Cholesterol and Coronary Heart Disease: Predicting Risks by Levels and Ratios

Methods

Patients

We used data from three sources: the placebo group of the Lipid Research Clinics (LRC) Coronary Primary Prevention Trial (CPPT); persons enrolled in the Framingham Heart Study who participated in the 1970-1971 biennial examination; and persons without coronary heart disease at the inception of the LRC Population Prevalence Study. Analyses were done with primary data provided by the National Heart, Lung and Blood Institute. We used a high-risk population (CPPT) and a general population (Framingham Heart Study) that monitored the full spectrum of coronary heart disease to test the discriminating ability of different measures. We used two general populations (Framingham Heart Study and LRC Prevalence Study) to evaluate the risk classification abilities of the measures.

The CPPT was a randomized, double-blind, placebo-controlled intervention trial that included 3806 men with primary hypercholesterolemia (type IIA) and without coronary heart disease, hypertension, hypertriglyceridemia (triglycerides >3.39 mmol/L [> 300 mg/dL]), or diabetes at trial inception. We restricted our analysis to the 1898 patients who received placebo and for whom there were complete data to minimize the confounding effects of the intervention on subsequent risk for coronary heart disease. The design of the CPPT and its results have been described previously [13, 14]. Data from this cohort of high-risk men were suitable for evaluating the ability of various cholesterol measures to discriminate among men according to their levels of risk for coronary heart disease, but these data could not be used for assessing the performance of those measures to classify risk for intervention in a more general population. Cholesterol levels (including LDL cholesterol levels) were directly measured by a central laboratory [13].

One of the general patient populations used in our analysis was made up of men and women in the Framingham Heart Study, a longitudinal observational study with biennial follow-up of 5209 men and women, 28 to 62 years of age at study inception. High-density lipoprotein and LDL cholesterol measurements were available only during the 1970-1971 examination, therefore, we studied 1025 men and 1442 women without coronary heart disease at that time who were all older than 50 years of age. The cohort study and methods for cholesterol measurement have been described previously [15]. Because the Framingham Study is a community-based inception cohort with excellent outcome assessment for the full spectrum of coronary heart disease events, data from the cohort are useful for evaluating the discriminating ability of various cholesterol measures as well as the performance of those measures in classifying risk for intervention in a population.

The LRC Prevalence Study, conducted between 1972 and 1976, used a previously described two-stage screening procedure to assemble a mixed cohort with and without existing coronary heart disease [16-18]. Our second general patient population was the subsample of this cohort who were screened in the second stage (1767 women, 1911 men), were free of existing coronary heart disease at the time of assessment, and were available for follow-up. The risk-factor distribution for coronary heart disease in this sample was similar to that of persons in the general population who were 35 to 72 years of age. Low-density lipoprotein cholesterol was directly measured for all participants. Although this subsample is representative of the general population of persons who do not have coronary heart disease, restricting the outcome to death from coronary heart disease meant the LRC Prevalence Study had inadequate power to statistically test the discriminating ability of alternative cholesterol measures, but could be used to classify risk for intervention.

Outcomes

The outcomes we used were coronary heart disease observed during 8 years of follow-up in the CPPT and Framingham studies and death from coronary heart disease observed during 10 years of follow-up in the LRC Prevalence Study, diagnosed using previously described assessment methods [13-16]. Coronary heart disease events in the CPPT included primary myocardial infarction, angina requiring hospitalization, Rose-questionnaire-defined angina, positive findings on an exercise tolerance test, coronary revascularization, and resuscitated coronary collapse. Deaths were classified as either coronary heart disease or non-coronary heart disease, with coronary heart disease-related deaths subclassified as sudden or nonsudden. Events in the Framingham Heart Study included myocardial infarction, angina, coronary insufficiency, and sudden and nonsudden death from coronary heart disease. In the LRC Prevalence Study, events were restricted to mortality follow-up, with vital status reported for more than 99% of the cohort.

Analysis Overview

We tested the hypothesis that the total cholesterol/HDL ratio is a superior measure of risk for coronary heart disease when compared with the total cholesterol level or LDL cholesterol level and that the total cholesterol/HDL ratio is an equal or better measure when compared with the LDL/HDL ratio, using grouped (stratified) and logistic regression (parametric) analyses.

To compare measures used for risk discrimination, we defined a superior measure as one that identifies low- and high-risk persons within risk groups identified by an inferior measure of risk for coronary heart disease. The corollary is that an inferior discriminating measure is unable to identify low- and high-risk persons within risk groups identified by a superior measure.

