- Email: [email protected]
The cardiovascular event reduction tool (CERT)—a simplified cardiac risk prediction model developed from the West of Scotland Coronary Prevention Study (WOSCOPS)
The Cardiovascular Event Reduction Tool (CERT)—A Simplified Cardiac Risk Prediction Model Developed from the West of Scotland Coronary Prevention Study (WOSCOPS) Gilbert L’Italien, PhD, Ian Ford, PhD, John Norrie, MSc, Pablo LaPuerta, MD, Jenifer Ehreth, PhD, Joseph Jackson, PhD, and James Shepherd, MD The clinical decision to treat hypercholesterolemia is premised on an awareness of patient risk, and cardiac risk prediction models offer a practical means of determining such risk. However, these models are based on observational cohorts where estimates of the treatment benefit are largely inferred. The West of Scotland Coronary Prevention Study (WOSCOPS) provides an opportunity to develop a risk-benefit prediction model from the actual observed primary event reduction seen in the trial. Five-year Cox model risk estimates were derived from all WOSCOPS subjects (n ⴝ 6,595 men, aged 45 to 64 years old at baseline) using factors previously shown to be predictive of definite fatal coronary heart disease or nonfatal myocardial infarction. Model risk factors included age, diastolic blood pressure, total cholesterol/ high-density lipoprotein ratio (TC/HDL), current smoking, diabetes, family history of fatal coronary heart disease, nitrate use or angina, and treatment (placebo/ 40-mg pravastatin). All risk factors were expressed as
categorical variables to facilitate risk assessment. Risk estimates were incorporated into a simple, hand-held slide rule or risk tool. Risk estimates were identified for 5-year age bands (45 to 65 years), 4 categories of TC/HDL ratio (<5.5, 5.5 to <6.5, 6.5 to <7.5, >7.5), 2 levels of diastolic blood pressure (<90, >90 mm Hg), from 0 to 3 additional risk factors (current smoking, diabetes, family history of premature fatal coronary heart disease, nitrate use or angina), and pravastatin treatment. Five-year risk estimates ranged from 2% in very low-risk subjects to 61% in the very high-risk subjects. Risk reduction due to pravastatin treatment averaged 31%. Thus, the Cardiovascular Event Reduction Tool (CERT) is a risk prediction model derived from the WOSCOPS trial. Its use will help physicians identify patients who will benefit from cholesterol reduction. 䊚2000 by Excerpta Medica, Inc. (Am J Cardiol 2000;85:720 –724)
xisting cardiac risk prediction models have been developed primarily from observational epidemiE ologic cohorts, and not from interventional trial
typically subjective and based upon a mix of quantitative (e.g., mean low-density lipoprotein [LDL] concentration) and qualitative information. Risk assessment models can help to inform the physician on how to treat patients, but they must be simple to use and practical in the busy office setting. This study describes the development of the CERT, which incorporates cardiac risk prediction based upon the real incidence of cardiac events (fatal coronary heart disease, nonfatal myocardial infarction) observed in WOSCOPS.
1,2
data. Although an epidemiologic model can reliably quantify baseline risk, model-derived estimates of the diminished risk due to therapeutic intervention may be inaccurate. For example, the observed event reduction in the West of Scotland Coronary Prevention Trial (WOSCOPS) was greater than that predicted by the Framingham equations.3 The WOSCOPS study evaluated the use of pravastatin for primary prevention of heart disease in patients with hypercholesterolemia. A disadvantage of current risk assessment methods is that they are often computationally complex. This complexity frequently renders them impractical for regular use in the office setting.4 – 6 Thus, despite the proliferation of baseline risk assessments and guidelines endorsing their use (e.g., National Cholesterol Education Program), physician assessment of risk is From Bristol-Myers Squibb, Princeton, New Jersey; and University of Glasgow, Glasgow, Scotland. Manuscript received June 8, 1999; revised manuscript received and accepted October 27, 1999. Address for reprints: Gilbert L’Italien, PhD, Outcomes Research, Bristol-Myers Squibb, 5 Research Parkway, Wallingford, Connecticut 06410. E-mail: [email protected].
