Abstract
Introduction
Twenty-five years ago, Geoffrey Rose [1] pointed out that, in a given community, most cardiovascular deaths will come from individuals at moderate risk because they are so numerous. Nevertheless, high-risk individuals gain most from risk factor modification and are therefore given the highest priority in clinical practice.
What is high risk?
The current European guidelines on cardiovascular disease (CVD) prevention [2] define high risk as: (i) patients with established coronary heart disease, peripheral artery disease and cerebrovascular atherosclerotic disease; (ii) asymptomatic individuals who are at high risk of developing atherosclerotic cardiovascular disease because of: multiple risk factors resulting in a 10-year risk of 5% or greater now (or if extrapolated to age 60 years) for developing a fatal CVD event; markedly raised levels of single risk factors: cholesterol 8 mmol/l or less (320 mg/dl), low density lipoprotein cholesterol 6 mmol/l or greater (240 mg/dl), blood pressure 180/110 mmHg or higher.
This definition of high risk immediately exposes a difficulty; young people, although at a low absolute risk, may have very high relative risk. By extrapolating their risk to age 60 years, their high-risk status will be emphasized, but such a procedure might encourage excessive drug usage in young people.
The need for total risk estimation
All current guidelines on prevention, particularly the current European ones [2], stress the need for total risk estimation in all individuals who have not declared themselves at high risk by the presence of established disease. This is because an individual's risk is, in the great majority of cases, the product of multiple risk factors. For example, a 60-year-old man with a cholesterol level of 5 mmol/l who smokes and is hypertensive may be at a 10 times higher risk than a 60-year-old lady with a cholesterol level of 8 mmol/l, if the latter is free of all other risk factors.
Risk prediction systems
A number of attempts have been made to facilitate CVD risk prediction for clinicians. Most of these have been based on the American Framingham study [3]. The Prospective Cardiovascular Munster Study (PROCAM) programme is a sophisticated risk prediction system based on a group of German employee volunteers [4]. More recently, the European Systematic COronary Risk Evaluation (SCORE) project attempted to improve the representativeness and thus the applicability of risk prediction to European populations by assembling a risk prevention system based upon 205 178 European individuals derived from 12 population studies [5].
Most of these risk prediction systems have used either Cox or Weibull-based methods to estimate risk. In the SCORE project, both approaches performed similarly. Although it is possible that a neural network-based approach might enhance risk prediction [6], the complexity of this approach has limited its acceptance to date.
Recently, Topol and Lauer [7] contrasted the Framingham (n = 5345) and SCORE (n = 205 178) systems, both with over 10 years of follow-up. One of the major differences is that the SCORE system estimates the 10 year absolute risk of cardiovascular death, whereas the Framingham system estimates total events, which includes a combination of coronary heart disease death, myocardial infarction and insufficiency, and angina.
Arising out of this, and out of subsequent debates and discussions with regard to the performance of current risk prediction systems and the potential for improvement, we suggest the following areas for further debate and examination by future researchers.
Which endpoints?
SCORE uses total atherosclerotic cardiovascular deaths, including myocardial infarction, stroke, aneurysms and peripheral vascular disease. Framingham-based systems usually use ‘hard' coronary endpoints, both fatal and non-fatal. Neither solution is either ‘right' or perfect.
The use of total cardiovascular mortality is relevant in that stroke deaths may prove proportionately more important in low-risk populations, and in that it allows recalibration of the charts to suit individual countries. Furthermore, a high risk of death automatically implies a higher risk for non-fatal events. However, clinicians are more comfortable with an actual figure for total events. This is less simple than it sounds, in that morbidity will be determined by event rates, diagnostic criteria, newer diagnostic techniques, secular changes in the natural history of cardiovascular disease, treatment effects, and ascertainment rates. These issues are being addressed in what is becoming known as the ‘SCORE-Plus' project.