To compare measures used for risk classification, we defined a superior cholesterol measure as one that identifies a larger proportion of a population at the same or greater risk than the current classification based on LDL cholesterol levels or as one that identifies the same proportion of the population at greater risk.

Grouped (Stratified) Analysis

We evaluated four pairs of cholesterol variables in the CPPT and the Framingham populations: total cholesterol compared with the total cholesterol/HDL ratio, LDL cholesterol compared with the LDL/HDL ratio, LDL cholesterol compared with the total cholesterol/HDL ratio, and the LDL/HDL ratio compared with the total cholesterol/HDL ratio.

We developed two sets of three-way contingency tables for each pair of cholesterol measures. Thus, we divided the population into deciles using the first measure in the pair (for example, total cholesterol). We then divided the population in each decile into tertiles using the second measure in the pair (for example, within each total cholesterol decile, we divided observations according to the total cholesterol/HDL ratio). Within each tertile, patients were separated into those who had a coronary heart disease event and those who did not.

We tested whether the total cholesterol/HDL ratio was able to identify different risks that were statistically significant within each of the deciles of total cholesterol or LDL cholesterol and whether any other measure identified differences in risk within deciles of the total cholesterol/HDL ratio, using a two-tailed P value of 0.05 for the Cochran-Mantel-Haenszel statistic [19]. Details of the grouped (stratified) analyses are provided in the Appendix.

Logistic Regression Analyses

We used three sets of 12 logistic regressions to evaluate the unique explanatory information about risk for coronary heart disease contained in the four measures of cholesterol. The same populations from the grouped (stratified) analysis were used.

First, for each pair, we chose one of the measures as our primary explanatory variable (all information shared between the two measures would be captured by the primary variable, for example, total cholesterol). Second, we developed a measure of the additional information provided by the second cholesterol variable, termed the residual (for example, information in the total cholesterol/HDL ratio that is not shared with total cholesterol), using ordinary least-squares regression analysis. Last, we used logistic regression to predict risk for coronary heart disease as a function of the primary explanatory variable and of the remaining, unique information provided by the second variable (residual). Details of the two-step regression approach are provided in the Appendix.

We tested whether the unique information provided by the total cholesterol/HDL ratio (total cholesterol/HDL residuals) was a statistically significant predictor of risk for coronary heart disease when added to the total cholesterol level, LDL cholesterol level, or LDL/HDL ratio used as the primary variable, whereas the unique information provided by total cholesterol level, LDL cholesterol level, or LDL/HDL ratio were not significant predictors of risk for coronary heart disease, when the total cholesterol/HDL ratio was the primary variable. A two-tailed P value of 0.05 for the residuals was used to indicate whether or not they added statistically significant information to the explanatory power of the primary variables.

Risk Classification for Intervention

To determine if using the total cholesterol/HDL ratio would have an effect on the risk classification of a population, we identified the proportion of men and women older than 45 years of age in the LRC Prevalence and the Framingham Heart studies who were classified as high risk using either the first set of guidelines from the National Cholesterol Education Program for initiating pharmacologic treatment regardless of other risk factors for coronary heart disease (LDL >4.91 mmol/L [>190 mg/dL] after diet treatment) or revised guidelines for persons older than 45 years, incorporating HDL cholesterol (LDL ≥ 4.91 mmol/L [≥ 190 mg/dL] or LDL ≥ 4.14 mmol/L (≥ 160 mg/dL) with other coronary heart disease risk factors including an HDL ≤ 0.91 mmol/L [≤ 35 mg/dL] after diet treatment) [2]. For those persons identified as high risk, we calculated the 10-year risk for death from coronary heart disease (LRC Prevalence Study) and the 8-year risk for coronary heart disease (the Framingham Study).

We also identified proportions of the population classified as high risk using several total cholesterol/HDL ratio thresholds. These thresholds, empirically derived based on the above definition of a superior measure for risk classification, ranged from 6.0 to 7.0 in men and from 5.6 to 6.0 in women. We estimated the risk for coronary heart disease events in these high-risk groups and evaluated whether the total cholesterol/HDL ratio satisfied the criteria for a superior measure for risk classification for either coronary heart disease or death from coronary heart disease.

Individual Risk Classification

We report average 8-year risks for coronary heart disease and likelihood ratios for five total cholesterol/HDL ratio strata derived from men and women in the Framingham Study (ratios <3, ≥ 3 and <5, ≥ 5 and <7, ≥ 7 and <9, and ≥ 9). Likelihood ratios were calculated as the proportion of persons developing coronary heart disease who had ratio values in the stratum divided by the proportion of those without coronary heart disease who had ratios in the stratum. Likelihood ratios can be combined with pretest probabilities of disease to develop an estimate of the post-test probability of disease.