720
©2000 by Excerpta Medica, Inc. All rights reserved. The American Journal of Cardiology Vol. 85 March 15, 2000
METHODS The design of WOSCOPS has been previously described.3,7 In the present analysis, Cox proportional hazards regression8 was performed for the entire study population (intention-to-treat), placebo and treated groups (n ⫽ 6,595 men, aged 45 to 64 years old at baseline) using factors previously shown to be predictive of the primary event for the trial population.9 Model risk factors included age, diastolic blood pressure, total cholesterol/high-density lipoprotein ratio (TC/HDL), current smoking, diabetes, family history of fatal coronary heart disease, nitrate use or angina, and treatment (placebo/40 mg pravastatin). The end 0002-9149/00/$–see front matter PII S0002-9149(99)00847-4
points of interest were definite fatal coronary heart disease or definite nonfatal myocardial infarction. Five-year risk estimates were obtained from the proportional hazards equation as follows: 1 ⫺ S(t;z) ⫽ [S0(t)]exp(Bz), where S0(t) is the baseline survivor function at time t (t ⫽ 5 years) corresponding to survival at z ⫽ 0. B is the set of Cox regression coefficients, and z is the pattern of risk factors. A table of risk estimates was developed from the Cox coefficients and incorporated into a simple, hand-held slide rule. The risk table provided estimates at 5-year age bands (between 45 to 65 years), 4 categories of TC/ HDL ratio (⬍5.5, 5.5 to ⬍6.5, 6.5 to ⬍7.5, ⱖ7.5), 2 levels of diastolic blood pressure (⬍90, ⱖ90 mm Hg), from 0 to 3 additional risk factors (current smoking, diabetes, family history of premature fatal coronary heart disease, nitrate use or angina), and treatment (placebo/40 mg pravastatin). The risk table also provided an estimate of the risk reduction due to treatment across all risk factor categories. All continuous risk factors (diastolic blood pressure, TC/HDL, age) were expressed as categorical variables that were defined as terms corresponding to combinations of any 1, 2, or 3 of these risk factors. There were no subjects with all 4 risk factors. This strategy was designed to facilitate risk estimation and permit the tabulation of risk estimates corresponding to a variety of risk factor combinations. However, this method may produce a discrepancy between risk computed using the CERT method and the standard proportional hazards regression, wherein continuous variables are regressed as such and specific binary variables are included. To evaluate this potential discrepancy, the following analysis was performed. (1) The 5-year risk of an event was computed for each patient using both the CERT method and the standard Cox regression described previously, and a correlation coefficient was computed to assess the degree of correlation between the 2 estimates. (2) Patients were grouped according to the Cox regression-derived risk into 5% risk categories (i.e., 0% to ⬍5%, 5% to ⬍10%, 10% to ⬍15%, and so forth). (3) Mean risks were computed for all patients within risk categories using both methods. (4) Plots of mean estimates were obtained (⫾ SD). The use of a clinical trial cohort to develop a risk model raises concern that the model may not be applicable to a more general population. To evaluate the performance of the model when applied to an external population, CERT-derived risk estimates were computed for a subset of men from the Third National Health and Nutrition Examination10 population (NHANES-III 1988 to 1994) aged 45 to 64 years with a LDL of 156 to 254 mg/dl. Subjects were grouped according to the 5% risk categories, and the proportion of subjects who fell within each risk category were compared between the NHANES-III and WOSCOPS placebo group. All computations were performed using SAS Statistical Analysis Software (SAS Institute, Cary, North Carolina). PROC PHREG was used to perform all Cox regression analyses and PROC CORR was used
TABLE I Risk Factors Associated With the Primary End Point Variables Placebo Pravastatin Age (45–⬍50) (yrs) 50–⬍55 55–⬍60 60–⬍65 TC/HDL ratio (⬍5) 5.5–⬍6.5 6.5–⬍7.5 ⱖ7.5 DBP ⬍90 mm Hg DBP ⱖ90 mm Hg Added risk factors None Any 1* Any 2* Any 3*
Risk Ratio
95% CI
p Value
n
Referent 0.68 Referent
0.58–0.83
0.0001
3,293 3,302 1,441
1.09–2.13 1.50–2.85 1.85–3.51
0.014 0.0001 0.0001
0.88–1.56 0.99–1.78 1.35–2.37
0.28 0.059 0.0001
1.05–1.55
0.014
1,784 1,803 1,567 1,750 1,957 1,535 1,353 4,182 2,413
1.31–2.00 2.59–4.84 3.31–15.2
0.0001 0.0001 0.0001
3,252 2,967 355 21
1.52 2.07 2.55 Referent 1.17 1.33 1.79 Referent 1.28 Referent 1.62 3.54 7.09
*Current smoking, diabetes, family history of premature fatal coronary heart disease, nitrate use or angina. CI ⫽ confidence interval; DBP ⫽ diastolic blood pressure.
to perform the correlation analysis between CERT and standard Cox regression.
