Cardiovascular Risk and Inhibition of Cyclooxygenase:  A Systematic Review of the Observational Studies of Selective and Nonselective Inhibitors of Cyclooxygenase 2 FREE

In the last 5 years, interest in the cardiovascular effects of the relatively selective inhibitors of cyclooxygenase 2 (COX-2) has been intense. In October 2004, rofecoxib was withdrawn from world markets after a randomized placebo-controlled trial found that in doses of 25 mg/d, it increased rates of cardiovascular events in patients with colorectal polyps.¹ The results were confirmed by several large pharmacoepidemiological studies.^2- 4 Celecoxib continues to be widely used, despite meta-analyses of randomized controlled trials showing an increased risk of myocardial infarction.^5- 6

Attention has turned to the cardiovascular safety of the older nonselective nonsteroidal anti-inflammatory drugs (NSAIDs).These agents are used extensively and some are available in many countries without prescription. NSAIDs reversibly block both isoforms of cyclooxygenase but vary in their degree of selectivity.⁷ In one trial, it was suggested that the apparent excess cardiovascular risk with rofecoxib may be explained by a “cardioprotective” effect of the comparator drug, naproxen.⁸ However, the results of another trial suggested that naproxen may increase the risk of myocardial infarction,⁹ and a recently published meta-analysis of randomized trials has implicated high doses of ibuprofen and diclofenac.⁶

Regulatory authorities have provided variable advice regarding the safety of NSAIDs. In the United States, the US Food and Drug Administration (FDA) requires that both selective COX-2 inhibitors and NSAIDs carry a warning highlighting the potential for increased risk of cardiovascular events.¹⁰ In contrast, the European Medicines Agency has required labeling of selective COX-2 inhibitors, but made no recommendation about the cardiovascular safety of the older NSAIDs.^11- 12

Meta-analyses of randomized placebo-controlled trials have provided information about cardiovascular risks with selective COX-2 inhibitors but in total, these have captured only 340 cardiovascular events in users of the drugs.⁶ A full evaluation of data on cardiovascular risks with these drugs requires an examination of controlled pharmaco-epidemiological studies.

Studies were eligible for inclusion if they were controlled (case-control or cohort design), and reported on cardiovascular risks associated with the use in population settings of selective COX-2 inhibitors, reported on cardiovascular risks associated with the use in population settings of conventional NSAIDs, with nonuse/remote exposure as the reference exposure for calculation of the relative risk (RR).

We searched electronic databases from 1985 until January 2006. These included MEDLINE, EMBASE, Cochrane Library, Google Scholar, epidemiological research Web sites, abstracts of scientific meetings, and bibliographies of relevant studies. The search terms were compiled from the names of individual drugs, the therapeutic class, mode of activity, cardiovascular and cerebrovascular outcome terms, and study design terms. We also searched on authors' names. Titles and abstracts of articles identified by the searches were reviewed by the authors. Searches were repeated using additional search terms identified from articles considered relevant to the review.

Quality assessment and data extraction were performed in duplicate with resolution of any discrepancies by consensus. Methodologic quality was assessed using an established instrument.¹³ The instrument assessed case-control studies in terms of methods of selection of cases and controls, comparability of cases and controls, and ascertainment of exposure to the agent of interest. Cohort studies were assessed in terms of selection of the exposed and nonexposed cohorts, comparability of the cohorts, and outcomes ascertainment.

All but 3 of the studies were conducted using linkage of large electronic databases or electronic medical records.^14- 16 Such studies use prescribing or dispensing as a proxy for drug consumption. All eligible cases occurring within the time frame were available for inclusion, and controls were selected randomly from the source populations; drug exposure and outcomes were recorded in real time, meaning these studies should have been free of the selection and recall biases that affect ad hoc case-control studies. Misclassification of clinical outcomes and comorbid states, and incomplete information regarding consumption of nonprescription drugs, alcohol, and tobacco are potential problems with database studies, but are equally likely whether data are analyzed prospectively or retrospectively. For these reasons we felt it appropriate to combine data from case-control and cohort analyses in order to increase the precision of our estimates and improve our ability to discriminate between individual drugs. The outcome under study is relatively uncommon, meaning the odds ratio is an accurate estimate of the RR. Many studies controlled for cohort enrollment date and applied the same index day to cases and controls. As this provides a control for calendar time, we felt it was justified to summarize odds ratios derived from the case-control analyses and hazard or rate ratios extracted from the cohort analyses. Point estimates and standard errors were extracted from each study and were combined by the random-effects model of DerSimonian and Laird using Cochrane Review Manager software (The Cochrane Collaboration, Oxford, England).

