Flow diagram showing the numbers of studies included and excluded from the analysis and the reasons for exclusion. FDA indicates Food and Drug Administration; GSK, GlaxoSmithKline; and RCT, randomized controlled trial. *Includes pediatric studies, terminated early, or summary analysis.
Risk for myocardial infarction and cardiovascular mortality in trials classified by study duration and comparator drug. CI indicates confidence interval; and OR, odds ratio. *Calculated by proportion of the total sample included in the meta-analysis. †The RECORD (Rosiglitazone Evaluated for Cardiac Outcomes and Regulation of Glycemia in Diabetes) trial4 was not included in the analysis of the comparator subgroup because of the combination treatment assignment, and ADOPT (A Diabetes Outcome Progression Trial)19 was included separately for both the metformin (n = 1454) and the sulfonylurea (n = 1441) subgroups, according to randomized assignment.
Alternative analysis of risk for myocardial infarction and cardiovascular mortality, including studies with no events. ADOPT indicates A Diabetes Outcome Progression Trial19; CI, confidence interval; DREAM, Diabetes Reduction Assessment With Ramipril and Rosiglitazone Medication18; OR, odds ratio; and RECORD, Rosiglitazone Evaluated for Cardiac Outcomes and Regulation of Glycemia in Diabetes.4 *Calculated by proportion of the total sample included in the meta-analysis. †Mantel-Haenszel method.
Nissen SE, Wolski K. Rosiglitazone RevisitedAn Updated Meta-analysis of Risk for Myocardial Infarction and Cardiovascular Mortality. Arch Intern Med. 2010;170(14):1191-1201. doi:10.1001/archinternmed.2010.207
Controversy regarding the effects of rosiglitazone therapy on myocardial infarction (MI) and cardiovascular (CV) mortality persists 3 years after a meta-analysis initially raised concerns about the use of this drug.
To systematically review the effects of rosiglitazone therapy on MI and mortality (CV and all-cause).
We searched MEDLINE, the Web site of the Food and Drug Administration, and the GlaxoSmithKline clinical trials registry for trials published through February 2010.
The study included all randomized controlled trials of rosiglitazone at least 24 weeks in duration that reported CV adverse events.
Odds ratios (ORs) for MI and mortality were estimated using a fixed-effects meta-analysis of 56 trials, which included 35 531 patients: 19 509 who received rosiglitazone and 16 022 who received control therapy.
Rosiglitazone therapy significantly increased the risk of MI (OR, 1.28; 95% confidence interval [CI], 1.02-1.63; P = .04) but not CV mortality (OR, 1.03; 95% CI, 0.78-1.36; P = .86). Exclusion of the RECORD (Rosiglitazone Evaluated for Cardiac Outcomes and Regulation of Glycemia in Diabetes) trial yielded similar results but with more elevated estimates of the OR for MI (OR, 1.39; 95% CI, 1.02-1.89; P = .04) and CV mortality (OR, 1.46; 95% CI, 0.92-2.33; P = .11). An alternative analysis pooling trials according to allocation ratios allowed inclusion of studies with no events, yielding similar results for MI (OR, 1.28; 95% CI, 1.01-1.62; P = .04) and CV mortality (OR 0.99; 95% CI, 0.75-1.32; P = .96).
Eleven years after the introduction of rosiglitazone, the totality of randomized clinical trials continue to demonstrate increased risk for MI although not for CV or all-cause mortality. The current findings suggest an unfavorable benefit to risk ratio for rosiglitazone.