RESULTS The CERT model results are shown in Table I for all model variables. All continuous risk factors (diastolic blood pressure, TC/HDL, age) were expressed as categorical variables that were defined as terms corresponding to combinations of any 1, 2, or 3 of these risk factors. To assess the benefit of therapy, a term for treatment was included in the model. The inclusion of treated patients and a treatment variable did not appreciably change the hazard ratios for the other variables. The hazard ratios shown in Table I illustrate the independent predictive value of increasing levels of risk factors such as age, diastolic blood pressure, and TC/HDL ratio. The incremental increase in risk with increasing numbers of additional categorical variables is also indicated in Table I. The 5-year estimates of the risk of a primary end point were derived according to the previously described equation and are displayed in Table II. These values are used to obtain patient-specific risk estimates as illustrated in the following. Example: The following example illustrates the use of CERT to obtain a risk estimate. The risk factor information is first obtained during the patient’s workup. CERT consists of an inner card containing the risk table (Figure 1) and an outer sleeve with “windows” that permits the clinician to match a patient’s specific set of risk factors to the 5-year risk of an event. The only computation that may be necessary is the TC/ HDL ratio, and this is often provided in the lab report. The patient is a 54-year-old smoker with a family history of fatal coronary heart disease, blood pressure of 140/90 mm Hg, and total cholesterol/HDL ratio of 6.6. To estimate risk, the age, diastolic blood pressure, and TC/HDL variables are first considered, then the
PREVENTIVE CARDIOLOGY/CARDIOVASCULAR EVENT REDUCTION TOOL
721
TABLE II Five-Year Risk Estimates According to Risk Factor Categories Age (yrs)
TC/HDL ⬍5.5
5.5–⬍6.5
6.5–⬍7.5
ⱖ7.5
⬍50
⬍50
50–⬍55
50–⬍55
DP⬍90
DP⬎90
DP⬍90
DP⬎90
Added Risk Factors*
Untreated
Pravastatin
Untreated
Pravastatin
Untreated
Pravastatin
Untreated
Pravastatin
0 1 2 3 0 1 2 3 0 1 2 3 0 1 2 3
2% 4% 8% 15% 3% 4% 9% 17% 3% 5% 10% 19% 4% 6% 13% 25%
2% 2% 5% 10% 2% 3% 6% 12% 2% 3% 7% 14% 3% 4% 9% 18%
3% 5% 10% 19% 3% 5% 11% 22% 4% 6% 13% 24% 5% 8% 17% 31%
2% 3% 7% 13% 2% 4% 8% 15% 3% 4% 9% 17% 3% 6% 12% 22%
3% 5% 12% 22% 4% 6% 13% 25% 5% 7% 15% 28% 6% 10% 20% 36%
2% 4% 8% 15% 3% 4% 9% 18% 3% 5% 11% 20% 4% 7% 14% 26%
4% 7% 15% 27% 5% 8% 17% 31% 6% 9% 19% 34% 8% 12% 24% 43%
3% 5% 10% 19% 3% 6% 12% 22% 4% 6% 13% 25% 5% 8% 17% 32%
Age (yrs)
TC/HDL
⬍5.5
5.5–⬍6.5
6.5–⬍7.5
ⱖ7.5
55–⬍60
55–⬍60
60–⬍65
60–⬍65
DP⬍90
DP⬎90
DP⬍90
DP⬎90
Added Risk factors*
Untreated
Pravastatin
Untreated
Pravastatin
Untreated
Pravastatin
Untreated
Pravastatin
0 1 2 3 0 1 2 3 0 1 2 3 0 1 2 3
5% 7% 15% 28% 5% 9% 18% 32% 6% 10% 20% 36% 8% 13% 26% 45%
3% 5% 11% 20% 4% 6% 12% 23% 4% 7% 14% 26% 6% 9% 18% 33%
6% 9% 19% 35% 7% 11% 22% 39% 8% 12% 25% 43% 10% 16% 32% 53%
4% 6% 14% 25% 5% 7% 16% 29% 5% 8% 17% 32% 7% 11% 23% 40%
6% 9% 19% 34% 7% 10% 21% 38% 7% 12% 24% 42% 10% 16% 31% 52%
4% 6% 13% 25% 5% 7% 15% 28% 5% 8% 17% 31% 7% 11% 22% 39%
7% 11% 23% 41% 8% 13% 27% 46% 9% 15% 29% 50% 12% 19% 38% 61%
5% 8% 16% 30% 6% 9% 19% 34% 6% 10% 21% 38% 9% 14% 27% 47%
*Current smoking, diabetes, family history of premature coronary heart disease, nitrate use, or angina. DP ⫽ diastolic pressure.
other risk variables are added to incrementally increase the risk estimate. The clinician adjusts the outer sleeve window to identify the age (step 1: 50 to 55 years), the diastolic blood pressure (step 2: ⬎90 mm Hg), and the TC/HDL ratio (step 3: 6.5 to 7.5) corresponding to those of the patient. Because this patient also has 2 additional risk factors (smoker, family history), the clinician further identifies the row corresponding to that number of risk factors (step 4: 2 risk factors). As shown in Figure 1, this patient’s 5-year risk of an event is 19%. If the patient is treated with pravastatin, the estimated risk decreases to 13%. Thus, using readily available data, the clinician obtains the 5-year risk of fatal coronary heart disease and/or nonfatal myocardial infarction (19%) and the 5-year risk after treatment with pravastatin (13%). 722 THE AMERICAN JOURNAL OF CARDIOLOGY姞
VOL. 85
These estimates may be derived without any computation by the clinician. The results of the analysis comparing the CERT method to standard Cox proportional hazards regression are presented in Table III and Figure 2. There was excellent correspondence between the mean CERT and standard Cox risk estimates (correlation coefficient ⫽ 0.942, n ⫽ 6,595). Beyond 15% risk, the CERT method tends to slightly underestimate the standard Cox risk, although the number of patients is much smaller in the higher risk categories. Also, the variability of the CERT estimates, as expressed by the SD, is greater than the standard Cox method variability at all risk levels. This is to be expected, because the variance of the Cox estimates is constrained by using risk categories derived from these estimates. MARCH 15, 2000
As shown in Table IV, the risk distributions of WOSCOPS patients who received placebo (n ⫽ 3,293) and NHANES-III subjects are very similar across the risk ranges shown. The estimated population of United States individuals who fall within this category is approximately 4 million.