Results

Grouped (Stratified) Analysis

Table 1 presents the results of our analysis when we divide men in the CPPT Study and men and women in the Framingham Study into deciles using a primary cholesterol measure and then divide the deciles into tertiles using a secondary cholesterol measure. It reports the absolute difference in risk between the highest and lowest tertiles averaged across the deciles in instances where the P value for the Cochran-Mantel-Haenszel statistic for the difference was ≤ 0.15.

When total cholesterol and LDL were used as the primary measures, the total cholesterol/HDL and LDL/HDL ratios identified statistically significant high- and low-risk groups within the deciles. For example, when CPPT men were divided into deciles based on LDL cholesterol (row 3, column 1), the absolute risk difference within the LDL deciles between those with the highest and lowest thirds of total cholesterol/HDL ratios was 13.0% (95% CI, 8.2% to 17.7%). Among men in the Framingham Study (row 3, column 2), the risk difference between the highest and lowest thirds of total cholesterol/HDL ratios within the LDL deciles was 4.3% (CI, −1.3% to 9.9%); among women in the Framingham Study (row 3, column 3), the risk difference was 9.3% (CI, 5.6% to 13.1%).

Little difference was noted between risk stratification using the total cholesterol/HDL and LDL/HDL ratios except among women in the Framingham Study. When the LDL/HDL ratio was used as the primary measure in this population, the total cholesterol/HDL ratio identified high- and low-risk groups among the LDL/HDL ratio deciles (risk difference, 7.9%; CI, 4.1% to 11.7%). When the total cholesterol/HDL and LDL/HDL ratios were used as the primary measures, neither total nor LDL cholesterol identified high- and low-risk groups among the deciles.

Logistic Regression Analyses

These analyses confirmed results from the grouped (stratified) analysis. In all logistic regressions that used total cholesterol or LDL cholesterol levels as the primary explanatory variable along with residuals of total cholesterol/HDL ratios or LDL/HDL ratios (which captured the residual information that was not shared with total or LDL cholesterol), the residual was statistically significant (Table 2). When the total cholesterol/HDL ratio was used as the primary variable in the regression and residuals for total cholesterol level, LDL cholesterol level, and LDL/HDL ratios were each added to the regressions, only in one instance (total cholesterol in the CPPT Study) did the residual tend toward adding statistically significant explanatory information to the predictive ability of the total cholesterol/HDL cholesterol ratio. When the LDL/HDL ratio was used as the primary measure, only in the CPPT did the total cholesterol/HDL ratio add substantial information.

Population Risk Classification for Intervention

We applied the current LDL threshold and HDL-modified LDL threshold [1, 2] as well as several total cholesterol/HDL ratio thresholds to two populations: the Framingham Study (measuring 8-year risk for coronary heart disease) and the LRC Prevalence Study (measuring 10-year risk for coronary heart disease death) (Table 3).

In the Framingham population, an LDL threshold of 4.91 mmol/L or more (≥ 190 mg/dL) identified groups of men (10.2%) and women (18.1%) whose risks for coronary heart disease were increased only slightly more than the average risk in the population (16.4% compared with 14.7% for men, 10.3% compared with 9.2% for women). The HDL-modified LDL threshold identified a larger proportion of the population (12.4% for men, 18.8% for women) who had lower risks for coronary heart disease (15% for men, 9.9% for women).

In contrast, in the LRC Prevalence Study Table 3, the LDL threshold identified a group at substantially greater risk for death from coronary heart disease than in the general population (7.9% compared with 2.9% for men, 4.7% compared with 1.3% for women). The HDL-modified LDL threshold identified a larger group at higher risk (10% risk for men, and 5.3% risk for women).

In the Framingham and the LRC Prevalence Studies, a total cholesterol/HDL ratio threshold of 6.0 or more identified substantially more men as high risk than were identified by using either an LDL threshold or HDL-modified LDL threshold. This ratio met the criterion for a superior measure in the Framingham Study, identifying a larger proportion of the population (26.5% compared with 10.2%) at similar or greater risk (16.9% compared with 16.4%). However, participants in the LRC Prevalence Study, identified as high risk using a total cholesterol/HDL ratio threshold of 6.0, were at lower risk for death from coronary heart disease. A total cholesterol/HDL ratio threshold of 6.4 or more satisfied the criteria for a superior measure for men in the Framingham Study and men in the LRC Prevalence Study using the LDL threshold. However, no ratio threshold satisfied these criteria for men in the LRC Prevalence Study, using the HDL-modified LDL threshold.