Controversy regarding the cardiovascular safety of the diabetes drug rosiglitazone arose in 2007 after the publication of a meta-analysis that demonstrated a significantly elevated risk for myocardial infarction (MI) and a borderline significant increased risk for cardiovascular (CV) mortality.1 The debate over the CV safety of rosiglitazone therapy has continued during the past 3 years, recently receiving renewed attention after the release of a report from the US Senate Committee on Finance that provided additional details about internal analyses conducted by the maker of rosiglitazone, GlaxoSmithKline (GSK), and the US Food and Drug Administration (FDA).2,3 No large, definitive CV outcomes trials have been conducted with rosiglitazone, although an open-label, noninferiority trial reported results in 2009 (RECORD [Rosiglitazone Evaluated for Cardiac Outcomes and Regulation of Glycemia in Diabetes]).4 That study was limited by low event rates, which resulted in insufficient statistical power to confirm or refute evidence of an increased risk for ischemic myocardial events.4- 8
Rosiglitazone was initially approved in 1999 to treat hyperglycemia in patients with type 2 diabetes mellitus. The standards for approval of diabetes drugs during that era required only demonstration of the reduction of hemoglobin A1c levels in trials of moderate duration (typically 24-52 weeks) in the absence of any apparent safety concerns. Controversy emerged soon after the drug's introduction, when reports first suggested that the use of rosiglitazone and a related drug, pioglitazone, could precipitate congestive heart failure in susceptible individuals.9 Then, in 2003, the Uppsala Drug Monitoring Group of World Health Organization alerted GSK about an unusually large number of spontaneous reports associating rosiglitazone use with congestive heart failure and ischemic myocardial events.2 GlaxoSmithKline subsequently conducted 2 internal patient-level meta-analyses of CV risk (the first completed in 2005 and the second in 2006), which were eventually published in 2008.10 Both analyses showed a significantly elevated risk for ischemic myocardial events. These analyses were supplied to the FDA and posted on the company's clinical trial registry, but neither GSK nor the FDA made any public statement about the findings.
As a consequence of a court settlement in 2004 with the State of New York, GSK was required to post clinical trial results on a public Web site.11 Using this data source, we performed a study-level meta-analysis that demonstrated a statistically significant 43% increase in the risk for MI and a borderline significant 64% increase in the risk for CV mortality.1 Subsequently, meta-analyses by other authors were published, some of which confirmed these findings, while others reported inconclusive results.12,13 The most comprehensive patient-level analysis was conducted by the FDA and confirmed an odds ratio (OR) of 1.38 for myocardial ischemic events (95% confidence interval [CI], 1.1-1.8; P = .02).14
An FDA Advisory Committee in 2007 concluded that the use of rosiglitazone increased the risk of MI but did not recommend removing the drug from the market.15 Subsequently, a consensus algorithm published by the American Diabetes Association and the European Association for the Study of Diabetes unanimously recommended against the use of rosiglitazone, but the drug continues to be marketed, with annual sales exceeding $1 billion in 2009.16 In the absence of definitive trials, a comprehensive meta-analysis remains the most appropriate means to evaluate the overall CV safety of rosiglitazone. Our objective was to update the 2007 meta-analysis that initiated the concerns about rosiglitazone therapy, using similar methods to the original study but also using alternative analyses to enable inclusion of trials with no CV events.
The studies included in the meta-analysis are listed in Table 1. The prespecified criteria for inclusion of trials required that studies have a randomized comparator group, a similar duration of treatment in all study groups, and more than 24 weeks of drug exposure. To identify potential studies for inclusion in the meta-analysis, we initially screened the clinical trials registry established by the manufacturer of rosiglitazone, GSK.20 Because this registry was mandated by a legal settlement,11 we assumed this registry to be complete but further searched for missing trials via MEDLINE and the FDA Web site. We screened 202 clinical trials for inclusion in the meta-analysis, excluding 146 for reasons summarized in Figure 1. The remaining 56 trials met all of the prespecified criteria for inclusion. Fifteen of the 56 trials did not report any MIs, and 30 of the trials did not report any CV mortality. The trials without events were not included in the primary prespecified analysis but were included in an alternative analysis.
The 56 trials included in the meta-analysis randomized 35 531 patients: 19 509 assigned to receive rosiglitazone and 16 022 assigned to comparator groups. Three groups of trials were identified. In 1999, GSK submitted 1 group of 5 studies to the FDA for presentation to the Advisory Committee, which recommended approval of rosiglitazone use.18 In these 5 trials, 1967 patients were randomly assigned to receive rosiglitazone and 793 patients received comparator drugs or placebo. A second group of 48 trials were primarily identified from the GSK clinical trials registry. In these 48 trials, 11 231 patients were randomly assigned to receive rosiglitazone and 7473 received comparator drugs or placebo. The third group of trials included 3 larger prospective randomized trials that were published in major medical journals.4,18,19 In these 3 trials, 6311 patients were randomly assigned to receive rosiglitazone and 7756 patients received comparator drugs. Four of the 5 registration trials and all 3 large trials were published, but most (35 of 48) trials identified via the GSK registry remain unpublished.