The searches returned 7086 potentially relevant articles. After review of the titles and removal of duplicates, 745 abstracts were read and 233 articles were selected for further evaluation (Figure 1). After application of inclusion criteria, 23 studies were eligible for inclusion (Table 1, Table 2).

Seventeen case-control analyses included 86 193 cases with cardiovascular events^{2- 4,14- 16,23- 33} (almost exclusively myocardial infarction or sudden cardiovascular death, Table 1), and at least 528 000 controls (1 study did not report the number of controls²⁴). All case-control analyses reported risks with nonselective NSAIDs and 9 also reported risks with selective COX-2 inhibitors. Two studies by Kimmel et al^15- 16 used the same population and had overlapping time frames, reporting on NSAIDs in the early study,¹⁵ and NSAIDs and COX-2 inhibitors in the later analysis.¹⁶ The risk estimates from the latter were used in the data summaries except for ibuprofen, which was reported only in the earlier study.¹⁵ Studies reporting on selective COX-2 inhibitors focused on celecoxib, rofecoxib, or meloxicam. Studies that provided data on individual NSAIDs reported mainly on ibuprofen, diclofenac, naproxen, indomethacin, and piroxicam.

Six studies were based on cohort analyses and included 75 520 users of selective COX-2 inhibitors,^17- 22 375 619 users of nonselective NSAIDs, and 594 720 unexposed participants (Table 2). The main outcomes reported were acute myocardial infarction and sudden cardiovascular death.

Most exclusions were because reports did not provide information on the study outcomes or the drugs of interest (Figure 1). Four studies were reported more than once and the most comprehensive analysis was used as the data source.^18,26- 28 As a consequence, 4 reports were excluded from the analyses but where these provided additional information relevant to the review, were included.^34- 37 Two additional studies were excluded; 1 compared parenteral ketorolac with opioids in acutely ill hospitalized patients³⁸; the second was the subject of allegations of scientific misconduct that remained unresolved at the time of compiling this report.³⁹

Eleven studies reported on first cardiovascular events^{2,15- 16,25- 27,29- 33} (Table 1, Table 2), 9 included both first and recurrent events (any event),^{3- 4,14,20- 24,28} 2 reported only on death following an event,^18- 19 and 1 on death or recurrent myocardial infarction following an event.¹⁷ Thirteen studies indicated that both fatal and nonfatal cardiovascular events were included,^{2- 4,17,21- 24,26,28,30,32- 33} 3 reported nonfatal events only^14- 16; the remaining studies did not specify whether both fatal and nonfatal events were included.

Study populations varied in their age and sex distributions (Table 1, Table 2). In all but 3 studies, information on the drugs of interest, hospitalization diagnoses, cardiovascular risk factors, and the presence of comorbid states was determined from linked health care databases or electronic medical records that included hospital discharge diagnoses, general practice encounters, and dispensing of prescriptions. The remaining studies collected information through face-to-face¹⁴ or telephone-based structured interviews.^15- 16

All studies reported adjusted estimates of risk for cardiovascular events. Factors adjusted for included age, sex, cardiovascular risk factors, and use of cardiac medications. Many studies also adjusted for comorbidity (Table 1, Table 2). Fifteen studies adjusted for prescribed aspirin use,^{2,14- 16,20- 23,25- 29,31,33} 2 did not report on aspirin adjustment,^3- 4 5 excluded users of prescribed aspirin or reported on a subgroup of non-users,^{15- 16,26,30,32} 7 could report on a subgroup of aspirin users or investigated only users of prescribed aspirin,^{15- 16,18- 19,26- 28} and 1 study assumed all participants were aspirin users.¹⁷ Adjustments for smoking, alcohol intake, and body mass index (calculated as weight in kilograms divided by height in meters squared) were undertaken variably across the studies. Studies that relied on an interview included information on over-the-counter drug use, including NSAIDs and aspirin, smoking, and alcohol intake.^14- 16

