You are on page 1of 1

CORRESPONDENCE

IS THE GENEVA RULE the axes. We also question whether knowledge that the dependence of the
USABLE OUTSIDE OF ROC curves are the best way to illus- Geneva rule on arterial blood gas
trate the diagnostic utility of noncon- analysis is a limitation.
GENEVA?
tinuous data when there are only Regarding the receiver operating
three ordered categories. characteristic (ROC) curves, the axes
To the Editor: Despite these issues, we feel that in part B of the Figure are inverted,
We read with interest the study by the Geneva rule represents the best although those in part A are correct.
Chagnon et al. (1) on clinical predic- standardized method of determining We apologize to the readers of The
tion rules in patients with suspected the pretest probability to date, and it
pulmonary embolism. Although we Green Journal for our error and in-
is for this reason that we are seeking to clude the corrected version below
admire the authors’ attempts to pro- validate its use in a population in New
duce a standardized prediction rule (Figure). However, that error does
Zealand. After comparing the vari- not change the information conveyed
for pulmonary embolism, we believe ability of both the Geneva rule and
that there are some flaws in the meth- by the Figure. Certainly, the absolute
the Wells’ score, we have found that
ods used that are important to iden- sensitivity and specificity of the two
the Wells’ score contains a highly sub-
tify for potential users of this score. cutoff points (between low and inter-
jective component: “Alternative diag-
These have come to light since we re- mediate-high probability, and low-
nosis less likely than pulmonary em-
cently embarked on a prospective intermediate and high probability)
bolism” that renders the score highly
study of pulmonary embolism at are of no direct clinical use, but anal-
variable among users and therefore
Christchurch Hospital, New Zealand, ysis of the area under the ROC curves
offers no advantage over implicit as-
in which we compare the Geneva rule sessment. allows comparisons of the predictive
with implicit risk assessment. Stephen Iles, MRCP accuracy (2), and these areas are un-
First, we would argue that the pres- Aidan Hodges changed even after rectifying the axes.
ence of an elevated hemidiaphragm Chris Frampton, PhD Contrary to Iles et al., we believe that
or atelectasis on an isolated chest ra- Ian Town, DM ROC curve analysis is the most ap-
diograph is of uncertain importance Canterbury Respiratory propriate means of comparing the di-
and not an entirely objective sign as Research Group agnostic accuracy when there are only
suggested, especially if interpreted by Christchurch School of Medicine and three ordered categories. We did not,
less experienced clinicians, who are Health Sciences however, state that the Geneva score
often the first to see such patients. Christchurch Hospital overridden by implicit judgment was
Second, the Geneva rule is highly de- New Zealand superior to the Wells’ score, even
pendent on measurement of arterial though our analyses showed a statis-
blood gases taken while the patient is 1. Chagnon I, Bounameaux H, Aujesky D, et
al. Comparison of two clinical prediction tically significant difference. We
breathing room air. As acknowledged rules and implicit assessment among pa- wrote: “The Geneva rule with clinical
by the authors, this criterion could tients with suspected pulmonary embolism. override was also marginally better
not be applied in 29% of their study Am J Med. 2002;113:269 –275.
than the Wells’ rule. . . , which al-
sample. Oxygen is often given while
ready includes a subjective element.
patients are being assessed as part of a
However, we believe that no conclu-
protocol for the treatment of hyp- The Reply:
sion can be drawn from that observa-
oxia. Finally, we are perplexed by the Iles et al. raise several interesting
tion because the Wells’ score was cal-
receiver operating characteristic points about our paper. Indeed, ele-
(ROC) curves in the Figure. If the vated hemidiaphragm and plate-like culated retrospectively. Indeed, if the
data given in Table 4 are used to re- atelectasis are not highly specific clinicians had had to rely on the
construct curve B, then there are sev- when considered singly, but they are Wells’ score to guide subsequent in-
eral errors. The axes may have been common in patients with proven pul- vestigations, they might have weighed
labelled incorrectly as “1 – specificity” monary embolism, and were clearly the likelihood of an alternative diag-
versus “sensitivity,” when calcula- associated with the disease in our der- nosis differently.” On behalf of the
tions indicate they are plotted the ivation sample, both in univariate authors, I thank Iles et al. for their
other way around. Are we to assume and multivariate analyses (1). More- useful comments and careful reading
that the same errors have been made over, in the original database, those of our work.
in curve A? Furthermore, there is no findings—which are easy to recog- Arnaud Perrier, MD
data point corresponding to the high- nize, even by less experienced clini- Medical Clinic
probability group (assuming the cians—were recorded by residents in Department of Internal Medicine
graph should be inverted), as the charge of patients in the emergency Geneva University Hospital
curves have not been extrapolated to department. In addition, we did ac- Geneva, Switzerland

©2003 by Excerpta Medica Inc. 0002-9343/03/$–see front matter 339


All rights reserved.

You might also like