS-CTnT would create a momentum towards the exploration of new pathophysiologic dynamics of heart disease.

The aim of this study was to test the ability of the Framingham, the Joint British Societies & British National Formulary (JBS/BNF), the Assign, and the Q-Risk 2 scores to identify individuals with detectable HS-CTnT plasma concentration in people with and without apparent atherosclerosis as defined by coronary artery calcification (CAC) and to assess to what extent their estimates are mediated by CAC.

Methods

Study design

Our cross-sectional study involved participants drawn from the Whitehall II epidemiological cohort (Marmot et al., 1991) between 2006 and 2008 in United Kingdom. The criteria for entry into the study included no history or objective signs of clinical or subclinical CVD, no previous diagnosis or treatment for hypertension, inflammatory diseases, allergies, or kidney disease. CVD was defined as prior myocardial infarction, stable or unstable angina, revascularization procedure, heart failure, transitory ischaemic attack, stroke, or electrocardiographic abnormalities (resting 12-lead electrocardiograms were taken). This information was confirmed by a telephone interview and verified from clinical data collected from the previous seven phases of the Whitehall II study. Volunteers were of white European origin, aged 53–76 years, and 56.5% were in full-time employment. Selection was stratified by grade of employment (current or most recent) to include higher and lower socioeconomic status participants. From the initially invited participants (n = 1169), 27.6% were not eligible (mainly because of prescribed medications) and 25.9% declined to take part. Participants were prohibited from using any medication from seven days before testing and were rescheduled if they reported colds or other infections on the day of testing. Participants gave full informed consent to participate in the study and ethical approval was obtained from the UCLH committee on the Ethics of Human Research. The study conformed to the principles of the declaration of Helsinki.

Data collection

Non-fasting blood samples were collected in EDTA tubes and centrifuged immediately at 2500 rpm for 10 min at room temperature. Plasma was removed from the tube and aliquoted into 0.5 ml portions and stored at 80 °C until analysis. We measured cardiac troponin T concentrations using a highly sensitive assay on an automated platform (Elecsys-2010 Troponin T hs STAT, Roche Diagnostics), with a lower detection limit of 3 ng/L and a reported 99th percentile value in apparently healthy individuals of 13.5 ng/L, at which the CV is 9%, confirmed by in house studies (Giannitsis et al., 2010; Collinson, 2011; Collinson et al., 2012).

The assessment of coronary artery calcification (CAC) was performed using electron beam computed tomography (GE Imatron C-150, San Francisco, CA, USA) as previously described (Anand et al., 2007). In brief, 40 contiguous 3 mm slices were obtained during a single breath-hold starting at the carina and proceeding to the level of the diaphragm. Scan time was 100 ms/slice, synchronized to 40% of the R-R interval. Agatston and volumetric calcium scores were calculated to quantify the extent of CAC by a single experienced investigator blinded to the psychophysiological and clinical data on an Aquarius workstation (TeraRecon Inc., San Mateo, CA, USA). Since calcified volume was very highly correlated with Agatston score (Spearman's rho = 0.99), we present data for Agatston score only.

Participants reported current smoking levels. We measured height and weight in light clothing for the calculation of body mass index (BMI). Fasting blood samples were taken during a separate clinical assessment. Total and high-density lipoprotein (HDL) cholesterol and triglycerides were measured within 72 h in serum stored at 4 °C using enzymatic colorimetric methods (Brunner et al., 1997). Low-density lipoprotein (LDL) cholesterol was derived using the Friedewald equation (Warnick et al., 1990). Glucose homeostasis was assessed from glycated haemoglobin (HbA1C) concentration, assayed using boronate affinity chromatography, a combination of boronate affinity and liquid chromatography.