Five case-control studies used the General Practice Research Database in Great Britain.^{27- 28,31- 33} Two large studies spanned the period 1992 to 2000 and apparently overlapped by 10 months.^28,31 The 3 remaining studies overlapped variably in the time periods studied.^27,32- 33 Two studies evaluated risk in specific patient groups (patients with rheumatoid arthritis,³² women³³). To allow for the effects of double counting, data summaries were made omitting studies in which the exposure periods overlapped with those of the 2 largest studies.

We selected studies that had non-use or remote use of the drugs of interest as the reference exposure category. Seven studies also reported RRs of cardiovascular events estimated from between-drug agent comparisons.^{3- 4,16,23- 24,27,32} Where possible, we also extracted and summarized data from these analyses.

Using the quality assessment instrument, case-control studies could score a maximum of 4 points in the selection and exposure categories and 2 points in the comparability category. Studies scored consistently well across categories (7-8 points in total from a possible 9), the exceptions being those reported only in abstract form^23- 24 (scoring 4/9 and 5/9, respectively). Cohort studies could score a maximum of 4 points in the selection and outcomes assessment categories and 2 for cohort comparability. All studies scored well (7-8 points).

The individual study values and summary estimates of RRs for the comparisons of COX-2 inhibitors and NSAIDs with remote or nonuse of anti-inflammatory drugs are presented in Table 3 and Table 4. The summary RR estimates from the case-control and cohort analyses were similar (Table 3, Table 4). Between-study heterogeneity in RR estimates was statistically significant, but quantitatively modest (Figure 2, Figure 3).

Rofecoxib was included in 9 case-control and 2 cohort studies (Table 3). The summary RRs for cardiovascular events were 1.31 (95% confidence interval [CI], 1.18-1.46) for the case-control studies and 1.53 (95% CI, 0.68-3.44) for the cohort studies. Combining across all studies, the summary RR was 1.35 (95% CI, 1.15-1.59) (Figure 2). A dose effect was apparent: the summary RR with doses in excess of 25 mg/d was 2.19 (95% CI, 1.64-2.91) compared with 1.33 (95% CI, 1.00-1.79) with 25 mg/d or less (Table 3).

Eight case-control and 3 cohort studies reported on celecoxib (Table 3). Celecoxib exposure did not lead to an elevation of the risk of cardiovascular events: summary RR 1.01 (95% CI, 0.90-1.13) for the case-control studies and 1.22 (95% CI, 0.69-2.16) for the cohort studies. Combining across all studies, the summary RR was 1.06 (95% CI, 0.91-1.23). Three studies provided dose-stratified RR estimates for celecoxib.^17,21,26 There were insufficient data to enable stable estimates of the effects of different doses of celecoxib.

Only 3 case-control studies provided data on meloxicam, of which 1 reported an elevation in vascular risk (Table 3). The summary RR was 1.25 (95% CI, 1.00-1.55).

Sixteen studies reported on ibuprofen and/or naproxen individually, 9 on diclofenac, 6 on indomethacin, and 4 provided data on piroxicam (Table 4). The summary RR with naproxen was close to 1 at 0.97 (95% CI, 0.87-1.07). There was no significant elevation in risk with use of ibuprofen or piroxicam at summary RRs of 1.07 (95% CI, 0.97-1.18) and 1.06 (95% CI, 0.70-1.59), respectively. Diclofenac and indomethacin were associated with increased risks of cardiovascular events. The summary RR for diclofenac from 9 studies was 1.40 (95% CI, 1.16-1.70). In all but 1 of the studies,¹⁹ the point estimate for the risk with diclofenac exceeded the risks simultaneously measured for ibuprofen and naproxen. Indomethacin was included in 6 studies and the summary RR was 1.30 (95% CI, 1.07-1.60).

