LAUSR

Circulation. 2019;140:e822–e823. DOI: 10.1161/CIRCULATIONAHA.119.042776 e822 Allan D. Sniderman, MD George Thanassoulis, MD, MSc Michael Pencina, PhD To the Editor: In their study of the UK Biobank database, Welsh et al1 concluded that “measurement of total cholesterol and HDL-C in the nonfasted state is sufficient to capture the lipid-associated risk in CVD prediction, with no meaningful improvement from addition of apolipoproteins, direct or calculated LDL-C.” Their conclusion is principally based on the failure of these alternate measures to increase the C-statistic meaningfully. Significant increase in the C-statistic—an estimate of the area under a receiver operating characteristic curve—has been commonly accepted as the test of whether a new measure should be adopted in clinical care. If the authors’ analysis and reasoning are correct, total cholesterol and high-density lipoprotein cholesterol (HDL-C) should displace low-density lipoprotein cholesterol (LDL-C), non-HDL-C, and apoB in clinical care. However, this conclusion is inconsistent with previous work and prevention guidelines. Whereas the C-statistic is a reasonable measure to assess rank-based discrimination of prognostic models, we have demonstrated that it is not the only lens through which to determine clinical value.2 Age is by far the most important determinant of the C-statistic. By contrast, causes of cardiovascular disease (CVD) only marginally influence the C-statistic even though their treatment has major effects on outcome.2 Thus, the C-statistic declines progressively from 0.75 at age 45 to 54 years to 0.62 at age 75 to 84 years, whereas adding systolic blood pressure, non-HDL-C, diabetes mellitus, or smoking to a model with other risk factors increases it by only 0.004 to 0.013.2 Metrics that improve estimation of expected treatment benefit are better gauges of clinical utility. These include calibration of risk estimates and effect on relative risk reduction. Discordance analysis compares risk estimates of highly correlated markers.3 Interestingly, Welsh et al1 demonstrate that in the participants in the UK Biobank discordant for apoB versus LDL-C or non-HDL-C, the hazard ratio for apoB was 1.23 and highly significant, whereas the hazard ratios for LDL-C and non-HDL-C were 1 and 1.08, respectively, and not significant after adjustment for age, sex, and the other conventional risk factors (see Tables X and XI in the online-only Data Supplement of Welsh et al1). The UK Biobank results are, therefore, consistent with multiple other discordance analyses, which demonstrate that in patients discordant for apoB and LDL-C or non-HDL-C, apoB correctly predicts risk, whereas LDL-C and non-HDL-C do not.3 The effect of lipid markers on relative risk reduction is beyond the scope of the article by Welsh et al,1 but mendelian randomization studies have suggested that apoB is the most accurate measure there as well.4 Risk is, and should remain, an important element in the CVD prevention paradigm. However, measuring the causes of CVD should also be a priority in clinical care. Causes of CVD can identify younger patients at high longer-term CVD risk who are not at high short-term risk who could benefit from statin preventive © 2019 American Heart Association, Inc. LETTER TO THE EDITOR