DISCUSSION This study was not intended to elucidate or identify specific predictors of cardiovascular disease among WOSCOPS patients. This study has been done previously for both cohorts using more appropriate epidemiologic methods.9 The goal of this study was to develop a utilitarian risk model for deriving risk and risk reduction estiFIGURE 1. Use of CERT example. mates associated with an intervention. Existing risk prediction models are based primarily on observational epidemiologic cohorts,1,2 not interventional trials, and typically require TABLE III Comparison of Risk Estimates: Standard Cox Regression Versus CERT lengthy calculations. Although baseMethod line risk assessments should be reliStandard Cox Regression CERT Method able, estimates of the benefit of an Cox Risk intervention may be inaccurate. Range n Mean SD Mean SD Minimum Maximum Based on the evidence from the 0–⬍5% 2,748 3.4% 1.0% 3.7% 1.2% 1.5% 8.9% WOSCOPS trial,3 the actual risk re5–⬍10% 2,642 7.0% 1.4% 6.8% 1.8% 3.4% 15.1% duction was greater than that which 10–⬍15% 815 12.0% 1.4% 11.1% 3.1% 5.0% 25.8% 15–⬍20% 241 17.0% 1.4% 15.0% 4.9% 8.4% 31.8% would have been predicted by the 20–⬍25% 84 22.0% 1.4% 20.5% 7.1% 8.2% 35.7% Framingham equations. Given that 25–⬍30% 36 27.1% 1.4% 25.2% 6.5% 11.8% 37.5% the parameter estimates were similar ⱖ30% 29 35.8% 7.9% 32.7% 9.1% 15.6% 53.4% for risk factors with and without inclusion of the treatment variable, CERT can be used for both baseline risk estimation and risk reduction with therapeutic intervention. The CERT model is based on the real event reductions seen in the WOSCOPS trial, using the same spectrum of risk factors which National Cholesterol Education Program guidelines recommend for consideration. In particular, use of the TC/HDL ratio has been advocated over LDL or TC alone because of its superior predictive value.6,9,11 A parsimonious model was developed to facilitate risk assessment and thus increase its ease of use in actual practice. The CERT method, which converts continuous to categorical variables to facilitate risk estimation, provides 5-year risk estimates that are comparable to those obtained using standard Cox regression techniques. However, the variability of the CERT estimates is uniformly greater than the standard estimates; this is most likely due to equal weighting of the discrete variables (current smoking, diabetes, family history of premature fatal coronary heart disease, nitrate use, or angina). Lastly, CERT-derived risk estimates are similarly distributed among an external population of persons (NHANES-III) with similar LDL and age ranges. Because the risk distributions are FIGURE 2. Comparison of 5-year risk estimates (%) (mean ⴞ SD) between the CERT method and standard Cox regression. comparable between the 2 populations, the CERT PREVENTIVE CARDIOLOGY/CARDIOVASCULAR EVENT REDUCTION TOOL
723
TABLE IV Comparison of CERT-Computed Risk Distribution* CERT Risk Range
U.S. Males (n ⫽ 4 million) (%)
WOSCOPS Placebo (n ⫽ 3,293) (%)
0–⬍5% 5–⬍10% 10–⬍15% 15–⬍20% 20–⬍25% 25–⬍30% ⱖ30%
27% 43% 13% 8% 3% 2% 4%
27% 52% 14% 5% 1% 1% 0%
*United States Men (NHANES-III) aged 45 to 64 years with LDL-cholesterol 156 to 254 mg/dl versus WOSCOPS placebo.
performance should be comparable to a model derived from a more general population. Also, the CERT method provides risk estimates that are similar to the Framingham Heart Study.2 For example, the 5-year Framingham-based risk of coronary heart disease for a 58-year-old male diabetic smoker with a diastolic blood pressure of 95 mm Hg and a TC/HDL ratio of 8.3 is 29.9%. This is very similar to the CERT-based estimate of 31.7%. There are 2 main limitations to the use of CERT. First, as stated previously, there could be a disparity between risk estimates computed using continuous values for age, TC/HDL ratio, diastolic blood pressure, and the specified discrete variables and risk estimates based on categorical and equally weighted variables as was done for CERT. Given the observed correspondence between CERT-derived and standard Cox regression estimates, the degree of this disparity seems acceptable. Second, because CERT is based on the exclusively male WOSCOPS trial, its applicability is limited to men, aged 45 to 64 years, as shown. However, a recent posthoc gender-specific analysis of the Air Force/Texas Coronary Atherosclerosis Prevention Study trial, implies a similar risk in women who are 10 years older than men with comparable risk factor profiles.12 Existing risk models are not commonly used in busy clinical practice5 despite efforts to
724 THE AMERICAN JOURNAL OF CARDIOLOGY姞
VOL. 85
facilitate this process.1,2,4,13 This may be due, in part, to the complexity of the risk computation. The main goal of CERT is to provide office or hospital-based clinicians with easily obtained, yet reliable quantitative risk and risk reduction estimates that have actually been demonstrated with pravastatin therapy in an interventional trial (WOSCOPS). Unlike more elaborate computer algorithms or risk scoring systems, CERT does not require extensive computation. The use of CERT should facilitate the identification of subjects who would benefit from pravastatin therapy. CERT is intended for use by clinicians during the patient’s examination, as an aid to decision making on whether treatment is needed. 1. Wilson PWF, D’Agostino, Levy D, Belanger, AM, Silbershatz, H, Kannel WB.
Prediction of coronary heart disease using risk factor categories. Circulation 1998;97:1837–1847. 2. Anderson KM, Wilson PWF, Odell PM, Kannel WB. An updated coronary risk profile: a statement for health professionals. Circulation 1991;83:357–363. 3. The West of Scotland Coronary Prevention Study Group. Influence of pravastatin and plasma lipids on clinical events in the West of Scotland Coronary Prevention trial. Circulation 1998;97:1440 –1445 4. Fager G. Cholesterol reduction and clinical benefit. Are there limits to our expectations? Arterioscler Thromb Vasc Biol 1997;17:3527–3533. 5. Greenland P, Grundy S, Pasternak RC, Lenfant C. Problems on the pathway from risk assessment to risk reduction. Circulation 1998;97:1761–1762. 6. Muldoon FH, Criqui MH. The emerging role of statins in the prevention of coronary heart disease. Statins are effective but we need better ways of assessing risk. BMJ 1997;315:1554 –1555. 7. The West of Scotland Coronary Prevention Study Group. A coronary primary prevention study of scottish men aged 45– 64 years. Trial design. J Clin Epidemiol 1992;34:849 – 860. 8. Collett D. Modeling Survival Data in Medical Research. London: Chapman and Hall, 1994. 9. The West of Scotland Coronary Prevention Study Group. Baseline risk factors and their association with outcome in the west of Scotland coronary prevention study. Am J Cardiol 1997;79:756 –762.. 10. U.S. Department of Health and Human Services (DHHS). National Center for Health Statistics. Third National Health and Nutrition Examination Survey, 1988 –1994. Hyattsville, MD: Centers for Disease Control and Prevention, 1996. 11. Kinosian B, Glick H, Garland G. Cholesterol and coronary heart disease: predicting risk by levels and ratios. Ann Intern Med 1994;121:641– 647. 12. Downs JR, Clearfield M, Weis S, Whitney E, Shapiro DR, Beere PA, Gotto AM. Air Force/Texas Coronary Atherosclerosis Prevention Study: primary prevention of coronary heart disease in women. Circulation 1998;98(suppl. 17):I-46. 13. Alvins AL, Browner WS. Improving prediction of coronary heart disease to aid the management of high cholesterol levels. What a difference a decade makes. JAMA 1998;279:445– 449.
