Internal Validity
The MEAD trials were similarly designed masked, sham-controlled RCTs that used appropriate methods to randomize patients (Interactive Voice/Web Response System). The MEAD trials were not initially designed to assess the average BCVA mean change from baseline as the primary end point. Rather, the original end point was the proportion of patients who achieved at least a 15-letter improvement by end of study, which the FDA still considered as the primary end point. Only subsequent to a protocol amendment was the primary end point changed to be the average BCVA mean change from baseline. The manufacturer provided adjustments to the sample size to accommodate the new end point. Based on the sample size calculations, the MEAD studies were designed as a superiority trial with the expectation to show a between-treatment difference of at least four letters. The primary end point used in the MEAD trials (average BCVA mean change from baseline) used the AUC approach. Although this method was considered to be more reliable and was expected to result in a more appropriate control of type I error compared with analysis at every individual time point according to the Health Canada reviewer report, it can also mask the variability of treatment effects across all time points.51 The FDA also commented on the robustness of the AUC approach, noting that the average mean change in BCVA during the study does not differentiate the short-term treatment effect (which the FDA indicated was prior to 36 months) from the long-term treatment effect.52 The FDA refused to accept the amendment to the primary end point and considered the original primary end point (i.e., proportion of patients who achieved at least a 15-letter improvement by end of study) more appropriate.52
The use of the ANCOVA method of analysis would have ensured that the results were adjusted for variables including baseline BCVA and CRT as measured by OCT. All efficacy analyses were conducted using ITT analysis defined as all randomized patients analyzed according to the treatment to which the patient was randomized. In the ANCOVA model, missing data were imputed using the LOCF approach for all end points with the exception of those based on the AUC approach. Excluding patients with missing data is inconsistent with the true definition of an ITT analysis, in which all randomized participants are included. The exclusion of these patients can potentially bias the results, given the pattern of missing data. Overall, more withdrawals occurred in the sham groups compared with the dexamethasone groups in both MEAD trials, especially due to lack of efficacy (more prominently in MEAD-011). Given that missing data were not imputed in the AUC approach used for the primary analysis, the treatment effect may have been biased in favour of dexamethasone 700 mcg given that patients that were doing well would be overrepresented in that group. Furthermore, patients were excluded from the primary analysis if they received escape treatment. Excluding these patients can bias the results if withdrawals due to escape therapy were imbalanced between treatment groups, although this was not the case given that the numbers were well balanced between treatment groups in both MEAD trials. In addition, a sensitivity analysis using a PP population was also conducted; however, this does not lessen the concerns related to exclusion of patients. Furthermore, given that more than 50% of patients discontinued the studies in the both MEAD trials, the LOCF method for imputing data may be biased given that it does not account for patients who had an initial response but could not tolerate treatment. Given that the reasons for withdrawal are related to the study drug and are not balanced between study groups, using the LOCF alone as an imputation method for certain end points may not be sufficient to address the missing data. Sensitivity analyses using multiple imputation methods were evaluated in the overall DME population; however, the results were not reported for the pre-specified subgroup of patients with DME who are pseudophakic. The Health Canada reviewers report states that the results were no longer statistically significant in the overall DME population when using the multiple imputation method to handle missing data.51 The lack of a statistically significant difference from the multiple imputation analysis is due to the larger variance produced by the analysis compared the LOCF method which artificially reduces variance by repeating the same data point when data are missing.51 Overall, per Health Canada, had the missing data actually been captured, they would have been expected to have had some variance — which indicates that the multiple imputation model may be more appropriate than the LOCF method.
Although, different investigators were used to perform the study treatment procedure, follow-up, data collection and data analysis throughout the trial to maintain masking, post-injection safety visits at day one, seven and 21 which were conducted by the treating investigator, resulting in unblinded safety evaluations. Furthermore, the adverse event profile associated with intravitreal steroids (i.e., IOP) is well known, therefore some accidental unblinding may have occurred.12,19 Given that prior intravitreal steroid experience was not an exclusion criterion in any of the trials, some patients with prior experience may have surmised that the allocated treatment was dexamethasone. However, considering that the primary end point of the MEAD trials is relatively objective, the potential for bias is of lesser concern. Unblinding may, however, lead to biases such as under or over reporting of subjective outcomes (i.e., AEs or health-related quality of life measures) which can impact the overall impressions with dexamethasone treatment. Treatment discontinuations during the studies were not reported in any of the MEAD trials. Overall, there were numerically more study discontinuations in the sham groups compared with the dexamethasone groups in both MEAD trials. The European Medicines Agency’s guideline on missing data in confirmatory clinical trials suggests that patients who do not complete a clinical trial may be more likely to have extreme values than patients who complete a trial.53 Therefore, excluding these patients could underestimate the variability and artificially narrow the confidence interval for the treatment effect, and neither the LOCF method nor sensitivity analysis using the PP population would have overcome this potential limitation.
Dexamethasone response at each study visit varied considerably across both MEAD trials. The reason for the relatively large difference in response rate between the two trials remains unclear; however, it may be due to underlying confounders due to the imbalance in baseline characteristics between treatment groups and across studies. In the MEAD trials, the ANCOVA model only adjusted for BCVA and retinal thickness as measure by OCT as covariates. Other baseline line factors which may confound the results were not adjusted, such as duration of diabetes, duration of DME, severity of DR, and type of DME among others. If not adjusted appropriately, such confounding factors may influence the comparative efficacy, though the direction of bias is unclear.