Data analysis

Data analysis was performed using Stata v.13. We checked the dataset for missing and inconsistent values, as well as normality, outliers, and digit preference for linear variables. We excluded 83/543 people (15.3%) with prescribed statins or diabetes. Out of the remaining 460 participants, 30 (6.5%) had missing information for HS-CTnT due to insufficient blood samples, and the final analytic sample therefore comprised 430 people. HS-CTnT was highly right-skewed and for 83.3% (n = 358) of the sample it was undetectable (below the lower detection limit of 3 ng/L) and so it was transformed into a binary variable (detectable vs undetectable). We calculated the Framingham (general CVD, primary model), the JBS/BNF, the Assign, and the Q-Risk2 scores for the risk of CVD events within ten years using information about age, gender, total cholesterol, HDL, systolic blood pressure, smoking, diabetes, history of CVD (not for Assign), family history of CVD (not for Framingham and JBS/BNF), BMI (Q-Risk2 only), ethnicity (Q-Risk2 only), and rheumatoid arthritis (Q-Risk2 only) (D'Agostino et al., 2008; British Cardiac Society et al., 2005; Woodward et al., 2007; Hippisley-Cox et al., 2008). The computation of the Assign and the Q-Risk2 algorithms includes optional variables consisting of area-based socio-economic deprivation scores for UK, and we opted to use their default values (20 for Assign and zero for Q-Risk2) so that the predictions were based on clinical variables only and were therefore comparable to the ones from the other algorithms, which do not consider such variables. We also calculated a simple risk score based on age and gender only. To do this we fitted a logistic regression model with age and gender as covariates and HS-CTnT as outcome and calculated the predicted probability of HS-CTnT prevalence (Cleves, 2002).

We performed non-parametric ROC analysis to assess the accuracy of each risk score in predicting the presence of detectable HS-CTnT and compared each of these estimates with the one based on age and gender alone (reference).

We used binary mediation analysis to assess to what extent the association between each risk score and HS-CTnT was mediated by CAC (Stata FAQ). In particular, we calculated the proportion of the effect that is mediated (Stata FAQ) and its 95% confidence intervals using bootstrapping (1000 replications) (Mooney, 1993). Similarly to the ROC analysis, we used the score based on age and gender as the referent equation for the comparisons.

We performed a statistical test for trend using the following strategy: we constructed a 5 × 2 table containing the results from the ROC analysis and from the mediation analysis for each score and calculated a P for trend using linear regression.

Sensitivity analyses

We considered CAC in several different ways: linear; log-linear; binary with a cut-off at zero; binary with a cut-off at 100 (this threshold was based on the St Francis Heart Study that demonstrated maximum sensitivity and specificity for detecting cardiovascular events at a threshold calcium score ≥ 100 (Arad et al., 2005)); and ordered categorical with cut-offs at 0, 100, and 400. The score based on age and gender was calculated including an interaction term between age and gender (multiplicative model) since risk algorithms are usually gender specific. A second score based on age and gender was calculated including measures of systemic inflammation (Interleukin-6 and C-reactive protein) since they are considered as novel risk factors for CVD. Instead of fitting the ROC models using the risk scores directly, we used the same procedure as for age and gender, i.e. we first fitted a logistic regression model using the risk score as exposure and HT-CTnT as outcome, we then calculated the predicted probabilities, and then used those to fit the ROC model (Cleves, 2002). We used linear (Stata FAQ) instead of binary mediation.

Results

Out of the 430 participants in our final dataset, none had been prescribed anti-hypertensive treatment or statins or had diabetes, and on average they were at low risk for CVD events within ten years according to any algorithm used (their mean risks ranged from 10.6% [s.d. = 6.1%] calculated using the JBS/BNF equation to 14.4% [s.d. = 7.4] calculated using the Assign equation). The prevalence of detectable HS-CTnT and CAC were 17.9% and 55.6% (77 and 239 participants, respectively). Out of the 191 participants without CAC, 11.0% (n = 21, 95%CI = 6.9% to 16.3%) had detectable HS-CTnT. The sample is further described in Table 1.

Table 1

Sample description: 430 CVD-free participants drawn from the Whitehall II epidemiological cohort between 2006 and 2008 in United Kingdom.