MARCH 15, 2000
categorical variables to facilitate risk assessment. Risk estimates were incorporated into a simple, hand-held slide rule or risk tool. Risk estimates were identified for 5-year age bands (45 to 65 years), 4 categories of TC/HDL ratio (<5.5, 5.5 to <6.5, 6.5 to <7.5, >7.5), 2 levels of diastolic blood pressure (<90, >90 mm Hg), from 0 to 3 additional risk factors (current smoking, diabetes, family history of premature fatal coronary heart disease, nitrate use or angina), and pravastatin treatment. Five-year risk estimates ranged from 2% in very low-risk subjects to 61% in the very high-risk subjects. Risk reduction due to pravastatin treatment averaged 31%. Thus, the Cardiovascular Event Reduction Tool (CERT) is a risk prediction model derived from the WOSCOPS trial. Its use will help physicians identify patients who will benefit from cholesterol reduction. 䊚2000 by Excerpta Medica, Inc. (Am J Cardiol 2000;85:720 –724)
xisting cardiac risk prediction models have been developed primarily from observational epidemiE ologic cohorts, and not from interventional trial
typically subjective and based upon a mix of quantitative (e.g., mean low-density lipoprotein [LDL] concentration) and qualitative information. Risk assessment models can help to inform the physician on how to treat patients, but they must be simple to use and practical in the busy office setting. This study describes the development of the CERT, which incorporates cardiac risk prediction based upon the real incidence of cardiac events (fatal coronary heart disease, nonfatal myocardial infarction) observed in WOSCOPS.
1,2
data. Although an epidemiologic model can reliably quantify baseline risk, model-derived estimates of the diminished risk due to therapeutic intervention may be inaccurate. For example, the observed event reduction in the West of Scotland Coronary Prevention Trial (WOSCOPS) was greater than that predicted by the Framingham equations.3 The WOSCOPS study evaluated the use of pravastatin for primary prevention of heart disease in patients with hypercholesterolemia. A disadvantage of current risk assessment methods is that they are often computationally complex. This complexity frequently renders them impractical for regular use in the office setting.4 – 6 Thus, despite the proliferation of baseline risk assessments and guidelines endorsing their use (e.g., National Cholesterol Education Program), physician assessment of risk is From Bristol-Myers Squibb, Princeton, New Jersey; and University of Glasgow, Glasgow, Scotland. Manuscript received June 8, 1999; revised manuscript received and accepted October 27, 1999. Address for reprints: Gilbert L’Italien, PhD, Outcomes Research, Bristol-Myers Squibb, 5 Research Parkway, Wallingford, Connecticut 06410. E-mail: [email protected].
720
©2000 by Excerpta Medica, Inc. All rights reserved. The American Journal of Cardiology Vol. 85 March 15, 2000
METHODS The design of WOSCOPS has been previously described.3,7 In the present analysis, Cox proportional hazards regression8 was performed for the entire study population (intention-to-treat), placebo and treated groups (n ⫽ 6,595 men, aged 45 to 64 years old at baseline) using factors previously shown to be predictive of the primary event for the trial population.9 Model risk factors included age, diastolic blood pressure, total cholesterol/high-density lipoprotein ratio (TC/HDL), current smoking, diabetes, family history of fatal coronary heart disease, nitrate use or angina, and treatment (placebo/40 mg pravastatin). The end 0002-9149/00/$–see front matter PII S0002-9149(99)00847-4
points of interest were definite fatal coronary heart disease or definite nonfatal myocardial infarction. Five-year risk estimates were obtained from the proportional hazards equation as follows: 1 ⫺ S(t;z) ⫽ [S0(t)]exp(Bz), where S0(t) is the baseline survivor function at time t (t ⫽ 5 years) corresponding to survival at z ⫽ 0. B is the set of Cox regression coefficients, and z is the pattern of risk factors. A table of risk estimates was developed from the Cox coefficients and incorporated into a simple, hand-held slide rule. The risk table provided estimates at 5-year age bands (between 45 to 65 years), 4 categories of TC/ HDL ratio (⬍5.5, 5.5 to ⬍6.5, 6.5 to ⬍7.5, ⱖ7.5), 2 levels of diastolic blood pressure (⬍90, ⱖ90 mm Hg), from 0 to 3 additional risk factors (current smoking, diabetes, family history of premature fatal coronary heart disease, nitrate use or angina), and treatment (placebo/40 mg pravastatin). The risk table also provided an estimate of the risk reduction due to treatment across all risk factor categories. All continuous risk factors (diastolic blood pressure, TC/HDL, age) were expressed as categorical variables that were defined as terms corresponding to combinations of any 1, 2, or 3 of these risk factors. There were no subjects with all 4 risk factors. This strategy was designed to facilitate risk estimation and permit the tabulation of risk estimates corresponding to a variety of risk factor combinations. However, this method may produce a discrepancy between risk computed using the CERT method and the standard proportional hazards regression, wherein continuous variables are regressed as such and specific binary variables are included. To evaluate this potential discrepancy, the following analysis was performed. (1) The 5-year risk of an event was