Response to Letter Regarding Article, “Initial Lesion Volume Is an Independent Predictor of Clinical Stroke Outcome at Day 90: An Analysis of the Virtual International Stroke Trials Archive (VISTA) Database”
We thank the authors for their interest in our study and interesting suggestions that we have followed in some reanalyses. We believe, however, that the criticism regarding methodology of our article is largely unjustified.
We have followed the suggestion to examine the interaction of the National Institutes of Health Stroke Scale (NIHSS) at baseline and lesion volume and indeed found an interaction (P=0.02), but only on NIHSS at Day 90 as the end point, not for modified Rankin Scale (mRS; linear) or mortality. The interaction for the NIHSS end point suggests a mutual influence of both baseline parameters resulting in steepening of the baseline NIHSS regression at higher lesion volumes and vice versa. This comparably weak finding should however be regarded with caution at present.
Considering the linear treatment of NIHSS, in biological systems, there is hardly any relationship of complex parameters that is really strictly linear; linear regression is a practical approximation that is most often used due to its simplicity and unproblematic use. In our case, linear regression was used to easily compare models and derive power calculations from the residual variance. Using spline modeling comes to essentially the same conclusions as described in the article. Logistic regression analyses using ordinal NIHSS lesion volume retain its strong influence on outcome.
König and Weimar suggest additional measures of accuracy of prediction. Both sensitivity and specificity and the test suggested by Pencina and colleagues refer to binary outcome prediction. These concepts are not equally useful for describing the quantitative relationship of baseline parameters to linear or ordinal multilevel outcome scales as we have done. However, to give an impression of this sort of predictivity measure, we have compared the receiver operating characteristic area under the curve for the dichotomized mRS (0–1 versus 3–6). Adding lesion volume increases this value from 0.76 (“fair” prediction value) to 0.81 (“good” prediction value).
The term validity was used in the sense of validation of the effect composition of the model, not for proving accuracy of exact equation parameters at this point. The model was based on >2500 patients, which we believe is a quite large data set irrespective of what percentage of the Virtual International Stroke Trials Archive (VISTA) that constitutes. We will provide additional validation with VISTA-independent high-quality data sets in the near future.
Regarding the value of other baseline factors, we do not claim and do not believe that our model is final in the sense that additional parameters may not improve it in the future. The suggested parameters by König and Weimar (early mRS, infarct location) are however not prime candidates at present for that; mRS is a scale designed to assess functioning of patients in a social context, and it is considered extremely difficult by stroke clinicians to get a meaningful readout early after stroke in an intensive care clinical setting. This is why there is virtually no trial we know of that has used mRS other than for late outcome assessment. In the work cited,1 early mRS also takes away any influence of NIHSS at baseline, meaning that one would have to exchange the well-validated NIHSS for the problematic early mRS in models. Using infarct location is very interesting but currently hampered by the lack of an accepted standardized classification system and by the need to enter multiple nominal effects into models, thereby complicating model analyses. Indeed, in the cited article,1 the authors did incorporate the presence or absence of lenticulostriate infarction but did not account for overall infarct location.
We have already discussed possible limitations of our model, one of them being that the model was derived from a database of clinical trials and therefore is likely most suited for examining similar stroke populations.
Finally, the authors question the general approach of adjusting for patients' variation in baseline characteristics in clinical trials on the grounds that this has not led to a successful stroke trial either (as opposed to nonadjusted analyses). This is an irrational argument. Use of adjustment of influential covariates is strongly recommended and discussed in the relevant regulatory guidelines for statistical aspects of clinical trials.2,3
Gerhard Vogt, PhD
Rico Laage, PhD
Armin Schneider, MD
G.V., R.L., and A.S. are employees of SYGNIS Bioscience.
Stroke welcomes Letters to the Editor and will publish them, if suitable, as space permits. Letters must reference a Stroke published-ahead-of-print article or an article printed within the past 3 weeks. The maximum length is 750 words including no more than 5 references and 3 authors. Please submit letters typed double-spaced. Letters may be shortened or edited. Include a completed copyright transfer agreement form (available online at http://stroke.ahajournals.org and http://submit-stroke.ahajournals.org).
- © 2012 American Heart Association, Inc.
German Stroke Study Collaboration. Predicting outcome after acute ischemic stroke: an external validation of prognostic models. Neurology. 2004;62:581–585.
CPMP. Points to Consider on Adjustment for Baseline Covariates (cpmp/ewp/2863/99). London, UK: EMA; 2003.
EMA. ICH Topic E9 Statistical Principles for Clinical Trials. CPMP/ICH/363/96. London, UK: EMA; 1998.