Exploratory analyses to guide inclusion, limitation of sample size and strengthening of endpoints in clinical stroke trials

Clinical trials for treatment of acute ischaemic stroke require large numbers of patients and are expensive to conduct. Treatment is typically administered within the first hours or days after stroke onset. Outcome is usually assessed by a single measure, the most common being the modified Rankin scale (mRS) at day 90. Any strategy that can reduce cost or deliver more reliable answers on safety and efficacy of the investigational treatment would be welcome for future exploratory testing of novel treatments. This thesis focused on the impact of applying different methods of design, inclusion and outcome measurement to limit sample size and strengthen analysis in clinical trials in acute stroke. Firstly, inclusion criteria were investigated to assess the impact on functional outcome. By assessing how the effect of thrombolysis changes over onset time to treatment (OTT) the relationship between OTT and age could be investigated. By looking across the entire range of OTT and assessing the interaction between the two covariates this provided complementary data to a previous VISTA analysis conducted by Mishra et al. It was found that across the full range of OTT, up to 3.5h, the treatment effect of thrombolysis in very elderly stroke patients (>80 years old) was comparable to that of their younger counterparts. The association of AF and modified Rankin Scale (mRS) at day 90 was then assessed. Multiple logistic regression analysis adjusted for age and baseline National Institutes of Health Stroke Scale (NIHSS) showed that history of AF had no independent impact on stroke outcome. Deferred selection of subjects for neurorestorative therapies from hyperacute (<6h) to 24h was then explored using a simulation approach. The sample size required to detect a ‘shift’ in mRS outcome equivalent to a 5% absolute difference in proportion achieving mRS 0-2 versus 3-6 was modelled, setting power at 80% and assuming adjustment for entry age and NIHSS. It was found that extending the time window for patient selection provides a measurement which has a stronger more predictive relationship with outcome. Trial inclusion was explored further by investigating selection for delayed treatment with thrombolysis. Prognostic scoring methods were proposed to identify a strategy for patient selection to be applied first to an existing trial dataset and then validated in the pooled RCT 4.5-6h data. ). Prognostic score limits were chosen to optimise the sample for a net treatment benefit significant at p=0.01 by Cochran Mantel Haenszel test and by ordinal logistic regression. More inclusive limits were also defined based on p=0.05 criteria. After finalising prognostic score limits, for validation they were applied by an independent statistician to the pooled RCT data for 4.5-6h. The validation analysis based on ordinal outcomes failed to deliver a population in whom treatment >4.5h was safe and effective; analysis based on net benefit (mRS 0-1) showed significance. Secondly, different strategies for endpoint selection were considered. In the past some trialists have investigated the use of earlier endpoints on single trial datasets and taken advantage of the fact that numerous outcome scales are available to measure various domains of neurological and functional recovery. The use of an earlier neurological endpoint for detecting futility in a trial was considered with validation on external RCT data. Global endpoints, investigating different aspects of functional recovery at different time-points were then considered. Simulations were undertaken to assess the relationship between sample size and power for ordinal scales and the corresponding global outcomes. Day 7 NIHSS was found to be the most sensitive individual ordinal endpoint. Dichotomised analyses supported these results. However this needed validation in a randomised trial dataset for use in exploratory stroke trials. The validation study reinforced the results from the non-randomised VISTA study. The global test combination of NIHSS90 with NIHSS7 appeared to offer incremental sensitivity to treatment effect compared to the ordinal scales alone. The combination of mRS90 with NIHSS7 did not increase the sensitivity to treatment effect when compared to NIHSS alone, but offers a broader clinical measure without loss of statistical power. Finally, alternatives to the traditional RCT were considered. Abandoning the rigour of the blinded RCT carries substantial penalty in loss of reliability and should not be undertaken lightly. If a placebo control is deemed impractical or unethical, investigators often consider comparisons against historical controls. A within-VISTA exploration of case-control matching is presented. The reliability of different matching methods and covariate combinations were assessed using a simulation approach. The results indicate that caution must be taken when using historical controls to generate a matched control group. Substantial further work matching to external data and validation to RCT data is needed. Cluster randomised trials, which randomise patients by groups, are becoming a more widely used approach. When evaluating strategies to promote the transfer of research findings into clinical practice, i.e. in "Implementation Research", an RCT is impractical and a cluster randomised trial design is of advantage. Some elements in the design and sample size calculation of cluster randomised trials were considered. Intra cluster correlation coefficients (ICCs) were estimated from linear and generalised linear mixed models using maximum likelihood estimation for common measures used in stroke research. These estimates of relevant ICCs should assist in the design and planning of cluster randomised trials. In conclusion, this research has shown that there are several areas in the design of clinical trials of acute stroke that merit further investigation. Several strategies have been highlighted that could potentially reduce sample size whilst retaining optimal levels of statistical power. However other aspects such as patient selection and the nature of the intervention under study can affect trial cost and statistical power and need to be taken under consideration