computed for each patient using both the CERT method and the standard Cox regression described previously, and a correlation coefficient was computed to assess the degree of correlation between the 2 estimates. (2) Patients were grouped according to the Cox regression-derived risk into 5% risk categories (i.e., 0% to ⬍5%, 5% to ⬍10%, 10% to ⬍15%, and so forth). (3) Mean risks were computed for all patients within risk categories using both methods. (4) Plots of mean estimates were obtained (⫾ SD). The use of a clinical trial cohort to develop a risk model raises concern that the model may not be applicable to a more general population. To evaluate the performance of the model when applied to an external population, CERT-derived risk estimates were computed for a subset of men from the Third National Health and Nutrition Examination10 population (NHANES-III 1988 to 1994) aged 45 to 64 years with a LDL of 156 to 254 mg/dl. Subjects were grouped according to the 5% risk categories, and the proportion of subjects who fell within each risk category were compared between the NHANES-III and WOSCOPS placebo group. All computations were performed using SAS Statistical Analysis Software (SAS Institute, Cary, North Carolina). PROC PHREG was used to perform all Cox regression analyses and PROC CORR was used
TABLE I Risk Factors Associated With the Primary End Point Variables Placebo Pravastatin Age (45–⬍50) (yrs) 50–⬍55 55–⬍60 60–⬍65 TC/HDL ratio (⬍5) 5.5–⬍6.5 6.5–⬍7.5 ⱖ7.5 DBP ⬍90 mm Hg DBP ⱖ90 mm Hg Added risk factors None Any 1* Any 2* Any 3*
Risk Ratio
95% CI
p Value
n
Referent 0.68 Referent
0.58–0.83
0.0001
3,293 3,302 1,441
1.09–2.13 1.50–2.85 1.85–3.51
0.014 0.0001 0.0001
0.88–1.56 0.99–1.78 1.35–2.37
0.28 0.059 0.0001
1.05–1.55
0.014
1,784 1,803 1,567 1,750 1,957 1,535 1,353 4,182 2,413
1.31–2.00 2.59–4.84 3.31–15.2
0.0001 0.0001 0.0001
3,252 2,967 355 21
1.52 2.07 2.55 Referent 1.17 1.33 1.79 Referent 1.28 Referent 1.62 3.54 7.09
*Current smoking, diabetes, family history of premature fatal coronary heart disease, nitrate use or angina. CI ⫽ confidence interval; DBP ⫽ diastolic blood pressure.
to perform the correlation analysis between CERT and standard Cox regression.
RESULTS The CERT model results are shown in Table I for all model variables. All continuous risk factors (diastolic blood pressure, TC/HDL, age) were expressed as categorical variables that were defined as terms corresponding to combinations of any 1, 2, or 3 of these risk factors. To assess the benefit of therapy, a term for treatment was included in the model. The inclusion of treated patients and a treatment variable did not appreciably change the hazard ratios for the other variables. The hazard ratios shown in Table I illustrate the independent predictive value of increasing levels of risk factors such as age, diastolic blood pressure, and TC/HDL ratio. The incremental increase in risk with increasing numbers of additional categorical variables is also indicated in Table I. The 5-year estimates of the risk of a primary end point were derived according to the previously described equation and are displayed in Table II. These values are used to obtain patient-specific risk estimates as illustrated in the following. Example: The following example illustrates the use of CERT to obtain a risk estimate. The risk factor information is first obtained during the patient’s workup. CERT consists of an inner card containing the risk table (Figure 1) and an outer sleeve with “windows” that permits the clinician to match a patient’s specific set of risk factors to the 5-year risk of an event. The only computation that may be necessary is the TC/ HDL ratio, and this is often provided in the lab report. The patient is a 54-year-old smoker with a family history of fatal coronary heart disease, blood pressure of 140/90 mm Hg, and total cholesterol/HDL ratio of 6.6. To estimate risk, the age, diastolic blood pressure, and TC/HDL variables are first considered, then the
PREVENTIVE CARDIOLOGY/CARDIOVASCULAR EVENT REDUCTION TOOL
721
TABLE II Five-Year Risk Estimates According to Risk Factor Categories Age (yrs)
TC/HDL ⬍5.5
5.5–⬍6.5
6.5–⬍7.5
ⱖ7.5
⬍50
⬍50
50–⬍55
50–⬍55
DP⬍90
DP⬎90
DP⬍90
DP⬎90
Added Risk Factors*
Untreated
Pravastatin
Untreated
Pravastatin
Untreated
Pravastatin
Untreated
Pravastatin
0 1 2 3 0 1 2 3 0 1 2 3 0 1 2 3
2% 4% 8% 15% 3% 4% 9% 17% 3% 5% 10% 19% 4% 6% 13% 25%
2% 2% 5% 10% 2% 3% 6% 12% 2% 3% 7% 14% 3% 4% 9% 18%
3% 5% 10% 19% 3% 5% 11% 22% 4% 6% 13% 24% 5% 8% 17% 31%
2% 3% 7% 13% 2% 4% 8% 15% 3% 4% 9% 17% 3% 6% 12% 22%
3% 5% 12% 22% 4% 6% 13% 25% 5% 7% 15% 28% 6% 10% 20% 36%
2% 4% 8% 15% 3% 4% 9% 18% 3% 5% 11% 20% 4% 7% 14% 26%
4% 7% 15% 27% 5% 8% 17% 31% 6% 9% 19% 34% 8% 12% 24% 43%
3% 5% 10% 19% 3% 6% 12% 22% 4% 6% 13% 25% 5% 8% 17% 32%
Age (yrs)
TC/HDL
⬍5.5
5.5–⬍6.5
6.5–⬍7.5
ⱖ7.5
55–⬍60
55–⬍60
60–⬍65
60–⬍65
DP⬍90
DP⬎90
DP⬍90
DP⬎90
Added Risk factors*
Untreated
Pravastatin
Untreated
Pravastatin
Untreated
Pravastatin
Untreated
Pravastatin
0 1 2 3 0 1 2 3 0 1 2 3 0 1 2 3
5% 7% 15% 28% 5% 9% 18% 32% 6% 10% 20% 36% 8% 13% 26% 45%
3% 5% 11% 20% 4% 6% 12% 23% 4% 7% 14% 26% 6% 9% 18% 33%
6% 9% 19% 35% 7% 11% 22% 39% 8% 12% 25% 43% 10% 16% 32% 53%
4% 6% 14% 25% 5% 7% 16% 29% 5% 8% 17% 32% 7% 11% 23% 40%
6% 9% 19% 34% 7% 10% 21% 38% 7% 12% 24% 42% 10% 16% 31% 52%
4% 6% 13% 25% 5% 7% 15% 28% 5% 8% 17% 31% 7% 11% 22% 39%
7% 11% 23% 41% 8% 13% 27% 46% 9% 15% 29% 50% 12% 19% 38% 61%