Variable	Mean or %	SD
Age (years)	62.8	5.7
Male (%)	55.6
White ethnicity (%)	100.0
Smoker (%)	4.4
No. of cigarettes per day if smoker	15.0	8.7
Body mass index (Kg/m2)	25.7	3.8
Systolic blood pressure (mmHg)	128.5	15.3
Diastolic blood pressure (mmHg)	69.6	8.4
HDL (mmol/L)	1.7	0.5
LDL (mmol/L)	3.1	0.8
Total cholesterol (mmol/L)	5.4	0.9
Total cholesterol /HDL ratio	3.4	1.0
Glycosylated haemoglobin (mean %)	5.4	0.3
Diabetes (%)	0.0
Family history of CVD (%)	52.7
On treatment for hypertension (%)	0.0
On treatment with statins (%)	0.0
Rheumatoid arthritis (%)	0.0
Estimated risk of CVD events within 10 years (mean %)
Framingham	13.2	7.1
JBS/BNF	10.6	6.1
Assign	14.4	7.4
Qrisk2	11.4	5.5
Agatston coronary calcium score (%)
None	44.4
< 100	33.7
100–399	14.2
400 +	7.7
Detectable HS-CTnT (%)	17.9
HS-CTnT concentration if detectable (geom. ng/L)	6.0	1.6

Table 2 and Fig. 1 show the results of the ROC analysis. The score based on age and gender (reference) was a better predictor for detectable HS-CTnT compared with any other: the area under the ROC curve for age and gender was 0.79 (95%CI = 0.75 to 0.83), for Q-Risk2 score was 0.76 (95%CI = 0.72 to 0.80; P value against age and gender = 0.047), for Framingham was 0.70 (95%CI = 0.66 to 0.75; P = 0.002), for JBS/BNF was 0.69 (95%CI = 0.64 to 0.73; P = 0.001), and for Assign score was 0.68 (95%CI = 0.64 to 0.73; P = 0.003).

Table 2

Comparison of areas under the ROC curve for the prediction of detectable concentrations of HS-CTnT in 430 CVD-free participants drawn from the Whitehall II epidemiological cohort between 2006 and 2008 in United Kingdom.

Risk score	ROC	Binomial exact		P
	Area	95% CI
Age and gender	0.79	0.75	0.83	Reference
Q-Risk 2	0.76	0.72	0.80	0.047
Framingham	0.70	0.66	0.75	0.002
JBS/BNF	0.69	0.64	0.73	0.001
Assign	0.68	0.64	0.73	0.003

Fig. 1

ROC analysis for the prediction of detectable concentrations of HS-CTnT, in 430 CVD-free participants drawn from the Whitehall II epidemiological cohort between 2006 and 2008 in United Kingdom.

Table 3 shows the results of the mediation models. The 6.8% (95%CI = − 0.1% to 13.8%) of the effect of age and gender on HS-CTnT was mediated by CAC (referent model), whereas this proportion was 9.7% for Q-Risk 2 score (95%CI = 1.6% to 17.8%; P value against age and gender = 0.245), 16.9% for Framingham (95%CI = 7.5% to 26.3%; P = 0.018), 17.8% for JBS/BNF (95%CI = 7.4% to 28.2%; P = 0.019), and 17.7% for Assign score (95%CI = 6.4% to 29.0%; P = 0.029). The full output from the mediation models is presented in an on-line supplementary data file.

Table 3

Comparison of mediation models for HS-CTnT: percentage of association between risk score and HS-CTnT that is mediated by CAC, for 430 CVD-free participants drawn from the Whitehall II epidemiological cohort between 2006 and 2008 in United Kingdom.

Risk score	Percentage	Bootstrap		P
	mediated	95% CI
Age and gender	6.8	− 1.4	13.8	Reference
Q-Risk 2	9.7	1.6	17.8	0.245
Framingham	16.9	7.5	26.3	0.018
JBS/BNF	17.8	7.4	28.2	0.019
Assign	17.7	6.4	29.0	0.029

Therefore, when the risk algorithms are ranked according to the magnitude of ROC areas, the age and gender model comes first, followed by Q-Risk2, Framingham, JBS/BNF, and Assign (Table 2). However, when the scores are ranked according to the extent of mediation by CAC, their order is essentially reversed (Table 3). Consequently, the more accurate a score is in predicting detectable HS-CTnT the less it is mediated by CAC. In other words, the more able a score is in capturing atherosclerosis the less it is able to predict cardiac damage. The P for trend was 0.009.

