Upper gastrointestinal bleeding and surrogate end points

In their paper, Hawkey et al (Gut 2001;49:372–9) report on the use of blood in the stomach as a surrogate of hard end points such as rebleeding, need for surgery, and death in a group of patients with upper gastrointestinal bleeding randomised to placebo, tranexamic acid, lansoprazole, or both lansoprazole and tranexamic acid. Using logistic regression analyses, they found that blood in the stomach was predictive of clinical outcome and that the active medications decreased the gastric pool size at endoscopy. The authors suggest that although the medications did not significantly improve the rates of clinical outcomes, they would probably do so in a trial of sufficient sample size given the effect achieved on the surrogate endoscopic end point. These conclusions should be viewed with caution, if not with skepticism, for several reasons. Firstly, it is interesting that the authors chose the presence of blood in the stomach as a substitute end point based on a previous study in which no statistical difference was found for the prevalence of this endoscopic sign alone between patients randomised to placebo and omeprazole.¹ Rather, in that study, the presence of a constellation of findings defined as “any sign” (blood, active bleeding, or stigmata) showed a significant reduction in favour of the omeprazole group. The current trial should then be viewed as an exploratory attempt to assess prospectively the independent prognostic contribution of blood in the stomach as a correlate of well defined clinical events. Thus the validity of extrapolating that the active medications will favourably impact on clinical outcomes because they reduce the gastric blood pool in the same cohort of patients in whom a correlation between the surrogate end point and clinical outcomes is being assessed is questionable.

Secondly, no clear definition is provided for what is meant by high or low risk groups, although this factor (risk classification) appears to be a consistent predictor of endoscopic and clinical outcomes. If this categorisation varies among different caregivers, how can we make sure what the regression analyses are truly predicting as it relates to risk status?

Thirdly, there seems to be an imbalance between the number of independent variables and the number of predicted outcomes in the logistic regression analyses. For instance, estimation of rebleeding in the model had only 39 predicted events and at least eight factors selected as potential determinants of clinical outcome (blood in the stomach among them). Such a disproportion between predictors and predicted variables is known to cause overestimation or underestimation of the regression coefficients, thus distorting the estimated effect.^2,3

Finally, application of a marker as a surrogate end point requires demonstration of its accuracy (correlation with the clinical end point) and precision (reproducibility of the marker), with rather restrictive criteria applying to the definition of “surrogate end point”.^4,5 Unfortunately, substitute end points often do not predict the true clinical effects of interventions, and sobering examples remind us that they also turn out to be inadequate markers for clinical practice.⁶ Hawkey et al address in the introduction of their paper the issue of sample size requirements to find significant differences on hard clinical end points. This seems to be a recurring academic exercise in the discussion of intervention studies for gastrointestinal bleeding because trials almost invariably end up with sample sizes that fall short of providing the answers to the relevant questions. The average sample size in a recent meta-analysis of 21 studies of pharmacological treatment for upper gastrointestinal bleeding was only 170,⁷ and that number is larger than the sample size included in the majority of endoscopic treatment trials.

Large scale studies have long been claimed in the field of therapeutic intervention for ulcer bleeding.⁸ This important clinical research area may not be advanced any further if conduction of trials with a small sample size continues; persistence with this methodological approach is sure to delay further progress. Perhaps we should follow the bold and altruistic examples of other fields in medicine, such as that of cardiology, which in one international effort alone assembled over 40 000 patients in order to clarify the benefits of different thrombolytics in the treatment of acute myocardial infarction.⁹ In the long run, the conduction of such trials will be the only defence against misleading reliance on uncertain substitute end points, and will provide at the same time the opportunity to reach firm conclusions derived from direct measures of outcome.¹⁰