| Summary: | <p><strong>Background</strong></p>
<p>Large randomised controlled trials (RCTs) in cardiovascular disease have provided reliable evidence for interventions whose widespread use has contributed to the secular declines in mortality. But, increasing costs of conducting RCTs threaten our ability to generate new randomised evidence. Linkage to health data collected during a person’s routine care (i.e. routine data) has the potential to substantially streamline trial follow-up, however, there remains considerable uncertainty about the reliability of such information.</p>
<p>The primary aims of this thesis were to investigate, whether for large cardiovascular RCTs in the UK, follow-up of their primary and secondary outcomes via routinely collected hospital admission and death registry data: (1) are ascertained as accurately and completely as adjudicated direct follow-up; and (2) can yield the same estimated treatment effects as adjudicated direct follow-up. Secondary aims were: (3) to assess whether clinical adjudication of direct trial follow-up is necessary; and (4) to use simulated data to investigate more generally the impact on randomised trials of removing true events and adding false events, as may happen if routine data are used for follow-up.</p>
<p><strong>Methods</strong></p>
<p>A systematic review of the existing literature was conducted, where MEDLINE and Embase were searched between 1st January 2010 and 12th August 2022 using a search strategy and selection criteria to identify studies assessing the reliability of UK routine data at ascertaining cardiovascular outcomes compared to trial follow-up data.</p>
<p>Analyses were then carried out using two large trials: (1) the ASCEND (A Study of Cardiovascular Events in Diabetes) primary prevention trial randomised 15480 UK people with diabetes using a 2x2 factorial design to aspirin versus matching placebo, and separately to omega-3 fatty acids versus placebo. The primary efficacy outcome was serious vascular events (SVEs; including non-fatal myocardial infarction [MI], ischaemic stroke or transient ischaemic attack [TIA], or vascular death [excluding intracranial haemorrhage]) and primary safety outcome for the aspirin comparison was major bleeding (including intracranial haemorrhage, sight-threatening eye bleeding, serious gastrointestinal bleeding, and other major bleeding). Adjudicated direct participant mail-based follow-up was the main source of outcome data, and participants were linked to their routinely collected hospital admission and death registry data, where an algorithm was applied to distinguish between major vs minor bleeding events. (2) SHARP (Study of Heart and Renal Protection) was a placebo-controlled trial of simvastatin plus ezetimibe in 9270 people with moderate-to-advanced chronic kidney disease (CKD). The primary outcome was major atherosclerotic events (MAEs; including non-fatal MI, coronary death, non-haemorrhagic stroke, or revascularisation), with the secondary outcome being kidney replacement therapy (KRT; including maintenance dialysis and kidney transplantation). Adjudicated direct participant clinic-based follow-up was the main source of outcome data, and 1576 participants living in England were linked to their routine data. Kappa statistics were used to assess agreement between follow-up data sources (where appropriate), and randomised comparisons were re-run using the outcomes dataset based on the different data sources.</p>
<p>Finally, simulation methods were used to investigate the impact on randomised trials more generally of removing true events and adding false events, as may happen if routine data are used for follow-up. Initially randomised trials were generated to match those of ASCEND’s (n=15480) aspirin comparisons for any SVE (risk ratio=0.88) and major bleeding (risk ratio=1.29). Scenarios using smaller (n=1000 and 5000) and even larger (n=25000) sample sizes were then considered with different treatment effects (RR 0.70, 0.80, 0.90, 1.10, 1.20, 1.30). These data were considered the ground-truth. To simulate routine data follow-up, true events were removed (0% to 40%) and/or false events added (0% to 40%), where participants in either arm had an equal probability of having an error occurring (i.e. errors were random rather than systematic). The primary performance measure was the mean difference between the ground-truth and routine data risk ratios, with the secondary performance measure being the difference in power.</p>