In the Framingham and the LRC Prevalence studies, at total cholesterol/HDL ratios of 6.0 or more, a smaller proportion of women were identified as high risk than were identified using either LDL threshold, but these women were at substantially higher risk for coronary heart disease or death from coronary heart disease. Decreasing the total cholesterol/HDL ratio threshold for women to 5.6 identified similar or larger proportions of the population at greater risk for coronary heart disease and death from coronary heart disease when compared with either LDL threshold.

Individual Risk Classification

Among men in the Framingham Study who were grouped (stratified) by total cholesterol/HDL ratios ranging from less than 3.0 to 9.0 or more, a general increase was noted in 8-year risk for coronary heart disease and in likelihood ratios for coronary heart disease (Table 4). Among women, the stratum-specific risks and likelihood ratios always increased. Stratum-specific risks for men ranged from a low of 5.3% (likelihood ratio, 0.32) for total cholesterol/HDL ratios less than 3.0 to a high of 35% (likelihood ratio, 3.12) for total cholesterol/HDL ratios of 9.0 or more. Among women, risks ranged between 5.6% (ratio, 0.59) and 23.1% (ratio, 2.98).

Discussion

Controversy exists about the best method for identifying those persons at increased risk for coronary heart disease. Although LDL and HDL cholesterol levels have been recommended for determining risk [1-3], revised guidelines from the National Cholesterol Education Program recommend that the LDL cholesterol level form the basis for risk stratification and that the HDL cholesterol level and other coronary heart disease risk factors (such as smoking status) should be used to refine risk estimates [2].

Our results suggest that the total cholesterol/HDL ratio is better for discriminating risk than is the LDL cholesterol level. Using stratified analyses, neither total nor LDL cholesterol levels conveyed additional information when ratios were used to predict coronary heart disease risk in a general and a high-risk population. However, ratios had additional predictive value when total cholesterol and LDL levels were used to predict risk. The results from the logistic regression analysis confirmed the stratified analyses: The total cholesterol/HDL ratio added discriminating ability to total and LDL cholesterol measures, although no measure added discriminating ability to the total cholesterol/HDL ratio.

The threshold value for LDL cholesterol (>4.91 mmol/L [>190 mg/dL]) used by the National Cholesterol Education Program was compared with revised guidelines (decreasing the LDL threshold to ≥ 4.14 mmol/L [≥ 160 mg/dL] if the HDL were ≤ 0.91 mmol/L [≤ 35 mg/dL]) and thresholds for total cholesterol/HDL ratios that varied from 6.0 to 7.0 for men and from 5.6 to 6.0 for women. We found that for men in Framingham and women in both studies ratio thresholds identified groups with a 2% to 5% greater absolute risk for coronary heart disease or death from coronary heart disease. These absolute risks are 20% to 50% greater than comparable LDL-threshold-defined risks.

In all populations other than men in the LRC Prevalence Study, the total cholesterol/HDL ratio identified a similar-sized or larger group at the same or greater risk for subsequent coronary heart disease events than did either LDL or HDL-modified LDL thresholds; thus, the ratio was a superior cholesterol measure for risk stratification. The improved performance of the HDL-modified LDL threshold among men in the LRC Prevalence Study may be related to the limited coronary heart disease outcome that was reported (death from coronary heart disease), but this performance needs to be studied in other populations [10].

The analyses provide less support for choosing the total cholesterol/HDL ratio instead of the LDL/HDL ratio as a discriminator of risk for coronary heart disease. However, results among women in the stratified analyses tend to support using the total cholesterol/HDL ratio, which is consistent with several recent analyses suggesting that the triglyceride level (included in the total cholesterol/HDL ratio but not in the LDL/HDL ratio) is an independent predictor of coronary heart disease and death from coronary heart disease among women [20-22]. Further, total cholesterol and HDL cholesterol can be measured in the nonfasting state, with reduced analytic variability, decreasing the cost of screening when compared with an LDL criterion [23, 24].

