CARDIOVASCULAR JOURNAL OF AFRICA • Vol 24, No 9/10, October/November 2013
378
AFRICA
for many could also be used in people with diabetes.
7
CVD risk
models specific to people with diabetes were also available,
particularly those from the UKPDS study.
7
However, the clinical
utility and comparative performance of these popular CVD risk
models in contemporary populations with diabetes in diverse
settings were still to be established.
Therefore, one of the major initial steps was to conduct
extensive validation studies of the Framingham and UKPDS
CVD risk models, using the unique features of the ADVANCE
cohort.
3
These validation studies revealed that, in the cohort of
ADVANCE participants who had no known history of CVD
at their enrolment in the trial, the four-year absolute risk of
cardiovascular events and components was largely overestimated
by the Framingham–Anderson,
30
Framingham–D’Agostino
31
and
UKPDS risk models.
9,19
This overestimation was also observed
in men and women, Caucasians and non-Caucasians, and the
double-placebo cohort (i.e. those assigned to the placebo group
in the blood pressure-lowering arm and the standard-care group
of the blood glucose control arm).
3
Discrimination of the Framingham and UKPDS risk models
in predicting CVD events in ADVANCE was poor for stroke,
and modest to acceptable for coronary heart disease and total
CVD. Recalibration substantially attenuated the magnitude of
risk overestimation by the Framingham and UKPDS risk models
in ADVANCE. Discrimination was unaffected as expected,
indicating the need for new CVD risk models with improved
predictive accuracy for people with diabetes, particularly those
who are receiving many contemporary cardiovascular risk-
reducing therapies.
Development of the ADVANCE cardiovascular
risk model
In developing a new model for risk prediction, it is critical to
account for the limitations of existing ones in order to improve
performance. The inclusion in ADVANCE of participants from
many countries provided the opportunity to account for the
substantial variation in the care of diabetes and CVD around the
world.Availablemodels so far had been derived fromhomogenous
populations. The ADVANCE model targets total CVD and
therefore captures the interrelation between components of CVD
such as CHD or stroke, unlike many existing models that have
focused specifically on these components.
The complexity of the relationship between chronic
hyperglycaemia and cardiovascular risk has been less fully
addressed in existing models. Some improvement was achieved
in the ADVANCE model through integration of risk factors to
capture both the exposure to chronic hyperglycaemia prior to and
after the clinical diagnosis of diabetes. Statistical method is an
important component of model development. Trusted statistical
methods were used to select the potential risk factors and test
their suitability for inclusion in the ADVANCE risk model.
14
Risk factors considered for inclusion in the ADVANCE model
were: age at clinical diagnosis of diabetes, duration of diagnosed
diabetes, gender, blood pressure (BP) indices [systolic BP,
diastolic BP, mean arterial (MAP) and pulse (PP) pressures], lipid
variables [total, high-density lipoprotein (HDL) and non-HDL
cholesterol, ratio of total:HDL cholesterol and triglycerides],
body mass index (BMI), waist circumference, waist-to-hip ratio,
BP-lowering medication (i.e. treated hypertension), statin use,
current smoking, retinopathy, atrial fibrillation (past or present),
urinary albumin:creatinine ratio (ACR), serum creatinine
(S
cr
), HbA
1c
and fasting blood glucose levels, and randomised
treatments (BP lowering and glucose control regimens).
Ten of these candidate risk factors were included in the
final ADVANCE risk model. Age at diabetes diagnosis and
known duration of diabetes were preferred to age at baseline to
improve the applicability of the ADVANCE risk model to other
populations. The beta coefficients and accompanying standard
error for risk factors in the ADVANCE risk model are shown in
Table 1.
14
Performance of the ADVANCE risk model
The applicability of the ADVANCE risk model
14
was tested on
the same population used to develop the model (i.e. internal
validation) and on an independent external sample for which the
DIAB-HYCAR cohort
32
was used. In both internal and external
validations, the discrimination of the ADVANCE model was
acceptable. In comparison with existing total CVD models,
the ADVANCE model largely outperformed the Framingham–
Anderson and Framingham–D’Agostino models. The calibration
of the ADVANCE model was excellent in internal validation
and good in external validation, with only a modest risk
underestimation. This is likely explained by the difference in the
levels of preventive therapies between ADVANCE and DIAB-
HYCAR population.
TABLE 1. BETA COEFFICIENTS (95% CONFIDENCE INTERVAL)
AND STANDARD ERRORS FOR PREDICTORS IN THE
ADVANCE CVD PREDICTION MODEL
14
Variable
Parameter
estimate
(standard error)
p
-value*
Age at diagnosis (per 1-year increase)
0.062 (0.008)
<
0.001
Gender (women vs men)
–0.474 (0.098)
<
0.001
Known duration of diabetes (per 1-year increase)
0.083 (0.010)
<
0.001
Pulse pressure (per 1-mmHg increase)
0.007 (0.003)
0.016
Retinopathy (yes vs no)
0.383 (0.101)
<
0.001
Atrial fibrillation (present vs absent)
0.601 (0.154)
<
0.001
HbA
1c
(per 1% increase)
0.099 (0.027)
<
0.001
Log of urinary albumin/creatinine ratio (per 1-log
mg/g increase)
0.193 (0.033)
<
0.001
Non-HDL cholesterol (per 1-mmol/l increase)
0.126 (0.034)
<
0.001
Treated hypertension (yes vs no)
0.242 (0.106)
0.022
*Mutually adjusted. Baseline survival probability at four years: S
0
(4)
=
0.951044.
Based on the Cox model, the probability
ˆ
Pof an event at t years of follow up is
defined by the following formula:
ˆ
P
=
1 – S
0
(
t
)
exp (
Σ
p
i
=1
β
i
χ
i
–
Σ
p
i
=1
β
i
–
χ
i
)
where S
0
(t) is the baseline survival at t years;
β
i
is the estimated regression coef-
ficient,
χ
i
is the value of the predictor;
–
χ
i
is the corresponding mean for continuous
predictors (to account for the fact that the value of S
0
(t) is estimated at the mean
level of predictors in the study population); and p denotes the number of predic-
tors.
Consider for example, a man diagnosed with diabetes at the age of 50, with a
known duration of diabetes of three years, a pulse pressure of 50 mmHg and treat-
ed for hypertension, a urinary albumin/creatinine ratio of 50 mg/g, an HbA
1c
level
of 7%, a non-HDL cholesterol level of 3.3 mmol/l, who has retinopathy and atrial
fibrillation. The estimated risk based on the ADVANCE model is:
p
Σ
i
= 1
β
i
χ
i
=
0.062*50
+
0.083*3
+
0.007*50
+
0.242*1
+
0.193*log(50)
+
0.099*7
+
0.126*3.3
+
0.383*1
+
0.601*1 – 0.474*0
=
6.78882
p
Σ
i
= 1
β
i
χ
i
=
0.062*57.94
+
0.083*7.90
+
0.007*64.59
+
0.242*0.644
+
0.193*2.83
+
0.099*7.54
+
0.126*4
+
0.383*0.239
+
0.601*0.054 – 0.474*0.464
=
6.55666
ˆ
P
=
1 – S
0
(
t
)
exp (
Σ
p
i
=1
β
i
χ
i
–
Σ
p
i
=1
β
i
–
χ
i
)
=
1 − 0.951044
exp
(6.78882−6.55666)
= 0.0613, or approximately 6.1%.