In the MEAD trials, only the primary analyses were controlled for multiple statistical testing using a gate-keeping procedure (i.e., average BCVA mean change from baseline in the general DME population). Adjustments for type I error were conducted using a hierarchical approach for the dexamethasone 700 mcg versus sham comparison at the 0.05 level of significance first, followed by the dexamethasone 350 mcg versus sham comparison (also at the 0.05 level of significance) if the prior analysis was statistically significant. As a result, all other outcomes were not appropriately adjusted for multiplicity, which increases the risk of making a type I error. Further, subgroups typically do not maintain randomization (unless used as stratification variables for randomization, which was not the case). Inadequate randomization introduces biases through the presence of confounders (known and unknown). The imbalances present in the baseline characteristics of the subgroup of adult patients who are pseudophakic may suggest that randomization may have been compromised. Given the distribution of known and unknown confounders, the direction of the bias remains unclear; however, these biases may explain the differing treatment affect between MEAD-010 and MEAD-011. Subgroups are also likely underpowered (small sample size) to detect a statistically significant difference.
The MEAD trials were originally designed to evaluate the effects of dexamethasone in the general DME population. This CDR report is based on the results of a subgroup (i.e., adults with DME who are pseudophakic) which only consisted of approximately 20% of overall enrolled patients. While the risks of type I error and bias remain, the validity of the results for the subgroup of adult patients with DME who are pseudophakic are strengthened by its pre-specified nature and the biological plausibility of the interaction effect. It is known that intravitreal steroid injections result in the development and progression of cataracts (clouding of the natural lens) in eyes that are phakic (natural lens).19 Furthermore, treatment with dexamethasone has been shown to be associated with increased frequency of cataracts, which is outlined in the warning and precautions section of the Health Canada–approved product monograph.12 It is believed that the formation of cataracts can potentially confound the overall treatment effect on the BCVA in patients with DME who are phakic; therefore, patients who have had their natural lens surgically replaced with an artificial lens (pseudophakic) would be expected to further benefit from treatment with dexamethasone given that the formation and progression of cataracts is no longer possible.51
External Validity
In the MEAD trials, 42% to 47% of patients were screening failures when considering the overall DME population. Stringent inclusion and exclusion criteria can result in a highly enriched population, which may not be completely representative of the DME population in Canada and can potentially limit the generalizability of the trial results. Furthermore, the MEAD trials were initially designed to assess the effects of dexamethasone in the general DME population. Only after a notice of non-compliance was issued (due to lack of efficacy due to confounding associated with cataracts) was an analysis subsequently conducted in the pre-specified subgroup of adult patients with DME who are pseudophakic. The Health Canada–approved indication was consequently limited to the treatment of patients with DME who are pseudophakic. Therefore, the focus of this CDR review is based on a subset of the overall DME population. In addition, the protocol for the present review considered further subgroups, including patients who are either unsuitable for anti-VEGF therapy or have had inadequate response to prior anti-VEGF therapy. Between ▬ and ▬ of the pseudophakic patients included in the MEAD trials had prior experience with anti-VEGF therapy, and it remains unclear if these patients responded to these treatments and were truly anti-VEGF refractory. It is also unclear if there were any patients included in the MEAD trials that were considered unsuitable for anti-VEGF therapy. According to the clinical expert consulted for this review, the date of conduct of the trials (between February 2005 to June 2012) was prior to the adoption of anti-VEGF therapies and may therefore may have influenced the number of patients having access to anti-VEGF therapy. It is therefore unclear if the results of the MEAD trials can be generalized to patients who are unsuitable for anti-VEGF therapy or have had an inadequate response to anti-VEGF therapy.
Both MEAD trials were multinational and included sites from Canada. The clinical expert consulted by CDR for this review highlighted that the MEAD trials appear to have recruited patients with characteristics similar to those of the overall DME population in Canada with some exceptions, however, the expert noted that the majority of patients recruited in the MEAD trials were aged > 65 years (▬ to ▬), which may represent an older DME population than what would be observed in Canadian clinical practice.
With respect to the study duration, the FDA suggested that 36 months is considered short term. The FDA recommends that the treatment effect be demonstrated at a time point of at least 36 months or later for the indication of DME given that earlier treatment success is not necessarily a good indicator of a later success.52 Therefore, it is unclear if the results of the MEAD trials would be representative of the long-term treatment effect.
Overall, the relatively large and imbalanced number of discontinuations in the MEAD trials may have led to a DME population that was generally healthier than those initially randomized into the study given that mostly patients who were doing well remained in the trial. This effect is artificially high in the sham groups of the MEAD trials because only patients who did not develop substantial visual deterioration without any treatment were included in the final analyses. Therefore, it remains unclear if the dexamethasone treatment effect observed in the MEAD trials is truly representative of the effect potentially observed in clinical practice.
The dexamethasone retreatment regimen could not occur more frequently than every six months in both MEAD trials. However, the clinical expert consulted for this CDR review suggests that the effects of intravitreal dexamethasone injections wane over time and are rarely sustained at month six, therefore dexamethasone may be used more frequently than every six months in clinical practice. The same expert suggested a retreatment regimen closer to every four months rather than a similar regimen included in the trials. The effects of dexamethasone associated with a more frequent injection regimen were not evaluated in the MEAD trials and therefore remain unclear.