<p><strong>Results</strong></p>
<p>The systematic review identified three studies that compared cardiovascular and bleeding outcomes ascertained via UK routine data with adjudicated direct follow-up, two of which assessed old routine data collected over two decades ago (i.e. 1989-2001), and the other followed-up participants for only 28 days. Nevertheless, for cardiovascular outcomes, agreement between data sources was strong for revascularisations, and moderate for MIs and strokes. Re-running randomised comparisons using routine data only provided relative treatment effect estimates similar to adjudicated direct follow-up. For major bleeding events, routine data showed moderate completeness but accuracy was low, because the study’s routine data outcome definition included major/minor bleeding and anaemia events. No studies were identified that assessed kidney disease outcomes.</p>
<p>In the ASCEND trial, agreement between modern routine data (i.e. 2005-2017) and adjudicated direct follow-up was strong for SVEs (kappa 0.78, 95% confidence interval [CI] 0.76-0.80), with similar levels of agreement for all SVE components except TIA (kappa 0.43, 95% CI 0.36-0.49). For the aspirin randomised comparison, rate ratios (RRs) for SVEs were similar for adjudicated direct follow-up versus routine data alone (adjudicated data: RR 0.88, 95% CI 0.79-0.97; routine data: RR 0.91, 95% CI 0.81-1.02), and almost identical for SVEs excluding TIA (adjudicated data: RR 0.92, 95% CI 0.82-1.03; routine data: RR 0.91, 95% CI 0.80-1.02). For the major bleeding outcome, despite only moderate agreement between datasets (kappa 0.53, 95% CI 0.49-0.57), RRs and absolute treatment effect estimates were similar for adjudicated direct follow-up and routine data (adjudicated data: RR 1.29, 95% CI 1.09-1.52; absolute excess +6.3/5000 person-years [mean standard error ±2.1]; routine data: RR 1.21, 95% CI 1.03-1.41; absolute excess +5.0/5000[±2.2]). Overall, the interpretation of ASCEND’s aspirin randomised comparison (i.e. the benefits of aspirin being largely counterbalanced by the bleeding hazard) would have not changed had follow-up been solely via routinely collected data, with similar findings for pre-adjudicated mail-based follow-up.</p>
<p>For the SHARP trial in people with moderate-to-severe CKD, agreement between routine data and adjudicated follow-up for MAEs was moderate-to-strong (kappa 0.67, 95% CI 0.61-0.72) and very strong for KRT (kappa 0.87, 95% CI 0.84-0.90). Analysis of SHARP’s pre-adjudicated data showed that about one-quarter of patient reported MAEs were refuted. Randomised analyses using SHARP’s pre-adjudicated data found similar results to the adjudicated data (pre-adjudication: RR 0.80, 95% CI 0.72-0.89; adjudicated: RR 0.83, 95% CI 0.74-0.94) and also suggested refuted MAEs were likely to represent atherosclerotic disease (refuted MAEs: RR 0.80, 95% CI 0.65−1.00).</p>
<p>Analysing simulated trial data from a wide range of trial scenarios showed that routine data sources with small-to-moderate random errors (i.e. true event removed and false events added ≤20%, which are equally distributed between trial arms) have little impact on relative treatment effect estimates in randomised trials. In the situation where routine data sources do introduce large amounts of error into a trial (i.e. true event removed and false events added >20%), larger trials (n≥5000) are better protected against random errors biasing the treatment effect towards the null and reducing trial power.</p>
<p><strong>Conclusions</strong></p>
<p>In summary, data from large cardiovascular trials in people with diabetes and CKD indicate that linkage to routinely collected UK hospital admission and death registry data may be a reliable method of streamlining follow-up for non-fatal MI, hospitalised ischaemic stroke, vascular death (including its subtypes), arterial revascularisation, kidney replacement therapy, and major bleeding outcomes. More generally, analysis of real and simulated randomised trial data consistently showed that small-to-moderate random errors in trial follow-up (which can be expected whatever the method of follow-up) had little impact on relative treatment effect estimates on account of the protection afforded by large-scale randomisation. This thesis provides a rationale to consider using UK routine data as the sole source of follow-up for future cardiovascular trials.</p>
|
|---|