Background
The risk of CVD can be assessed using tools such as QRisk2,52 ASSIGN54 and Framingham 1991,53 all of which include the TC/HDL cholesterol ratio. As noted previously (see Chapters 1 and 5), the latest NICE guidance on lipid monitoring for primary prevention populations2 recommends using such risk scores to estimate 10-year risk of developing CVD to identify patients at ‘high risk’ and hence guide treatment decisions, such as the prescription of statins. The 2010 high-risk threshold was a 10-year risk of ≥ 20%, with the choice of risk calculator left to the GP. Updated guidance in 2014 recommended lowering the treatment threshold to 10% and the use of QRisk252 to calculate risk.56 The methods in this chapter therefore focus on 10-year CVD risk calculation using QRisk2,52 and include both 10% and 20% risk thresholds.
Although findings in Chapter 2 suggest that there is unlikely to be much improvement in risk calculators if other common lipids (TC, HDL, LDL cholesterol as single measurements or ratios) were used in place of TC/HDL cholesterol, there was some indication that alternative measures such as non-HDL cholesterol or Apo B may be more predictive in some population subgroups. If risk scores were recalculated to use different lipid measurements, such as non-HDL cholesterol or Apo B, some patients, currently classified as low risk, would be reclassified as high CVD risk, and vice versa; we examine the effect that modifying QRisk252 in this way would have on the proportions of patients in different risk categories, and whether or not reclassified subjects are predominantly those close to the risk threshold. As Apo B is not widely available in general practice, analyses presented in this chapter therefore focus on modification of risk scores to use non-HDL cholesterol, which can be calculated from TC and HDL measurements and does not need fasting blood samples.
There is also potential to refine cardiovascular risk estimation within a lipid monitoring programme by taking the average of lipid measurements on two or more separate occasions; as shown in Chapter 4, the high level of within-person variability means that using the average of a number of measurements, taken over a short time period, might better represent the true underlying lipid level for that patient at that time point. This would reduce the within-measurement variability (see Table 23, column 4) of the resulting estimate. Regression dilution theory shows that coefficients in a generalised linear model, such as a risk score, are affected by within-person variability.212 Specifically, they are attenuated (closer to 0) with increased variability. It follows that if the variability of the measure is to be reduced, for example by taking an average, the coefficient in the risk score must be increased.213 Here we use the results of Chapter 2 to estimate the modifications necessary to use CVD risk calculators with the mean of three and five lipid measurements.
Methods
Open source C code to calculate QRisk252 is available online,214 and version QRisk2–2012 (Q68_QRisk2_2012, created on 3 January 2012) was downloaded. QRisk252 is updated approximately annually; we did not update to the QRisk2–2013, as the latest date in CPRD data available in this project was 21 September 2012. The C code was compiled for use with both Windows and Linux machines using Microsoft Visual C++ 2010 Express (Microsoft Corporation, Redmond, WA, USA) and the GNU Compiler Collection (GCC) compiler, respectively. A plug-in was then created using Stata, version 12.1, to calculate QRisk252 directly from the C code.215
Alternative lipid measures: non-HDL cholesterol
Risk estimates calculated using a single measurement of TC/HDL cholesterol in the original QRisk252 equation were compared with those calculated using a modified risk equation utilising a single measurement of non-HDL cholesterol. The modified equation was created by substituting – for the centring constant and beta coefficients (log-HRs) for TC/HDL cholesterol in the original C code – values relevant to non-HDL cholesterol, derived as follows.
For each gender, new centring values (means) were taken from the CPRD primary prevention subjects not taking statins, having excluded individuals with measurements outside the allowed ranges for the QRisk252 variables.214 New beta coefficients were taken from the EPIC-Norfolk study,216 as this was the largest study in our systematic review (see Chapter 2) to investigate the prognostic significance of non-HDL cholesterol in a UK primary prevention population who were not taking statins. This study reports HRs per 1 SD increase in non-HDL cholesterol separately for men and women, and for both genders combined, and reported the SD for non-HDL cholesterol by gender and the occurrence of a CVD event. The SDs across the CVD and non-CVD event groups for men and women can be combined to obtain the overall SD of non-HDL cholesterol for men and women, as shown in Appendix 21.