For the 5 studies included in the original submission for FDA approval, study-level data were extracted from publicly available briefing documents that were downloaded from the FDA Web site.21 Data from these same trials were also reported in a summary fashion on the Clinical Trial Registry Web site maintained by GSK.20 For the 4 registration trials that were published in peer-reviewed medical journals, we obtained adverse event data from data tables in the manuscripts.22- 25 These 3 sources (FDA documents, GSK Web site, and publications) were cross-checked for consistency. In cases of disagreement between published and unpublished data, data derived from the manufacturer's Web site were used. For the 3 larger outcome trials, study-level data were extracted from data tables contained in the published manuscripts.4,18,19 For the remaining group of 48 trials that were available primarily on the GSK clinical trials registry, data on reported adverse outcomes were extracted from the study summaries included in the posted individual trial reports.
In cases in which several treatment groups received rosiglitazone within a single trial, the rosiglitazone-exposed groups were pooled together for analysis. For each study, the control group was defined as patients receiving any drug regimen other than rosiglitazone, including placebo. With the exception of the DREAM (Diabetes Reduction Assessment With Ramipril and Rosiglitazone Medication)18 and RECORD trials, the included studies did not centrally adjudicate either MI or mortality. We reviewed data summaries provided in the FDA review documents, the GSK clinical-trial registry Web site, and published trial results and then abstracted data from the adverse event tabulations for MI and mortality (CV and all-cause). Because we did not have access to individual patient data, time-to-event analyses for adverse events could not be performed, which precluded the calculation of hazard ratios (HRs). Because only summary data tables were available, it was not possible to determine whether the same patient had both an MI and CV mortality, which precluded calculation of outcomes based on the composite of MI or CV death. Accordingly, the 3 outcomes (MI, CV death, and all-cause mortality) are reported separately.
Because most of the included trials had few events, we prespecified the use of the Peto method to calculate ORs and 95% CIs.26- 28 This approach excludes trials with no adverse events, since an OR cannot be calculated for such studies. To account for trials with no events, we also determined ORs and 95% CI's using an alternative method in which smaller studies were grouped by randomization ratios and larger trials were considered individually. This alternative analysis used the Mantel-Haenszel method to calculate ORs and 95% CIs.29 With this approach, all 56 trials meeting inclusion criteria were included in the alternative analysis. All reported P values are 2-sided. Statistical heterogeneity across the 56 trials was tested using the Cochran Q statistic. P >.10 was considered evidence of a lack of heterogeneity. This analysis revealed no heterogeneity, allowing use of a fixed-effects model. Additional analyses classified by trial duration (shorter or longer than 1 year) and type of comparator drug (insulin, metformin, sulfonylurea, or placebo) were assessed using the Peto method in a similar fashion to the overall analyses. The number needed to harm (NNH) was estimated based on the annualized rate of MI that was observed in the ACCORD (Action to Control Cardiovascular Risk in Diabetes) trial,5 normalizing this rate to the 5-year duration typically used in NNH calculations. Data were analyzed with Comprehensive Meta-analysis software (Version 2.2; Biostat, Englewood, New Jersey).
The dosages of study drugs, baseline demographic characteristics, study periods, and hemoglobin A1c levels are listed in Table 2.
. The mean age of patients was approximately 57 years; 55% were male; and more than 80% were white. The mean hemoglobin A1c level at baseline averaged 8.2%. Dosages of rosiglitazone varied from 2 to 8 mg/d, with most studies titrating patients to 8 mg/d during the course of the study. Although nearly all trials enrolled patients with type 2 diabetes mellitus, a few studies included patients treated for investigational indications, including psoriasis, Alzheimer disease, rheumatoid arthritis, or multiple sclerosis. A single study, the DREAM trial, was designed as a diabetes prevention study to assess whether rosiglitazone therapy could prevent the new onset of type 2 diabetes in high-risk individuals
The individual trials contributing to the meta-analysis and the numbers of MIs and CV deaths are listed in Table 3. For MI, the Cochran Q statistic, including the RECORD trial, was 30.3 (P = .87; I2 = 0%). Excluding the RECORD trial, the Q statistic for MI was 29.7 (P = .86; I2 = 0%. For CV mortality, including the RECORD Trial, the Q statistic was 16.2 (P = .91; I2 = 0%). Excluding the RECORD Trial, the Q statistic was 12.8 (P = .97; I2 = 0%).
