CARDIOVASCULAR JOURNAL OF AFRICA • Vol 23, No 3, April 2012
162
AFRICA
research focusing on improvement of patient care and outcome.
19
The STS NCD is unparalleled in terms of its size and
comprehensiveness: data were collected prospectively from
more than 950 participating centres in the United States.
3,20
The
STS NCD now also includes more than 3.6 million surgical
procedures.
20
STS risk models for various cardiac procedures have been
developed since 1999 and have undergone periodic revisions.
1,20
A wide variety of endpoints are included in some of the models
calculating risk for isolated coronary artery bypass grafting,
valve surgery or combined surgeries.
1
Twenty-seven new STS
adult cardiac surgery models for 2008 have been developed and
validated.
21
The predictive performance of the STS algorithms is in
general comparable with other systems and remains the most
widely used model in the United States.
1,3
The STS NCD also
does not predict possible morbidity and does not include relation
to any intra-operative variables.
Parsonnet score
The Parsonnet score was first described in 1989 by Victor
Parsonnet. The aim was to construct a straightforward uniform
reporting system for levels of operative mortality risk in all
cardiac surgical procedures, which included data that are readily
available. It includes objective risk factors in order to leave little
room for bias.
3
Development took place in the United States and included
data from 3 500 patients collected between 1982 and 1987.
Retrospectively, analyses included uni- and multivariate logistic
regression models. The model was prospectively tested in an
additional 1 332 procedures at a single site. A second, additive
model was also developed. This method was tested at two other
centres and the outcomes were comparable to those of the
hospitals.
3
The Parsonnet score received widespread acceptance, but
the predictive accuracy has been diminished as a result of
advances in treatment.
1
The original score was later modified in
1994 to include 30 new risk factors, according to the SUMMIT
system, and is known as the ‘modified Parsonnet score’.
22
Again,
no morbidity or relation to intra-operative events are being
predicted.
Major critique of current models
In recent years, several models have predicted a rising
probability of operative mortality while the observed mortality
has decreased.
23
This is due to an increasing prevalence of high-
risk patients, believed to be attributed to significant advances
made in diagnostic and interventional cardiology.
3,5
Risk models
from earlier periods (or retrospectively collected data) can as a
result not be used when the goal of the outcome analysis includes
determination of the trend of mortality over time. Retrospective
data do not only fail to take into calculation the advances in
treatment, but also the evolution of the case mix. Therefore, the
gold standard for data collection should be speciality-specific,
prospectively maintained clinical databases that ought to contain
a core set of variables that have been demonstrated to be
associated with outcome.
24
It is furthermore believed that risk models usually predict
outcome more accurately in the setting where it was originally
developed.
5
Socio-economic conditions, living standards,
healthcare funding, and geographic and ethnic origins affect
the applicability of risk models in different regions.
3
To date no
sub-Saharan African country has developed a risk-stratification
model applicable to the unique pathology of their native
population.
Risk models have diverse clinical aims. The choice of
inclusion/exclusion of risk factors as well as the number of risk
factors included in the model is influenced by the clinical aim.
25
Variables that may affect patient outcome but which are not
necessarily related to pre-operative patient characteristics are
often not taken into account. These include variables related to
adverse intra-operative events as well as co-morbid diseases and
aspects of the disease progression not included in the calculation
of risk.
1
There is no general agreement about the inclusion and
exclusion of these factors.
8
Risk factors associated with outcomes generally are likely
to reflect concurrent, disease-specific variables whereas factors
associated with increased resource utilisation reflect serious
co-morbid disease.
26
It has been suggested that the strength of
scores should be that some kind of grouping is provided for
patient cohorts.
27
Models are sometimes criticised for multicollinearity.
Intercorrelations between independent variables included in risk
models are known as multicollinearity (e.g. obesity and diabetes
mellitus). Including large numbers of independent variables
increases the risk of multicollinearity and the consequent
inclusion of redundant information in the model.
8
Excessively complex models with too many variables will
appear to have an extremely good fit in the training set, but
generalise poorly to test samples and have limited predictive
abilities. This is known as overfitting.
18
It is recommended that
instead of including all statistically significant variables, one
should confine the model to the most powerful predictors or
combinations of variables that are the most powerful predictors.
8
Different operators will provide different interpretations to
categorical risk factors, such as chronic obstructive pulmonary
disease and unstable angina. Even with clearly stated definitions,
a degree of personal interpretation takes place, resulting in
different final risk scores.
8
Wherever practical, continuous data
should be used and there should be strict standardisation of
definitions for the risk factors and the outcomes measured.
18
Some models have been criticised for not being able to predict
individual risk. Currently utilised models are derived from the
studies of very large populations and although very effective at
predicting population outcomes, are not necessarily suited for the
prediction of risk of an individual patient.
1,26
As previously stated,
it is generally accepted that the number of independent variables
that can be included in a multivariate logistic regression depends
on the number of events: there should be a variable-to-event ratio
of 1:10 .
8,25
For that reason, to contemplate a 15 risk-factor model
with a mortality rate of 3%, at least 5 000 cases are required to
achieve adequate sample size.
25
This also means that even in a
unit performing 500 surgeries per annum, it would take at least
10 years to meet the required sample size.
Most of the scores are unsuitable for individual risk
prediction despite the sample size. This is due to a simple
methodological reason: the application of logistic regression
models mathematically describes a multiphasic, more complex
behaviour of a survival curve that cannot achieve enough
