Pitfalls in the reporting of clinical events in unblinded device trials: analysing performance bias in the BIO-GUARD-MI trial
C Jons, P Sogaard, T Smilde, M Taborksy, F Falukozy, M Wiemer, P D Christensen, A Erglis, M Khan, J Marti Almor, S Del Prado Diaz, I Rybar, L Vitali-Serdoz, J Schrader, S BehrensAbstract
Background/Introduction
The BIO|GUARD-MI study was an unblinded trial that investigated the clinical effect of continuous arrhythmia monitoring with an implantable cardiac monitor in high-risk patients after myocardial infarction. The primary endpoint was a composite of cardiovascular (CV) death or CV hospitalisation. A pre-specified interim analysis identified a higher incidence of non-CV adverse events (AEs) in the treatment group. Since these events should be balanced between the study groups, a performance bias was assumed and ultimately the study was prematurely terminated. In the final data set, non-CV AEs were balanced between the study groups (HR 1.08, 95% CI 0.91 – 1.29). The primary analysis of the study did not show a benefit of arrhythmia monitoring (HR = 0.84, 95%-CI 0.65–1.10). [1]
Purpose
While a performance bias in an unblinded trial is not implausible, its nature has not been described. We analyse a specific aspect of the patient’s behaviour, which may be the basis of the presumed bias. Further, we attempt to identify a subgroup of patients that is responsible for it.
Methods
By study protocol, patients were expected to visit the study site only when invited after detected arrhythmias. We compare the numbers of patients with uninvited visits between the study groups in the total cohort and in 14 national subgroups and 10 subgroups based on clinical or demographic characteristics. Patients with uninvited visits were compared with Fisher’s exact test. Time to event analysis was done with the Kaplan-Meier method.
Results
In the total population, similar percentages of patients of treatment and control group visited the site on their own initiative (13.3% vs. 9.7%, P = 0.12). However, in patients from Denmark (21% of the total population) the difference was very large (24.4% vs. 2.3%) and statistically significant after correction for multiple testing (P = 0.0004). No other subgroup showed a difference between the study groups. Visual inspection of the time to the first non-CV AE suggests that these events were more frequent in Danish treatment group patients than in Danish control patients. The latter were similar to patients of both groups from the remaining countries (Figure).
Conclusion
A performance bias is evident in the unblinded BIO|GUARD-MI trial. It was caused by Danish patients, who visited the investigational site more ofter without invitation if they were in the treatment group. In the Danish health system, the general practitioner acts as a gate keeper for secondary and tertiary care, which treatment group patients could bypass. There is compatible evidence that this behaviour led to more clinical events being reported in them. When conducting an unblinded diagnostic trial, continuous assessment of characteristics that indicate unexpected participant behaviour may be reasonable. Such differences may be caused by characteristics of national health systems.