Table 4 shows the results of the primary prespecified analyses, and Figure 2 lists the results in subgroups classified by trial duration or comparator drug. For the 41 trials with at least 1 MI, the Peto OR for rosiglitazone relative to comparator drugs was 1.28 (95% CI, 1.02-1.63; P = .04). Excluding the RECORD trial from the analysis yielded similar results, but with a higher OR (OR, 1.39; 95% CI, 1.02-1.89; P = .04). The results were similar in subgroups that were classified by trial duration. For trials of less than 12 months' duration, the OR for MI was 1.76 (95% CI, 0.93-3.33). For trials that were longer than 12 months, the OR for MI was 1.22 (95% CI, 0.95-1.57). All subgroups that were classified by comparator drug(s) show elevated estimates for the OR for MI, ranging from 1.26 to 3.49, but with wide confidence intervals (Figure 2).
For the 26 trials with at least 1 event, the Peto OR for CV mortality for rosiglitazone relative to comparator drugs was 1.03 (95% CI, 0.78-1.36; P = .86) (Table 4). However, excluding the RECORD trial had a large effect on the results. For the 25 trials with at least 1 CV death, excluding the RECORD trial, the OR was 1.46 (95% CI, 0.92-2.33; P = .11). For trials with a duration of less than 12 months, the estimated OR was 2.32 (95% CI, 0.93-5.78). For trials with a duration of longer than 12 months, the OR was 0.94 (95% CI, 0.70-1.27) (Figure 2). All subgroups that were classified by comparator drug(s) show elevated estimates for the OR for CV mortality, ranging from 1.13 to 2.02, but with wide confidence intervals (Figure 2).
To permit inclusion of studies with no CV events, an alternative analysis was performed in which smaller studies were pooled by randomization ratios. The pooling of smaller studies allows inclusion of all 56 trials meeting the prespecified inclusion criteria, regardless of the presence or absence of adverse events. Figure 3 shows Mantel-Haenszel ORs for the pooled smaller studies classified by randomization ratio, with the 3 larger trials considered individually. The results of this alternative analysis are very similar to the primary results obtained using the Peto method. For MI, the OR was 1.28 (95% CI, 1.01-1.62), including the RECORD trial, and 1.38 (95% CI, 1.01-1.87), excluding the RECORD trial. For CV mortality, the OR, including the RECORD trial, was 0.99 (95% CI, 0.75-1.32), and the OR was 1.36 (95% CI, 0.84-2.21), excluding the RECORD trial.
Substituting all-cause mortality for CV death in the primary analysis (Peto method) yielded similar results: the OR was 0.98 (95% CI, 0.82-1.17), including the RECORD trial, and 1.16 (95% CI 0.89-1.52), excluding the RECORD trial. For NNH calculations, the background rate for MI based on the ACCORD trial was 1.38% per year, or 6.9% over the 5-year period typically used in NNH estimates. Applying the odds of increased risk for MI with rosiglitazone from the current study, the NNH is estimated to be 1 additional MI per 52 patients (assuming OR = 1.28 after inclusion of the RECORD trial) or 1 additional MI per 37 patients, (assuming OR = 1.39 after exclusion of the RECORD trial) treated with rosiglitazone for 5 years.
The CV safety of rosiglitazone therapy has remained controversial after the publication of initial reports that suggested that the use of rosiglitazone increases the risk for MI and other ischemic myocardial events.1,14,20 Despite 11 years on the market, rosiglitazone has not been studied in any definitive randomized controlled CV outcomes trials. Accordingly, a meta-analysis of existing clinical trials represents the most robust available approach to determining the CV safety of this drug. The current study analyzed MI and CV mortality for 56 randomized trials involving 35 531 patients. Using a study-level meta-analysis, the OR for MI was significantly increased but without evidence of an increase in CV or all-cause mortality. An alternative analysis that included trials with no CV events found a similar hazard (Figure 3). Subgroups classified by study duration and comparator drug also showed elevated OR estimates (Figure 2). These findings are consistent with prior meta-analyses conducted by GSK, the FDA, and most independent investigators demonstrating an increased risk of MI in patients treated with rosiglitazone.1,10,14,20 The FDA has announced that it will conduct an advisory committee meeting in July 2010 to consider whether to remove rosiglitazone from the market.