5% 8% 16% 30% 6% 9% 19% 34% 6% 10% 21% 38% 9% 14% 27% 47%
*Current smoking, diabetes, family history of premature coronary heart disease, nitrate use, or angina. DP ⫽ diastolic pressure.
other risk variables are added to incrementally increase the risk estimate. The clinician adjusts the outer sleeve window to identify the age (step 1: 50 to 55 years), the diastolic blood pressure (step 2: ⬎90 mm Hg), and the TC/HDL ratio (step 3: 6.5 to 7.5) corresponding to those of the patient. Because this patient also has 2 additional risk factors (smoker, family history), the clinician further identifies the row corresponding to that number of risk factors (step 4: 2 risk factors). As shown in Figure 1, this patient’s 5-year risk of an event is 19%. If the patient is treated with pravastatin, the estimated risk decreases to 13%. Thus, using readily available data, the clinician obtains the 5-year risk of fatal coronary heart disease and/or nonfatal myocardial infarction (19%) and the 5-year risk after treatment with pravastatin (13%). 722 THE AMERICAN JOURNAL OF CARDIOLOGY姞
VOL. 85
These estimates may be derived without any computation by the clinician. The results of the analysis comparing the CERT method to standard Cox proportional hazards regression are presented in Table III and Figure 2. There was excellent correspondence between the mean CERT and standard Cox risk estimates (correlation coefficient ⫽ 0.942, n ⫽ 6,595). Beyond 15% risk, the CERT method tends to slightly underestimate the standard Cox risk, although the number of patients is much smaller in the higher risk categories. Also, the variability of the CERT estimates, as expressed by the SD, is greater than the standard Cox method variability at all risk levels. This is to be expected, because the variance of the Cox estimates is constrained by using risk categories derived from these estimates. MARCH 15, 2000
As shown in Table IV, the risk distributions of WOSCOPS patients who received placebo (n ⫽ 3,293) and NHANES-III subjects are very similar across the risk ranges shown. The estimated population of United States individuals who fall within this category is approximately 4 million.
DISCUSSION This study was not intended to elucidate or identify specific predictors of cardiovascular disease among WOSCOPS patients. This study has been done previously for both cohorts using more appropriate epidemiologic methods.9 The goal of this study was to develop a utilitarian risk model for deriving risk and risk reduction estiFIGURE 1. Use of CERT example. mates associated with an intervention. Existing risk prediction models are based primarily on observational epidemiologic cohorts,1,2 not interventional trials, and typically require TABLE III Comparison of Risk Estimates: Standard Cox Regression Versus CERT lengthy calculations. Although baseMethod line risk assessments should be reliStandard Cox Regression CERT Method able, estimates of the benefit of an Cox Risk intervention may be inaccurate. Range n Mean SD Mean SD Minimum Maximum Based on the evidence from the 0–⬍5% 2,748 3.4% 1.0% 3.7% 1.2% 1.5% 8.9% WOSCOPS trial,3 the actual risk re5–⬍10% 2,642 7.0% 1.4% 6.8% 1.8% 3.4% 15.1% duction was greater than that which 10–⬍15% 815 12.0% 1.4% 11.1% 3.1% 5.0% 25.8% 15–⬍20% 241 17.0% 1.4% 15.0% 4.9% 8.4% 31.8% would have been predicted by the 20–⬍25% 84 22.0% 1.4% 20.5% 7.1% 8.2% 35.7% Framingham equations. Given that 25–⬍30% 36 27.1% 1.4% 25.2% 6.5% 11.8% 37.5% the parameter estimates were similar ⱖ30% 29 35.8% 7.9% 32.7% 9.1% 15.6% 53.4% for risk factors with and without inclusion of the treatment variable, CERT can be used for both baseline risk estimation and risk reduction with therapeutic intervention. The CERT model is based on the real event reductions seen in the WOSCOPS trial, using the same spectrum of risk factors which National Cholesterol Education Program guidelines recommend for consideration. In particular, use of the TC/HDL ratio has been advocated over LDL or TC alone because of its superior predictive value.6,9,11 A parsimonious model was developed to facilitate risk assessment and thus increase its ease of use in actual practice. The CERT method, which converts continuous to categorical variables to facilitate risk estimation, provides 5-year risk estimates that are comparable to those obtained using standard Cox regression techniques. However, the variability of the CERT estimates is uniformly greater than the standard estimates; this is most likely due to equal weighting of the discrete variables (current smoking, diabetes, family history of premature fatal coronary heart disease, nitrate use, or angina). Lastly, CERT-derived risk estimates are similarly distributed among an external population of persons (NHANES-III) with similar LDL and age ranges. Because the risk distributions are FIGURE 2. Comparison of 5-year risk estimates (%) (mean ⴞ SD) between the CERT method and standard Cox regression. comparable between the 2 populations, the CERT PREVENTIVE CARDIOLOGY/CARDIOVASCULAR EVENT REDUCTION TOOL
723
TABLE IV Comparison of CERT-Computed Risk Distribution* CERT Risk Range
U.S. Males (n ⫽ 4 million) (%)
WOSCOPS Placebo (n ⫽ 3,293) (%)
0–⬍5% 5–⬍10% 10–⬍15% 15–⬍20% 20–⬍25% 25–⬍30% ⱖ30%
27% 43% 13% 8% 3% 2% 4%
27% 52% 14% 5% 1% 1% 0%
*United States Men (NHANES-III) aged 45 to 64 years with LDL-cholesterol 156 to 254 mg/dl versus WOSCOPS placebo.
