CARDIOVASCULAR JOURNAL OF AFRICA • Vol 24, No 9/10, October/November 2013
AFRICA
377
importance of absolute risk estimation in people with diabetes
as the appropriate basis for CVD risk-factor modification. Such
an approach is further supported by the gradual shift in the
management of diabetes mellitus from a glucocentric focus to
an intensive multifactorial strategy targeting reduction in the risk
of both macro- and microvascular complications of diabetes.
12,13
The growing recognition of the importance of global CVD
risk in people with diabetes has generated interest among
researchers to develop tools with improved performance to
estimate absolute risk in people with diabetes, or to establish
the validity of the existing ones and refine their performance.
7
The following development is a discussion on the rationale and
strategies for global CVD risk estimation in people with diabetes,
with emphasis on the specificities and limitations of these
strategies. The discussion is largely inspired by new knowledge
gained from CVD risk modelling in the ADVANCE study.
3,14
Overview of global cardiovascular risk
assessment
Global cardiovascular risk assessment is based on the combination
of predictive information from several cardiovascular risk factors
using mathematical equations (also called models). In those
models, the coefficient of each included risk factor indicates
its relative contribution to the overall (global) CVD risk.
2,15
A
model can be used to estimate the risk that a disease is present
(diagnostic model) or to estimate the risk that a particular disease
or health event will occur within a given time period (prognostic
models). The focus of the current article is on prognostic models.
Once developed, a cardiovascular risk model normally
requires a validation in both the sample population that was used
to develop the model (internal validation) and in independent
populations (external validation). Validation consists of testing
whether the prognostic model accurately estimates the risk of
future events in one or several populations.
2,15
The performance of absolute cardiovascular risk models
in validation studies is commonly assessed in terms of
discrimination, calibration and, more recently, reclassification.
2,15
Discrimination is the ability of the model to distinguish people
who go on to develop a cardiovascular event and those who
remain event free.
2,15
For example, for two individuals with
diabetes with one developing a cardiovascular event after 10
years of follow up and the other remaining CVD free within that
same time period, a discriminating model will systematically
assign, at the start of the follow up, a higher absolute risk to the
first subject compared to the second.
Discrimination is commonly assessed using the
C
-statistic,
which ranges from 0.5 (lack of discrimination) to 1.0 (perfect
discrimination).
1,2,15
In general, a
C
-statistic of 0.7 or greater is
considered acceptable.
Calibration describes the agreement between estimated and
observed risks. It is assessed by comparing absolute risk
estimates from the model with the actual event rates in the test
population.
1,2,15
For illustration, a 10-year estimated absolute risk
of CVD of 20% for a patient indicates that, in a given group
of patients with similar characteristics, 20% will experience a
cardiovascular event within a 10-year period of follow up.
The most commonly reported measure of calibration is
the Hosmer-Lemeshow statistic. Estimates of calibration are
sensitive to differences in background levels of risk across
populations. For example, if a given CVD risk model is
developed in a high-risk population but tested in a low-risk
population, the estimated absolute risks will be unreliably high.
Recalibration of the risk model by adjusting the baseline risk
estimates to fit the target population may help correcting the
over- or underestimation of risk.
1,15
Global cardiovascular risk estimation in
people with diabetes
Global CVD risk has been estimated in people with diabetes
using essentially three main approaches.
16
In the ‘CVD risk-
equivalent’ approach described above, the presence of diabetes
mellitus is considered to confer a 10-year absolute CVD risk of
20% or more, which is approximately the 10-year CVD event
rate observed in non-diabetic individuals with a prior history
of CVD. Such an approach appears to be counter-intuitive as
the CVD risk is not uniformly distributed among people with
diabetes. This is further supported by many studies showing
multivariable risk estimation to be significantly better than
classification of diabetes as a cardiovascular risk equivalent.
17,18
In the second approach, also termed ‘step approach’, unifying
CVD risk-estimation models are developed for both people
with diabetes and those without the condition. This approach
assumes that major risk factors for CVD are related to future
occurrence of CVD in a similar way, regardless of the status for
diabetes mellitus. Stated otherwise, everything else being equal,
an individual with diabetes will always have a higher risk of
CVD (by a constant amount) than the non-diabetic subject with
the same level of other risk factors (e.g. blood pressure or lipid
levels). This has been the basis for models such as the popular
Framingham cardiovascular absolute-risk models.
16
In the last approach, also known as the ‘interaction approach’,
CVD risk models are constructed separately for people with
and without diabetes. This approach suggests that risk factors
are related to future CVD risk in different ways in people with
and without diabetes. This approach in people with diabetes was
initially used by the UKPDS investigators.
9,19
Available studies
largely suggest that classical cardiovascular risk factors (including
smoking, blood pressure and lipid variables) and even some novel
risk factors,
16,20-23
affect the risk of CVD in similar ways in people
with and without diabetes with no evidence of interaction.
Some risk factors or characteristics are likely to be more
frequent in people with diabetes and may justify separate
cardiovascular risk models for people with diabetes. These
diabetes-specific characteristics include prescriptions of
cardiovascular risk-reducing therapies, which may differ in
people with and without diabetes. Additional specific factors
are haemoglobin A
1c
(HbA
1c
) levels, urinary albumin excretion
rate and markers of microvascular complications of diabetes in
general (especially retinopathy). These have been demonstrated
to be associated with CVD risk and can contribute useful
information to predictions.
24-29
Performance of popular CVD risk models and
the ADVANCE study
At the time the ADVANCE study was conducted, CVD risk-
prediction models in the general population were dominated
by models developed from the Framingham Heart study, which
