Abstract
Measurement of C-reactive protein (CRP), an inflammatory marker associated with coronary heart disease (CHD) events, has been proposed as a means of screening for future CHD. In prospective studies about a three-fold increase in risk of CHD observed between the top fifth and bottom fifth of the CRP distribution has been taken to support the use of CRP as a screening test. This however gives an over-optimistic impression of its value, because people in the middle of the distribution, where most CHD events occur, are excluded from the analysis. A different analysis is needed to assess whether screening is worthwhile. Examination of the relative frequency distributions of CRP in individuals from 22 prospective studies of individuals without previous cardiovascular disease who subsequently did and did not have a CHD event shows that the detection rate (or sensitivity) was 18% for a false-positive rate of 10% (CRP cut-off 6.65 mg/L); a poor screening test. Whatever CRP cut-off is used, the overlap in CRP values between affected and unaffected individuals is too great for CRP to usefully discriminate between those who will and will not have a CHD event.
An examination of the relative frequency distributions of CRP in individuals with and without subsequent CHD events provides direct estimates of screening performance: the detection rate (the proportion of affected individuals with a positive CRP test result) for specified false-positive rates (proportion of unaffected individuals with a positive CRP test result). Shah et al. 1 showed that CRP is a poor screening test with a detection rate of about 16% (95% CI 15–17%) for a false-positive rate of 10%, based on a meta-analysis of 25 published studies. Thirteen of these included some individuals with and others without cardiovascular disease at baseline. We conducted a screening meta-analysis of 22 studies of individuals without prior cardiovascular disease,5–26 including 12 studies considered by Shah and colleagues and a further 10 from a Medline search of prospective studies published up to May 2009. Figure 1 shows the detection rates for a 10% false-positive rate in each of the 22 studies ordered by the mean age of participant together with the summary estimate based on all 22 studies. This showed an 18% detection rate (16–20%) for a 10% false-positive rate.
C-reactive protein (CRP) as a screening test for future coronary heart disease (CHD) events: estimates of detection rate for a 10% false-positive rate from prospective studies of individuals without CHD at baseline and a summary estimate for all studies
Plots of detection rate against false-positive rate (so-called receiver operating characterstic [ROC] curves) are often used to assess the value of a screening test. Figure 2 shows this based on the meta-analysis of 22 CRP studies. The closer the curve to the diagonal, the worse the test. The ‘area under the curve’ (or so-called ‘c’ statistic) is often used as a measure of screening performance, with values ranging from 0.5 (a useless test, where detection rate always equals false-positive rate) to 1.0 (a perfect test) – for CRP it is 0.57. This measure is not very informative, because it does not indicate directly what matters, namely the CRP cut-offs that are needed to yield detection rates for specified false-positive rates, and because the value is influenced by false-positive rates that are unlikely to be adopted in practice.
C-reactive protein (CRP) as a screening test for future coronary heart disease: plot of detection rate against false-positive rate, based on 22 prospective studies
Figure 3 shows the distributions of CRP in individuals who did and did not have a CHD event over time based on the meta-analysis of 22 studies. These give a simple visual assessment of the screening performance of a test (the smaller the overlap the better the test). From the distributions it is easy to get an accurate indication of the detection rate for different false-positive rates according to the CRP cut-off – clinically relevant information that is not given in an ROC curve. The detection rates for 5%, 10%, 15%, 20% and 30% false-positive rates were 11%, 18%, 26%, 32% and 42%, requiring CRP cut-offs of 9.7, 6.7, 4.8, 3.8 and 2.6 mg/L respectively. Whatever CRP cut-off is used, the overlap is too great for CRP to usefully discriminate between those who will and will not have a CHD event.
Relative frequency distributions of C-reactive protein (CRP) values in individuals who subsequently had a coronary heart disease event (affected) and unaffected individuals; detection rates with 3 CRP cut-offs (mg/L) shown corresponding to false-positive rates of 5%, 10% and 20%, based on 22 prospective studies
In 2003 the American Heart Association concluded that CRP may be used at the discretion of a physician as part of standard risk factor measurement to assess a person's risk of CHD and select individuals for preventive treatment, but acknowledged that the benefits of this strategy remained uncertain. 27 In spite of the uncertainty, the JUPITER trial adopted CRP as a screening test in a systematic way, using a cut-off of 2 mg/L to select men (≥50 years of age) and women (≥60) for inclusion in the trial of rosuvastatin versus placebo. 28 Rather than resolving the uncertainty over whether CRP screening is worthwhile, the trial assumed that it was. The opposite is the case. Figure 3 shows that a CRP cut-off of 2 mg/L yields a detection rate of about 50% for a false-positive rate of about 40%. Screening using CRP is little better than selecting people at random for preventive treatment.