Discussion
Like many other disorders in medicine, the prognosis of a particular
patient with HF has a stochastic - rather than deterministic - nature.
As a direct result, a risk model could never have a perfect
discriminatory ability for mortality, regardless of the complexity of
the model. Using too many variables for a risk model not only makes it
less useful for clinical practice, but also increases the risk of
‘overfitting’ - which threatens the accuracy of a model when applied to
populations other than the original derivation sample [20].
Preferably, a model should follow the “law of parsimony” and contain
least number of variables that has the most value, rather than including
every variable that only provides a marginal increase in accuracy.
Present study showed that a simple risk score only consisting of three
variables have a good predictive accuracy for one-year mortality and
performs rather comparably to more complex risk scores such as GWTG-HF
model.
Risk models have important drawbacks that limit their usefulness. A HF
risk model could give inaccurate results when applied to populations
beyond their initial derivation, they are rarely accurate to predict
prognosis for individual patients with HF and they can become obsolete
with time [21,22]. However, they are still convenient as risk models
enable a more objective assessment of the average life expectancy and
they could be useful for selecting optimal management strategy for a
given HF patient [21,22]. Even risk models with external validation
are underutilized in daily clinical practice, perhaps not only because
of the limitations but because of the inconvenience of finding and
entering multiple data to calculate the final score [23]. MAGGIC
risk score, which has a good evidence base for validity and a formidable
c-score of 0.74 for mortality when applied to other HF cohort, needs 13
different variables to be entered [24]. GWTG-HF score had an
acceptable predictive ability for one-year mortality (c-score varied
between 0.64 - 0.67 for HFrEF and HFpEF, respectively), though it needed
a mere 7 variables that made GWTG-HF score somewhat easier to calculate
and more compatible with the law of parsimony [25]. Present findings
indicate that ACEF-MDRD score could predict one-year mortality with an
accuracy comparable to the GWTG score, and similar to the GWTG-HF score
it could be applied to HF populations regardless of the presenting
phenotype. ACEF-MDRD score had the additional advantage of using three
simple and universally available parameters that makes it convenient to
calculate, thus making it somewhat better suited to move beyond the
“research realm” to the real world, as compared to other risk models.
The components of the ACEF score are not only used as standalone
predictors of prognosis in HF, but also one or more of these variables
are commonly found in nearly all HF risk scores [3,4,16,26].
Combining these variables allows an overall estimation of life
expectancy, comorbidities, end-organ function and left ventricular
performance. Despite the availability of multiple studies demonstrating
the predictive ability of ACEF score in a multitude of different
cardiovascular conditions, including patients with recent myocardial
infarction or those undergoing cardiovascular surgery or percutaneous
interventions, data on the prognostic usefulness of ACEF score in
patients with HF is extremely limited [8-12]. Chen and associates
have studied ACEF and ACEF-MDRD in 862 patients with ischemic
cardiomyopathy and found that both scores had a good discriminative
ability (c-statistics were 0.73 for ACEF and 0.72 for ACEF-MDRD,
respectively), though it was not clear whether these patients had
accompanying HF or not as this study was only presented as an abstract
[27]. Present findings suggest that ACEF-MDRD score is an
independent predictor of mortality in all HF patients, regardless of the
underlying etiology, presentation, or phenotype, thus making it a
potentially useful tool for a wide variety of patients.
To note, ACEF-MDRD score was not developed from the present sample but
rather applied to it, and as such present analysis itself should be
considered as a validation study. While there were many studies that
have reported a more impressive predictive accuracy for their models
than the figures provided in this study, they either lack external
validation or their predictive accuracy is substantially lower when
tested in samples other than their derivation cohorts [28]. Given
that provided c-statistics rarely exceed 0.8 for nearly all models,
using an index with a rather modest predictive accuracy could be
justified given the sheer simplicity of the calculation (which could be
done even with a pen and paper) making it practical for daily use and
the lack of “overfitting” - making it suitable for use in different HF
populations [22].
Available treatments for HF are numerous in the contemporary era and
algorithms provided to guide management strategies are not evidence
based. While the main expectation from a risk model is estimation of
overall mortality, it is nonetheless more useful when it could guide
treatment decisions. Several studies have already shown that risk models
could indeed be utilized for this aim. For example, Seattle Heart
Failure Model (SHFM) has been shown to predict mortality after left
ventricular assist device implantation [29]. Whether ACEF-MDRD score
could be utilized in a similar manner would be an interesting prospect
to research in future studies.
Present findings indicate that ACEF-MDRD score had a rather modest
discriminative ability for mortality. Adding new variables to the
equation would be one way to improve the accuracy, since our findings
indicate that ACEF score itself does not explain all the variability in
mortality. However, this approach would violate the founding principle
of ACEF score, which was using a limited number of predictors rather
than every variable with statistical significance on multivariate
analysis. Another way would be finding similar yet more powerful
predictors of mortality to redesign ACEF-MDRD score. Although individual
components of ACEF score are standalone predictors of mortality, it is
not clear whether they are the best predictors, as ACEF score was not
developed to predict mortality after HF. As such, better predictors
could be used to replace core components of the ACEF score, but the law
of parsimony should still be applied to keep the predictors at a
minimum.