A Method to Determine Stroke Trial Success Using Multidimensional Pooled Control Functions

Abstract

Many early phase trials in stroke have not been subsequently confirmed. Randomization balance in baseline factors that influence outcome are difficult to achieve and may be partly responsible for misleading early results. We hypothesized that comparison with an outcome function derived from a large number of pooled control arms would mitigate these randomization problems and provide a reliable predictor for decision-making before proceeding to later phase trials. We developed such a model and added a novel feature of generation of multidimensional 95% prediction surfaces by which individual studies could be compared. We performed a proof-of-principle study with published clinical trials, determining whether our method correctly identified known outcomes. The control arms from all randomized, controlled trials for acute stroke with >or=10 subjects, including baseline National Institute of Health Stroke Scale, age, and 3-month outcomes published between 1994 and May 2008, were identified. A Matlab program (PPREDICTS) was written to generate outcome functions based on these parameters. Published treatment trials were compared with these 95% intervals to determine whether it successfully identified positive and negative trials. Models of mortality and functional outcome were successfully generated (mortality: R(2)=0.69; functional outcome, modified Rankin Scale 0 to 2: R(2)=0.81; both P<0.0001). The National Institute of Neurological Diseases and Stroke intravenous recombinant tissue plasminogen activator trial and 3 studies yet to be subjected to Phase III study had modified Rankin Scale 0 to 2 outcomes above the 95% prediction interval. Sixteen treatment arm outcomes fell within prediction surface bounds. This group included 2 major trials, Stroke-Acute Ischemic NXY Treatment and Abciximab Emergent Stroke Treatment Trial, that initially appeared promising but went on to negative Phase III results. This proof-of-principle analysis confirmed all positive and negative clinical stroke trial results and identified some promising therapies. The use of a pooled standard treatment group function combined with statistical bounds may improve selection of early studies for further study. This method may be applicable to any condition in which baseline factors influence outcome and at any stage of the development process.