Several clinical lessons can be derived from our work. For example, among persons with similar ratios, those with total cholesterol or LDL levels classified as “high risk” (using guidelines from the National Cholesterol Education Program) have risks for coronary heart disease that are similar to risks of those persons with total cholesterol or LDL levels classified as “low risk” or “desirable” (as defined by the guidelines). Thus, a justification may not exist for providing either the low- or high-risk group with different treatments. In addition, meaningful changes in risk predictions (from 60% to 200% of the average risk for coronary heart disease) can be made for individual patients based on total cholesterol/HDL likelihood ratios that are stratum specific. For 2% of men with the highest ratios (consistent with HDL levels from 0.78 to 0.91 mmol/L [30 to 35 mg/dL] and total cholesterol levels >7.24 mmol/L [>280 mg/dL]), one in three men will likely have a coronary heart disease event during the subsequent 8 years.

We intentionally considered cholesterol levels alone in our analysis in order to address specific controversies surrounding claims that the total cholesterol/HDL ratio should be the preferred means of risk stratification [5-10]. Additionally, our results do not address whether changes in the total cholesterol/HDL ratio are better predictors of change in risk for coronary heart disease than are changes in total cholesterol or LDL levels [2]. Answering this question requires analysis of clinical trial data [25]. However, because guidelines from the National Cholesterol Education Program use post-diet risk classification to decide on subsequent pharmacologic interventions, our results should prompt serious scrutiny of interventions that leave the total cholesterol/HDL ratio unchanged by lowering LDL and HDL cholesterol levels [26, 27].

In two general populations as well as in a population of high-risk men, the total cholesterol/HDL ratio was superior to either the total cholesterol level or the LDL cholesterol level for identifying people at greater risk for developing subsequent coronary heart disease events. Current practice guidelines could be more efficient if risk stratification were based on the total cholesterol/HDL ratio rather than primarily on the LDL cholesterol level.

Appendix

Grouped (Stratified) Analysis

For each of the four pairs of cholesterol measures (for example, total cholesterol compared with the total cholesterol/HDL ratio), we started with one of the measures (for example, total cholesterol), and we rank-ordered observations from the lowest cholesterol value to the highest. We then divided the sample into deciles, such that the first decile contained the 10% of observations with lowest cholesterol values and the tenth decile contained the 10% with highest cholesterol values.

Within each decile of cholesterol values, we rank-ordered the observations from lowest to highest based on the second cholesterol measure (for example, within the total cholesterol strata, we rank-ordered the observations according to the total cholesterol/HDL ratio). We then divided the sample within each decile into thirds, or tertiles, such that the first tertile contained the third of observations with the lowest total cholesterol/HDL ratios and the third tertile contained the third with the highest ratios. Within each tertile, patients either had a coronary heart disease event or they did not. Thus, we constructed ten 3 × 2 tables (10 strata; three tertiles within each stratum, event or no event).

The Cochran-Mantel-Haenszel statistical test determines whether tertiles of the second cholesterol measure discriminate risk across the 10 strata of the first cholesterol measure. After completing the evaluation for one of two measures of cholesterol (for example, the total cholesterol/HDL ratio within the total cholesterol strata), we used the second measure (for example, total cholesterol/HDL ratio) to develop the 10 strata and used the first measure (for example, total cholesterol) to determine whether the first measure was able to discriminate risk within the strata of the second by using the same evaluation tests.

Logistic Regression Analysis

We used three sets of 12 logistic regressions to evaluate the unique explanatory information about risk for coronary heart disease contained in four measures of cholesterol values. First, for each pair (for example, total cholesterol level compared with total cholesterol/HDL ratio), we chose one of the measures as our primary explanatory variable (that is, all information shared between the two variables would be captured by the primary variable). Second, we developed a measure of the unique information in the second cholesterol variable (for example, information in the total cholesterol/HDL ratio that is not shared, or residual, with the total cholesterol value) by regressing the primary cholesterol variable on the second (that is, making the primary cholesterol variable the independent predictor of the second cholesterol variable) and by computing a set of residuals (for example, the actual total cholesterol/HDL ratio minus the predicted total cholesterol/HDL ratio). By definition, residuals have zero correlation (are orthogonal) with the primary cholesterol measure. Last, we ran a logistic regression that predicted risk for coronary heart disease as a function of the primary cholesterol measure and the residual secondary measure. The P value on the residual indicates whether or not it added statistical significance to the explanatory power of the primary cholesterol measure.

After completing the evaluation for one of the two measures of cholesterol (for example, total cholesterol level and the total cholesterol/HDL residual), we repeated the analysis using the second cholesterol measure as the primary explanatory variable and using the residual of the first cholesterol measure (for example, the total cholesterol/HDL ratio and the residual for total cholesterol).