The sensitivity analyses gave similar results compared with the main analysis.

Discussion

Our results show that a risk model based on age and gender is more accurate in predicting the presence of detectable concentrations of cardiac Troponin T using a high-sensitivity assay compared with several widely-used algorithms for the prediction of CVD events within ten years that use additional risk factors such as lipid profile and blood pressure. We have also ascertained that the predictions are mediated by coronary calcification to a greater extent for these algorithms compared with age and gender alone, and we found a trend showing that the more accurate a score is in predicting detectable HS-CTnT the less it is mediated by CAC. This suggests that mechanisms other than extent of coronary atherosclerosis may operate in determining ischaemia and/or that phenomena other that ischaemia may cause CVD. Moreover, since the risk scores typically take age and sex into account, the poorer ability of these algorithms to identify people with detectable troponin indicates that the inclusion of standard cardiovascular risk factors may reduce the accuracy of predicting the processes that lead to release of troponin T. About 11% of our participants with absent coronary calcification had detectable levels of HS-CTnT.

A strength of our study is that all participants were free from any disease (including diabetes) and symptoms, and that no participants were taking anti-hypertensive medications or statins, limiting issues of confounding, interaction, and bias. On the other hand, the fact that we have used a low risk sample may limit the applicability of our results to the general population (broader spectrum of risk profiles).

The prevalence of detectable HS-CTnT (17.9%) in our British sample was similar to levels reported (15.7%) in a large-scale nationally-representative CVD-free population sample in USA (De Lemos et al., 2010). This suggests that our selection strategy was relatively unbiased, and that the biochemical analyses in our study were comparable to those in other investigations.

A single measure of plasma HS-CTnT concentration cannot be regarded as a robust test if it is not stable over time, i.e. if it shows high intra-individual short-term variation. However, the results from the ARIC study showed that HS-CTnT intra-individual variability over 6 weeks is small, with a correlation coefficient of 0.94 (Agarwal et al., 2011).

A limitation of our study is that we did not assess the prediction of actual future CVD events in a prospective manner, but we only considered HS-CTnT cross-sectionally. However, Troponin T is a contractile protein that normally is not found in serum and it is released only when myocardial necrosis occurs and it is therefore highly sensitive and specific for cardiac damage (Hillis & Fox, 1999). As a consequence, HS-CTnT is considered to be a proximal marker of heart disease, since the Troponin T test is at the core of the diagnosis of AMI itself (Thygesen et al., 2007). Furthermore, our study design (cross-sectional on disease-free and symptoms-free people) limits the possibility of selection bias that is typical of follow-up studies. Moreover, the possibility of reverse causality (elevated levels of HS-CTnT causing increased risk scores because of elevated levels of cholesterol, high blood pressure, etc.) is not biologically plausible, especially for the score based on age and gender, which are non-modifiable parameters.

Our results could be distorted if our CAC measurements did not mark atherosclerosis accurately (information bias). However, although it has been argued that plaque calcification may be actually protective against plaque rupture, the role of calcification in the progression and as a marker of coronary atherosclerosis has been confirmed, and CAC score is considered an excellent marker of atherosclerosis (Shimizu et al., 2012).

Non-calcified coronary plaques are not detectable using cardiac computerized tomography and therefore an explanation to our results may be that the ‘classic’ risk factors may have an effect of plaque calcification, whereas plaque formation and instability may be mainly due to some other unknown factors. In fact, it has been argued that raised troponin T may be due to occult or undetected plaque rupture (Korosoglou et al., 2011) and it is known that plaque rupture is a relatively common event that is usually not followed by an acute cardiac event (Arbab-Zadeh et al., 2012).

In conclusion, our results suggest that the ‘classic’ risk factors for CVD, on which several internationally-used algorithms are based, are not able to explain all dynamics through which cardiac cell damage is caused. Further research is needed to identify other determinants of CVD that may become part of risk scores and may make CVD risk prediction more accurate.

Conflict of interest statement

All authors declare no conflict of interest of any kind.