Narrative Review: Assessment of C-Reactive Protein in Risk Prediction for Cardiovascular Disease

Key Summary Points Novel risk markers for cardiovascular disease (CVD) are often said to add independent predictive value for risk prediction, based on the finding of a significant relative risk after adjustment for traditional risk factors. However, the utility of novel risk markers for screening and risk prediction should be judged not by relative risks but by test characteristics such as sensitivity, specificity, predictive values, likelihood ratios, model calibration, c-statistics, and areas under receiver-operating characteristic curves. Inflammation has emerged as a key mediator of atherogenesis and triggering of CVD events. C-reactive protein (CRP) is a nonspecific marker of inflammation that, in healthy individuals, has been shown to be associated with future incidence of CVD. Few studies have reported test characteristics for CRP, particularly in the context of traditional risk prediction algorithms such as the Framingham risk score. In the overall adult population, CRP appears to add little to risk prediction that uses the Framingham risk score. Likewise, among subgroups of individuals predicted to be at high (>20%) or low (<10%) risk by Framingham, CRP levels contribute little further risk discrimination. Among those predicted to be at intermediate 10-year risk (10% to 20%) by Framingham, CRP levels greater than 3.0 mg/L may indicate high risk and need for more intensive preventive therapy. Many questions remain before CRP can be accepted as a standard CVD risk factor, incorporated into risk prediction algorithms, or used for universal screening. Future studies of CRP and other novel CVD risk markers should focus on test characteristics, not just relative risks, in order to better define their utility for risk prediction when added to traditional CVD risk factors. Hundreds of putative risk factors have been associated with cardiovascular disease (CVD) (1). With remarkable frequency, investigators continue to propose additional behavioral, biochemical, environmental, and genetic risk markers. Recent extensive evidence supports inflammation as a key pathogenetic mechanism in the development and progression of atherosclerosis and in triggering clinical atherothrombotic CVD events. C-reactive protein (CRP) is one possible marker of vascular inflammation, and some investigators hypothesize that CRP plays a direct role in promoting vascular inflammation, vessel damage, and clinical CVD events (2-4), although this remains controversial. Concomitantly, some experts recommend routine measurement of CRP using high-sensitivity assays as a major new population screening test for prediction of CVD risk (5, 6). Scientific (7-9) and lay (10-13) media have enthusiastically embraced, advocated, and promoted CRP for risk prediction. We believe that this enthusiasm is premature. Instead, we recommend thorough consideration of what is known about CRP's test characteristics and the effectiveness, costs, and benefits of its measurement. The literature examining CRP levels and incident cardiovascular disease among healthy individuals focuses mainly on associations between CRP and CVD, without adequately considering CRP's test characteristics and its additional utility over and above that provided by traditional risk factor measurement. In this critical review of literature published before January 2006, we use a framework for evaluating diagnostic tests in risk prediction to show what is and is not known about the role for CRP in cardiovascular risk prediction. Differences between Association and Predictive Utility Countless studies report strong, independent associations between novel risk markers and CVD. Typically, these studies find unadjusted odds ratios for disease in the range of 2.0 to 3.0, and occasionally a bit higher, for individuals with the highest compared with the lowest levels of the factor. After adjustment for age, sex, and other established CVD risk factors, the relative risks (which may be reported as rate ratios, odds ratios, or hazards ratios, depending on the study design) for the association are typically attenuated to within the range of 1.5 to 2.0. Such findings indicate that an independent association exists between the marker and disease, provided one assumes that the multivariable model adjusted for all important confounding. However, whether such an association indicates clinical utility for measuring the factor is an entirely different issue. Decisions about the predictive utility of new tests should not focus on associations and relative risks. The consequences of decision making based solely on significant associations would include screening for and treating hundreds of risk markers and risk factors. Rather, we should base decisions about the additional utility of a new test for risk prediction on the test's performance (that is, sensitivity, specificity, predictive values, and clinical likelihood ratios) in the context of existing