performance should be comparable to a model derived from a more general population. Also, the CERT method provides risk estimates that are similar to the Framingham Heart Study.2 For example, the 5-year Framingham-based risk of coronary heart disease for a 58-year-old male diabetic smoker with a diastolic blood pressure of 95 mm Hg and a TC/HDL ratio of 8.3 is 29.9%. This is very similar to the CERT-based estimate of 31.7%. There are 2 main limitations to the use of CERT. First, as stated previously, there could be a disparity between risk estimates computed using continuous values for age, TC/HDL ratio, diastolic blood pressure, and the specified discrete variables and risk estimates based on categorical and equally weighted variables as was done for CERT. Given the observed correspondence between CERT-derived and standard Cox regression estimates, the degree of this disparity seems acceptable. Second, because CERT is based on the exclusively male WOSCOPS trial, its applicability is limited to men, aged 45 to 64 years, as shown. However, a recent posthoc gender-specific analysis of the Air Force/Texas Coronary Atherosclerosis Prevention Study trial, implies a similar risk in women who are 10 years older than men with comparable risk factor profiles.12 Existing risk models are not commonly used in busy clinical practice5 despite efforts to
724 THE AMERICAN JOURNAL OF CARDIOLOGY姞
VOL. 85
facilitate this process.1,2,4,13 This may be due, in part, to the complexity of the risk computation. The main goal of CERT is to provide office or hospital-based clinicians with easily obtained, yet reliable quantitative risk and risk reduction estimates that have actually been demonstrated with pravastatin therapy in an interventional trial (WOSCOPS). Unlike more elaborate computer algorithms or risk scoring systems, CERT does not require extensive computation. The use of CERT should facilitate the identification of subjects who would benefit from pravastatin therapy. CERT is intended for use by clinicians during the patient’s examination, as an aid to decision making on whether treatment is needed. 1. Wilson PWF, D’Agostino, Levy D, Belanger, AM, Silbershatz, H, Kannel WB.
Prediction of coronary heart disease using risk factor categories. Circulation 1998;97:1837–1847. 2. Anderson KM, Wilson PWF, Odell PM, Kannel WB. An updated coronary risk profile: a statement for health professionals. Circulation 1991;83:357–363. 3. The West of Scotland Coronary Prevention Study Group. Influence of pravastatin and plasma lipids on clinical events in the West of Scotland Coronary Prevention trial. Circulation 1998;97:1440 –1445 4. Fager G. Cholesterol reduction and clinical benefit. Are there limits to our expectations? Arterioscler Thromb Vasc Biol 1997;17:3527–3533. 5. Greenland P, Grundy S, Pasternak RC, Lenfant C. Problems on the pathway from risk assessment to risk reduction. Circulation 1998;97:1761–1762. 6. Muldoon FH, Criqui MH. The emerging role of statins in the prevention of coronary heart disease. Statins are effective but we need better ways of assessing risk. BMJ 1997;315:1554 –1555. 7. The West of Scotland Coronary Prevention Study Group. A coronary primary prevention study of scottish men aged 45– 64 years. Trial design. J Clin Epidemiol 1992;34:849 – 860. 8. Collett D. Modeling Survival Data in Medical Research. London: Chapman and Hall, 1994. 9. The West of Scotland Coronary Prevention Study Group. Baseline risk factors and their association with outcome in the west of Scotland coronary prevention study. Am J Cardiol 1997;79:756 –762.. 10. U.S. Department of Health and Human Services (DHHS). National Center for Health Statistics. Third National Health and Nutrition Examination Survey, 1988 –1994. Hyattsville, MD: Centers for Disease Control and Prevention, 1996. 11. Kinosian B, Glick H, Garland G. Cholesterol and coronary heart disease: predicting risk by levels and ratios. Ann Intern Med 1994;121:641– 647. 12. Downs JR, Clearfield M, Weis S, Whitney E, Shapiro DR, Beere PA, Gotto AM. Air Force/Texas Coronary Atherosclerosis Prevention Study: primary prevention of coronary heart disease in women. Circulation 1998;98(suppl. 17):I-46. 13. Alvins AL, Browner WS. Improving prediction of coronary heart disease to aid the management of high cholesterol levels. What a difference a decade makes. JAMA 1998;279:445– 449.
MARCH 15, 2000