The public health implications of these results are considerable. There are more than 23 million persons with diabetes in the United States alone and nearly 300 million worldwide.30,31 Cardiovascular disease is the leading cause of death in patients with type 2 diabetes, representing approximately 68% of all causes of mortality.30 The estimated 28% to 39% increase in the risk for MI observed for rosiglitazone use in the current analysis and the NNH of 52 or 37 (with and without the RECORD trial) represent a significant potential health burden. The magnitude of the observed effect is larger than might be anticipated in a safety analysis using intent-to-treat (ITT) methods. In ITT efficacy studies, discontinuation of therapy or crossovers between treatment groups bias the study toward the null hypothesis, thereby favoring the control treatment. However, in safety studies, similar flaws in study conduct bias the investigation toward a relative risk of 1.0, providing the potential for a false declaration of safety. Accordingly, using standardized ITT methods, it is statistically much more difficult to conclude that a therapy is unsafe than to demonstrate efficacy. Because we did not have access to patient-level data, we were unable to perform a useful alternative analysis that is commonly used in drug-safety studies, a “per protocol” approach that includes events that occurred “on-treatment” or within 30 days after discontinuation of treatment.
We elected to present analyses with and without the RECORD trial. Several of the concerns about the RECORD trial have been reported elsewhere.5- 8 The study was an open-label, randomized noninferiority trial that compared rosiglitazone with metformin or sulfonlyurea. The primary efficacy parameter was unconventional, CV hospitalization or death. The study postulated an annual event rate of 11% but observed an event rate of only 2.6%, a large mismatch that substantially reduced statistical power. The MI rate for the control group in the RECORD trial was 0.52% per year compared with 1.38% for a similar population in the ACCORD trial, raising the concern that MIs may have been incompletely ascertained. By the end of the trial, 40% of patients randomized to rosiglitazone therapy were no longer taking the drug. Nonadherence to randomized therapy represents an important issue in a safety trial because, as noted above, dropouts and crossovers bias the result toward the null hypothesis. Finally, the company compromised data integrity by publishing an unplanned interim analysis32 and appears to have had access to ongoing study data at a time when the trial should have remained blinded.2,6,7
The limitations of our meta-analysis are notable. We had access to study-level data that were diclosed as a result of a court settlement and subsequently posted on a company Web site. The unavailability of patient-level data precluded a more statistically powerful analysis using time-to-event methods. However, it should be noted that the original 2007 meta-analysis was subsequently replicated by the FDA using time-to-event data, resulting a nearly identical relative risk. There are important strengths to the study. The number of patients and studies included in the analysis is substantially larger than was available for our original meta-analysis, which was completed in 2007. Furthermore, because disclosure of all clinical trials by the maker of rosiglitazone was mandated by a court order, the common problem of publication bias did not confound our analyses. The original 2007 analysis was criticized by some authors because it did not include clinical trials in which there were no events.13 Therefore, in the current effort, we provided an alternative approach that includes all 56 trials, regardless of whether there were adverse events. With both methods, the OR was nearly identical.
A related issue involves the question of whether use of the other marketed thiazolidinedione, pioglitazone, carries similar risks. A large CV outcomes trial with pioglita zone, the PROACTIVE (Prospective Pioglitazone Clinical Trial in Macrovascular Events) trial, which was published in 2005, did not show statistically significant benefits.33 It showed a trend toward reduction in the primary efficacy parameter, a broad composite of CV events (HR, 0.90; 95% CI, 0.80-1.02; P = .10). However, a prespecified secondary end point of death, MI, and stroke showed a benefit (HR, 0.84; 95% CI, 0.72-0.98; P = .03). A patient-level meta-analysis of CV outcomes with pioglitazone analyzed 19 trials, with a total enrollment of 16 390 patients, and showed a statistically significant benefit on the composite of death, MI, and stroke (OR, 0.82; 95% CI, 0.72-0.94; P = .005). These findings effectively rule out a CV hazard for pioglitazone use and suggest the possibility of a CV benefit. However, it must be noted that the use of both rosiglitazone and pioglitazone has been associated with an increased risk of congestive heart failure.