A sensitivity analysis was carried out using the overall HR per SD increase in non-HDL cholesterol for the primary prevention group not taking statins calculated in the systematic review (see Chapter 2). To convert to a HR per millimole/litre increase, we raised the overall HR per SD to the power of 1/SD, with the SD calculated from the primary prevention off-statin group in the CPRD data (having excluded individuals with measurements outside the allowed ranges for the QRisk252 variables and different lipid measures). Centring values were also taken from this data set, as in the primary analysis.
Predicted risk using non-HDL cholesterol in the modified score was compared with predicted risk using TC/HDL cholesterol ratio in the original QRisk252 score, in primary prevention patients in CPRD. Results are presented as scatterplots and Bland–Altman plots.217 We also grouped CVD risk into 10-year risk categories in 5% increments from 0% to ≥ 25%, and used tables to compare the numbers of subjects reclassified.
Repeated lipid measures
We estimate the regression dilution ratios for a single measure of any lipid fraction as:
where and are the within-measurement and between-measurement variability. The values of and for TC and the TC/HDL cholesterol have been estimated in Chapter 4.
The regression dilution ratio for the mean of k measures is:
and hence the coefficient of a single measurement can be converted to the coefficient of the mean by multiplying by:
as has been reported previously.213 The regression dilution ratio method is exact for linear models and has been shown to apply approximately to other generalised linear models.212 It can be shown that it applies approximately in the case of univariate survival models with censoring.218 We conducted further simulations to confirm the approximation also holds for multivariate (multiple covariate) survival models (details available on request).
For consistency with the simulation models examining alternative lipid-monitoring strategies, presented in Chapter 5, we initially compare risk estimates from the original QRisk252 equation (that uses a single measurement of TC/HDL cholesterol) with estimates calculated using the mean of three and the mean of five TC/HDL cholesterol measurements in the original QRisk252 equation. We then compare estimates from the original QRisk252 equation, using one measurement of TC/HDL cholesterol with the equations modified to adjust for regression dilution using the mean of three and the mean of five repeated TC/HDL cholesterol measurements. Analyses were carried out using Stata 12.1 and R 3.0.1 (The R Foundation for Statistical Computing, Vienna, Austria).219
Results
Alternative lipid measure: non-HDL cholesterol
presents frequency counts of subjects in each of six categories of risk estimated using the original QRisk252 score, calculated using a single measure of TC/HDL cholesterol, compared with risk estimated using QRisk252 modified for use with a single measurement of non-HDL cholesterol, using estimates from EPIC-Norfolk.216 A large proportion of patients have very high risk (> 25%) according to the original QRisk252 score. These high-risk individuals are primarily aged ≥ 70 years and have at least one of the following conditions: atrial fibrillation, diabetes, treated hypertension. and show scatterplots comparing the two QRisk252 scores, with the number of subjects reclassified at 20% and 10% CVD risk thresholds, respectively. Approximately 5% of subjects are reclassified from low to high risk, and 5% from high to low risk using a threshold of 20%; smaller proportions are reclassified when the threshold is lowered to 10%. shows a Bland–Altman plot comparing the QRisk252 estimates using TC/HDL cholesterol and non-HDL cholesterol; absolute differences in the risk estimates are less than approximately 10% for 95% of subjects. Absolute differences of > 10% were generally observed in people over the 20% risk threshold. Of the 1757 people with risk differences outside the normal range, 80.9% had risk estimates of > 20% according to both the original and modified equations. Similar results were obtained when the systematic review results from Chapter 2 were used to modify the risk equation ( and –).
Frequency counts of individuals by risk estimate category: comparing QRisk2 based on single TC/HDL cholesterol measure with QRisk2 based on single non-HDL cholesterol measure, modified using estimates from EPIC-Norfolk analysis
Comparison of estimates of 10-year cardiovascular risk calculated using TC/HDL cholesterol in the original QRisk2 equation or using non-HDL cholesterol in a modified equation, with the low-/high-risk threshold set at 20%.
Comparison of estimates of 10-year cardiovascular risk calculated using TC/HDL cholesterol in the original QRisk2 equation or using non-HDL cholesterol in a modified equation, with the low-/high-risk threshold set at 10%.