predictors. Useful ways to examine potential additional utility in the context of existing predictors include examination of the calibration of models with the new risk factor and comparisons of the areas under receiver-operating characteristic curves (AUCs) or c-statistics for risk scores calculated without and with the novel risk marker. (See the Glossary for definitions of these and other terms relevant to test utility and prediction.) All these test characteristics must be carefully considered before it can be determined whether a novel risk marker truly adds prognostic value above that provided by existing risk prediction, whether it helps to identify and reclassify risk status appropriately, and whether it has utility across the population or only in specified subgroups. Ultimately, these characteristics will determine whether a novel prognostic screening test is useful in informing clinical decision making regarding treatment. The AUC and the c-statistic are equivalent functions of the sensitivity and specificity of a test across the spectrum of all possible cut-points defining a positive and a negative test result. They represent the ability of a test to discriminate cases from noncases. For the purposes of this review, we refer to c-statistics, although most investigators to date have published their findings as AUCs. Although reliance on the c-statistic alone as a measure of test performance has limitations (14), one can think of the c-statistic as the probability that a randomly selected person from the affected population will have a higher test score or value than a randomly selected person from the unaffected population. The c-statistic ranges from 0.50 (no discrimination beyond random chance) to 1.0 (perfect discrimination). C-statistics between 0.70 and 0.80 are considered acceptable, and those between 0.80 and 0.90 are considered excellent (15). As recently demonstrated by Pepe and colleagues (16), extremely large odds ratios are required for clinically meaningful increases in these test characteristics. For example, for a binary risk marker considered in isolation, a univariate odds ratio of 9.0 or greater would be required for excellent discrimination of cases from noncases. When the marker is considered in the context of preexisting risk factors or a risk score, multivariable (independent) odds ratios in excess of 3.0 for the marker would typically be required to increase the c-statistic (and hence improve discrimination of cases from noncases) by an additional 5% or more (16). Thus, only unusually strong, independent risk factors can markedly improve risk discrimination in the population as a whole; most new risk factors do not achieve this level of added risk above standard risk factors. It is true that 2 prediction strategies may yield the same c-statistic but have different utility. A given strategy may have a higher sensitivity at a given specificity, or vice versa. Thus, examination of the 2 receiver-operating characteristic curves and consideration of where the maximal clinical benefit lies are useful to determine the overall utility of a test in the population and for subgroups. The observation that single risk factors predict CVD risk poorly can be explained by the fact that CVD is a complex disease with multiple antecedents. Accordingly, investigators developed multivariable risk prediction algorithms to predict the absolute and relative risks for occurrence of CVD and coronary heart disease (CHD) events over the next decade (17-20). The Framingham risk score (20, 21)), which uses age, sex, total and high-density lipoprotein cholesterol levels, blood pressure, smoking, and diabetes as variables to predict risk, is widely recommended. When a model containing these traditional risk factors was applied to several diverse epidemiologic cohorts, the AUC ranged from 0.66 to 0.83 in men and from 0.72 to 0.88 in women (17). Thus, despite some contrary claims (22-24), traditional risk factors, when considered in combination, provide valuable approximations of CVD risk that are superior to approximations based on single risk factors (17,25). Given the improvement in risk discrimination provided by multivariable models, these multivariable risk estimates are the logical standard to which new risk factors must be added and compared. The National Cholesterol Education Program's Third Adult Treatment Panel (21) used a modified form of the Framingham risk score as its risk prediction tool. The panel suggested thresholds of 10-year risk on which treatment decisions regarding lipid-lowering therapy should be based, in order to match the intensity of therapy to the magnitude of absolute risk. In the panel's scheme, individuals with a predicted 10-year risk for CHD less than 10% are considered to be at low risk, those with a risk of 10% to 20% are considered to be at intermediate risk, and those with a risk greater than 20% (or with existing CHD or CHD risk equivalents) are considered to be at high risk. The focus is on identifying individuals at hi