The potential mechanism(s) for CV harm from rosiglitazone use (and the differences from pioglitazone use) remains to be elucidated, but there are several reasonable hypotheses. Rosiglitazone therapy increased low-density lipoprotein cholesterol (LDL-C) levels as much as 23% in trials, leading to approval.21 Current FDA guidelines consider a drug that lowers LDL-C levels by at least 15% “approvable” for presumed CV benefits. Although the FDA has not established a level of increase in LDL-C that is presumed to cause harm, a drug that increases LDL-C levels would reasonably be expected to increase CV adverse events. Interestingly, the lipid effects of the 2 marketed thiazolidinediones, pioglitazone and rosiglitazone, are markedly different.34 In a comparative efficacy trial, rosiglitazone therapy produced greater increases in LDL-C levels and raised triglyceride levels, while pioglitazone therapy reduced triglyceride levels. Pioglitazone therapy also produced significantly greater increases in high-density lipoprotein cholesterol compared with rosiglitazone therapy. Thiazolidinediones are nuclear receptor agonists that modulate expression of a large number of genes. There are major differences in the pattern of gene modulation for pioglitazone vs rosiglitazone.35 Rosiglitazone activates a gene associated with production of matrix metalloproteinase 3, an enzyme linked to plaque rupture.36
There are also implications of these findings on the traditional approach used by regulatory authorities to approve drugs that are used to treat diabetes. Historically, evidence of a glucose-lowering effect, with no evidence for obvious safety issues, was sufficient for approval. In the wake of the rosiglitazone controversy, the FDA has mandated that sponsors of all new diabetes drugs perform CV outcomes studies sufficient to rule out an HR with an upper 95% CI of 1.8 before approval and 1.3 after approval.37 Had such requirements been in place at the time rosiglitazone was developed, it seems likely that the drug would never have been approved.
The results of the current meta-analysis suggest an unfavorable benefit to risk ratio for rosiglitazone use. The implications of this finding warrant further discussion. Even a modest increase in the risk of MI in a diabetic population would have serious consequences. Reviewers within the FDA Office of Surveillance and Epidemiology calculated the number of major CV events potentially attributable to rosiglitazone therapy from 1999 to 2006, reporting a range from 41 000 to 205 000.2 More recently, using lower estimates of the rate of drug use after the 2007 controversy, FDA reviewers have calculated the number of excess MIs (6000 annually) potentially attributable to rosiglitazone use relative to treatment with the alternative thiazolidinedione, pioglitazone.3 Although hyperglycemia has been associated with an increased risk of microvascular adverse events, there are now 12 classes of drugs that are approved to lower blood glucose levels, including insulin. Because no unique benefits of rosiglitazone use have been identified, administration of this agent solely to lower blood glucose levels is difficult to justify.
Correspondence: Steven E. Nissen, MD, Department of Cardiovascular Medicine, The Cleveland Clinic Foundation, 9500 Euclid Ave, Cleveland, OH 44195 (firstname.lastname@example.org).
Accepted for Publication: May 10, 2010.
Published Online: June 28, 2010. doi:10.1001/archinternmed.2010.207
Author Contributions:Study concept and design: Nissen and Wolski. Acquisition of data: Nissen and Wolski. Analysis and interpretation of data: Nissen and Wolski. Drafting of the manuscript: Nissen and Wolski. Critical revision of the manuscript for important intellectual content: Nissen and Wolski. Administrative, technical, and material support: Nissen and Wolski. Study supervision: Nissen.
Financial Disclosure: Dr Nissen has received research support from AstraZeneca, Atherogenics, Eli Lilly, Novartis, Pfizer, Resverlogix, Takeda, Daiichi-Sankyo, and Sanofi-Aventis through The Cleveland Clinic Center for Clinical Research (C5) within the last 5 years. He has consulted for a number of pharmaceutical companies without financial compensation. All honoraria, consulting fees, or any other payments from any for-profit entity are paid directly to charity, so he receives neither income nor tax deduction