Bland–Altman plot of estimates of 10-year cardiovascular risk calculated using TC/HDL cholesterol in the original QRisk2 equation or using non-HDL cholesterol in a modified equation.
Frequency counts of individuals by risk estimate category: comparing QRisk2 based on single TC/HDL cholesterol measure with QRisk2 based on single non-HDL cholesterol measure, modified using estimates calculated in Chapter 2
Comparison of estimates of 10-year cardiovascular risk calculated using TC/HDL cholesterol in the original QRisk2 equation or using non-HDL cholesterol in a modified equation (sensitivity analysis), with the low-/high-risk threshold set at 20%.
Bland–Altman plot of estimates of 10-year cardiovascular risk calculated using TC/HDL cholesterol in the original QRisk2 equation or using non-HDL cholesterol in a modified equation (sensitivity analysis).
Comparison of estimates of 10-year cardiovascular risk calculated using TC/HDL cholesterol in the original QRisk2 equation or using non-HDL cholesterol in a modified equation (sensitivity analysis), with the low-/high-risk threshold set at 10%.
Repeated lipid measures
Before adjusting for regression dilution
and and show the numbers of subjects reclassified between risk categories when CVD risk is calculated from one measurement or the mean of three measurements using the original QRisk252 equation, with no adjustment for regression dilution. Only a very small number of subjects are reclassified between risk categories, and for 95% of subjects any absolute change in calculated risk is < 3% (). Similar results were seen when CVD risk estimated using one measurement was compared with that calculated using the mean of five measurements in the original QRisk252 equation without adjusting for regression dilution; corresponding estimates are shown in and –.
Frequency counts of individuals by risk estimate category: comparing QRisk2 based on single TC/HDL cholesterol measure with QRisk2 based on mean of three TC/HDL cholesterol measures, without adjusting for regression dilution
Comparison of estimates of 10-year cardiovascular risk calculated using a single TC/HDL cholesterol measure or the mean of three TC/HDL cholesterol measures in the original QRisk2 equation, with the low-/high-risk threshold set at 20%.
Comparison of estimates of 10-year cardiovascular risk calculated using a single TC/HDL cholesterol measure or the mean of three TC/HDL cholesterol measures in the original QRisk2 equation, with the low-/high-risk threshold set at 10%.
Bland–Altman plot of estimates of 10-year cardiovascular risk calculated using a single TC/HDL cholesterol measure or the mean of three TC/HDL cholesterol measures in the original QRisk2 equation.
Frequency counts of individuals by risk estimate category: comparing QRisk2 based on single TC/HDL cholesterol measure with QRisk2 based on mean of five TC/HDL cholesterol measures, without adjusting for regression dilution
Comparison of estimates of 10-year cardiovascular risk calculated using a single TC/HDL cholesterol measure or the mean of five TC/HDL cholesterol measures in the original QRisk2 equation, with the low-/high-risk threshold set at 20%.
Bland–Altman plot of estimates of 10-year cardiovascular risk calculated using a single TC/HDL cholesterol measure or the mean of five TC/HDL cholesterol measures in the original QRisk2 equation.
Comparison of estimates of 10-year cardiovascular risk calculated using a single TC/HDL cholesterol measure or the mean of five TC/HDL cholesterol measures in the original QRisk2 equation, with the low-/high-risk threshold set at 10%.
After adjusting for regression dilution
The estimates of parameters σw and σa in Table 23 give a regression dilution ratio of 1.281 for the TC/HDL cholesterol ratio in men. Coefficients for TC/HDL cholesterol should be therefore multiplied by 1.281/(1 + 0.281/3) = 1.171 for the mean of three measurements, and by 1.213 for the mean of five measurements. We therefore increased the log-HR in the QRisk252 equation from 0.17 to 0.199 (corresponding to HR 1.22) for three measurements and to 0.206 (HR 1.23) for five measurements. In women, the regression dilution ratio from Table 24 is 1.213 and the QRisk252 log-HR is 0.16, giving rise to log-HRs (and HRs) of 0.181 (1.20) for the mean of three measurements and 0.186 (1.20 to two decimal places) for the mean of five measurements.
and –, and and –, show the effect of these changes on risk classification. Differences in the estimated risk were small, and few subjects were reclassified between risk categories.
