To the Editor:
Dr Mehta and colleagues provided an informative overview of
incidence, risk factors, time trends, and outcomes regarding reoperations
for bleeding after coronary surgery.
In addition, they performed a logistic regression analysis to provide
a model that predicts the risk of reoperation for bleeding. The resulting
model was derived from an impressive database and consists of 19
variables. With a c statistic of only 0.60, this model showed a very
modest discrimination. A c statistic of 1 indicates perfect discrimination
while 0.5 equals flipping a coin. Generally, prediction models with a c
statistic of at least 0.75 are considered to discriminate well.
A more parsimonious bedside tool was also constructed based upon the
derived model. The number of variables was reduced to 12, but
discriminatory performance was not reported.
When keeping the tradeoff between parsimony and performance in mind,
either discriminatory performance should be better or the level of
parsimony of both model and bedside tool should be higher to justify their
implementation in clinical practice.
One of the variables used is serum creatinin level. As this is a
continuous variable, readily available, and a well known predictor for
many adverse outcomes, we chose to construct a model that predicts the
risk of reoperation for bleeding based upon this variable alone. Between
January 2003 and January 2008, 1873 patients underwent coronary surgery in
our center, of which 108 (5.8%) required a reoperation for bleeding. We
used 80% of the cohort for the development of the model and the remaining
20% of the cohort for model validation. The derived model already obtained
a c statistic of 0.63 when applied to the validation cohort.
Although these results are from a cohort of different size and
characteristics, comparable results are likely to be obtained when
applying this approach to the dataset used by Dr Mehta and colleagues.
In other words: the simple clinical rule that increased serum creatinin
levels result in a higher risk of reoperation for bleeding could well have
a discriminatory capability comparable to the model described by Mehta et
al.
In Dr. Mehta's study, what are the discriminatory results for the bedside tool and what c statistic is obtained when, for example, only serum creatinin is used in the model?
In any case one can argue that with a c statistic of only 0.60 common
clinical sense will likely outperform preoperative models that predict the
risk of reoperation for bleeding after coronary surgery in terms of
parsimony and discrimination.
Menno van Gameren, MD
A. Pieter Kappetein, MD, PhD
Ad J.J.C. Bogers, MD, PhD
Johanna J.M. Takkenberg, MD, PhD
Erasmus University Medical Center
Rotterdam, The Netherlands
Conflict of Interest Disclosures: None
Read 