Risk prediction in the young
As indicated above, current risk charts correctly indicate that all young people are at low absolute risk. However, this may conceal substantial relative risks. The SCORE solution of extrapolating risk predictions to age 60 years will emphasize that these young people are at a high relative risk, but could lead to overtreatment with drugs. Other approaches to this problem include presentations of relative risk, prediction of ‘risk age', the use of rate advancement periods and expressing relative risk as the position in a queue. A risk score to predict atherosclerotic lesions in young persons was recently described [8].
Risk prediction in the old
As one ages, absolute risk eventually approaches 100% and the estimated risk score will have no discriminatory value. Further cohort studies in older individuals are needed to be sure how far into old age risk prediction can be extended, particularly in societies in which longevity is increasing.
The impact of secular changes in mortality
If a risk chart is based on cohorts 10 or 20 years old, its predictive ability will vary depending on the secular changes in subsequent mortality. As recently pointed out, the SCORE charts could overpredict risk in Norway [9], in that the charts appear to assign a large proportion of middle-aged Norwegian individuals to a high-risk status. This may indeed be the truth, but in view of the longevity of Norwegians, this point needs to be examined further. This issue can be addressed by recalibrating the risk charts on the basis of up-to-date mortality data and current risk factor levels. This has been undertaken for a number of countries, including Sweden [10] and Germany.
The applicability of risk charts to different cultures
The recalibration process described above can deal with this issue to some extent, but individual countries should also be encouraged to develop their own cohort studies to allow a more accurate local risk prediction. For example, a risk score derived from Caucasian cohorts may substantially overpredict risk in a Chinese population [11].
Blood sugar and diabetes
Most risk prediction systems handle diabetes rather crudely, particularly as it is now clear that the relationship with a 2-h post-prandial blood sugar and risk is graded. The incorporation of newer cohort data into risk prediction systems such as SCORE will be an important future development.
Incorporation of newer risk markers
Most risk prediction systems do not allow adequately for lipoprotein (a), homocysteine, fibrinogen, CD-40 ligand or other newer risk markers. Although data from studies such as INTERHEART indicate that simpler markers account for most of the occurrence of vascular disease [12], this needs to be treated with some caution as case-control studies, which estimate relative rather than absolute risk, are less reliable than cohort studies for the estimation of attributable risks.
Interaction effects
Most risk prediction systems assume independent risk factor effects, and interactions between particular combinations of risk factors are difficult to model. This is a major challenge for biostatistics.
Electronic risk prediction and relationship to management guidelines
The accessibility and flexibility of risk prediction has been substantially assisted by the development of HeartScore [13]. HeartScore is an electronic, interactive risk prediction system derived from a partnership between SCORE and the Danish electronic PRECARD [14] system. In its original Windows-based format, HeartScore was slow and rather clumsy to use, but the recently launched web-based version is both flexible and easy to use [13]. Such an approach also facilitates the incorporation of data from newer cohorts, the use of regionally adapted charts, and in particular facilitates interaction with risk-managed guidelines.
What about a ‘lay' version?
It is our belief that risk prediction should be made readily accessible and comprehensible to individuals without formal medical training. Anecdotal experience suggests that many individuals can understand a risk chart almost intuitively. The possibility of developing a fully accessible ‘lay' version of HeartScore was discussed at a recent European Heart Network-European Society Cardiology Workshop and preliminary plans were made to develop this issue further.
Conclusion
Current cardiovascular risk prediction systems work adequately well at assigning individuals to low, medium and high-risk groups. Even confining them to ‘conventional' risk factors, considerable refinement is both possible and planned. No single risk prediction system is currently ideal. Individual countries are strongly encouraged to choose the risk prediction system that matches their culture best. To this end, the European Society of Cardiology has sought the appointment of national coordinators to assist in this process. Recalibrated SCORE charts have been produced or are in development for a number of countries, including Germany, Sweden, Belgium, the Czech Republic, Spain, Ireland, Greece and Italy.