Frequency counts of individuals by risk estimate category: comparing QRisk2 based on single TC/HDL cholesterol measure with QRisk2 based on mean of three TC/HDL cholesterol measures, adjusted for regression dilution
Comparison of estimates of 10-year cardiovascular risk calculated using a single TC/HDL cholesterol measure in the original QRisk2 equation or the mean of three TC/HDL cholesterol measures in a modified equation, with the low-/high-risk threshold set (more...)
Bland–Altman plot of estimates of 10-year cardiovascular risk calculated using a single TC/HDL cholesterol measure in the original QRisk2 equation or the mean of three TC/HDL cholesterol measures in a modified equation.
Frequency counts of individuals by risk estimate category: Comparing QRisk2 based on single TC/HDL cholesterol measure with QRisk2 based on mean of five TC/HDL cholesterol measures, adjusted for regression dilution
Comparison of estimates of 10-year cardiovascular risk calculated using a single TC/HDL cholesterol measure in the original QRisk2 equation or the mean of five TC/HDL cholesterol measures in a modified equation, with the low-/high-risk threshold set at (more...)
Bland–Altman plot of estimates of 10-year cardiovascular risk calculated using a single TC/HDL cholesterol measure in the original QRisk2 equation or the mean of five TC/HDL cholesterol measures in a modified equation.
Comparison of estimates of 10-year cardiovascular risk calculated using a single TC/HDL cholesterol measure in the original QRisk2 equation or the mean of three TC/HDL cholesterol measures in a modified equation, with the low-/high-risk threshold set (more...)
Comparison of estimates of 10-year cardiovascular risk calculated using a single TC/HDL cholesterol measure in the original QRisk2 equation or the mean of five TC/HDL cholesterol measures in a modified equation, with the low-/high-risk threshold set at (more...)
Discussion
We have considered two possible strategies for improving the lipid measurement within a single cardiovascular risk estimate. We first considered whether or not risk estimates could be improved by an alternative lipid measurement. Motivated by Chapter 2, we posited a risk equation based on non-HDL cholesterol, and found potential for about 1 person in 10 to be classified differently (high vs. low risk and vice versa) compared with a risk calculator based on TC/HDL cholesterol ratio. These results are only illustrative but at least demonstrate the potential for a change. This contrasts with our second analysis, in which we considered whether or not risk estimates could be improved by taking multiple blood samples and using the average of a lipid measure. This appears to make negligible difference to risk estimation, probably because the within-measurement variability of TC/HDL cholesterol, although not negligible, is not large compared with other sources of variation.
We did not have primary data on which to build a risk equation with non-HDL cholesterol. A further, and major, limitation of these analyses is that we can estimate the degree of reclassification but not the extent to which our modifications represent an improvement to risk estimation. To quantify this would require a large cohort with cardiovascular risk factors including both lipid measurements, hard outcome measures and sufficient follow-up, and without treatment changes during follow-up.220,221 Databases of routine clinical data, such as the CPRD,222 can provide cohorts of sufficient size, coverage and follow-up; however, they will be subject to a large amount of treatment initiation or change during follow-up, especially in those, or those estimated to be, at high risk of CVD. Particularly problematic for a comparison of TC/HDL cholesterol to another lipid is that under current practice much of such treatment change will have been triggered by high risk according to TC/HDL cholesterol. Therefore, we can estimate that using non-HDL cholesterol could produce appreciably different risk estimates but cannot determine whether or not these would be more predictive of future CVD. Other limitations of our analyses include the use of an approximate risk equation for non-HDL cholesterol, and an approximate though well-established method for adjusting for regression dilution.218,223,224
Taking multiple blood samples in order to average a lipid measurement such as TC/HDL cholesterol ratio would increase financial and practical costs but is unlikely to appreciably improve cardiovascular risk estimation. Improvements to cardiovascular risk estimates, within a lipid monitoring programme, could instead be sought through alternative choices of lipid measure. Our illustrative results, using non-HDL cholesterol in a modified cardiovascular risk equation, show that this has potential to appreciably alter risk estimates. However, at present the data do not exist to definitively determine whether or not the estimates based on non-HDL cholesterol or the estimates based on the ratio are more useful in clinical practice.