To minimize the effects of between-study heterogeneity, we undertook a series of “pairwise” comparisons, summarizing only data from studies that simultaneously evaluated both a selective COX-2 inhibitor and a nonselective NSAID, with nonuse or remote use as the reference exposure. The RRs were as follows: rofecoxib, 1.37 (95% CI, 1.04-1.79) and ibuprofen, 1.05 (95% CI, 0.88-1.25) in 6 studies; rofecoxib, 1.32 (95% CI, 1.18-1.48) and naproxen, 1.11 (95% CI, 0.96-1.28) in 7 studies; celecoxib, 1.07 (95% CI, 0.82-1.38) and ibuprofen, 1.03 (95% CI, 0.88-1.21) in 7 studies; celecoxib, 1.00 (95% CI, 0.89-1.12) and naproxen, 1.07 (95% CI, 0.93-1.22) in 7 studies. Three studies provided data on celecoxib, rofecoxib, and diclofenac; the summary RRs were: celecoxib, 1.28 (95% CI, 0.74-2.22); rofecoxib, 1.61 (95% CI, 1.07-2.43); and diclofenac, 1.80 (95% CI, 1.34-2.40). Two studies evaluated celecoxib, rofecoxib, and indomethacin; the summary RRs were: celecoxib, 0.98 (95% CI, 0.76-1.26); rofecoxib, 1.32 (95% CI, 1.23-1.42); indomethacin, 1.48 (95% CI, 1.13-1.94).

A number of studies provided analyses of risk where the reference exposure was another drug rather than nonuse. Compared with any nonselective NSAID, the summary RR for rofecoxib was 1.21 (95% CI, 1.13-1.29) and for celecoxib it was 0.95 (95% CI, 0.85-1.05). With naproxen as the reference, the summary RR for rofecoxib was 1.68 (95% CI, 1.22-2.32) and with celecoxib it was 0.94 (95% CI, 0.74-1.18). An elevation in risk was evident with rofecoxib when it was compared directly with celecoxib: summary RR, 1.34 (95% CI, 1.14-1.56). When compared directly with other NSAIDs (including diclofenac, indomethacin, or both) naproxen was associated with a lower risk of cardiovascular complications: summary RR, 0.75 (95% CI, 0.63-0.88).

Three studies evaluated vascular risk among new users of COX-2 inhibitors, with remote or nonuse as the reference^21,25- 26 and 1 study compared new users of rofecoxib directly with new users of celecoxib.⁴ The exposure category was the first 30 days of treatment. An early risk was evident with rofecoxib (summary RR, 1.66; 95% CI, 1.09-2.51) but not with celcoxib, 1.32 (95% CI, 0.80-2.19). The data from the cohort study by Ray et al²¹ suggest an early risk with rofecoxib 25 mg/d or more. Comparing first-time users of celecoxib and rofecoxib, Solomon et al⁴ reported that relative to celecoxib, the odds ratio for acute myocardial infarction associated with first-time use of rofecoxib was 1.43 (95% CI, 1.12-1.83) for use in the first 30 days.

In an attempt to explore whether cardiovascular risk might vary with aspirin use, we attempted to extract data from studies of COX-2 inhibitors and NSAIDs in groups defined by aspirin exposure. Unfortunately, there were insufficient data to provide stable estimates of whether aspirin protects against the cardiovascular risk with rofecoxib or other risk-inducing drugs. In the case of ibuprofen, the summary RR in the absence of aspirin was 0.82 (95% CI, 0.50-1.32) in 3 studies, compared with 0.99 (95% CI, 0.75-1.31) in aspirin users in 5 studies.

Of the 5 studies that used the General Practice Research Database, all reported on NSAIDS^{27- 28,31- 33} and 1 also evaluated COX-2 inhibitors²⁸ (Table 1). Including in the pooled analyses only the data from the 2 least overlapping studies,^28,31 the summary RRs were little different from the overall estimates (Table 4) for naproxen, 0.98 (95% CI, 0.89-1.08); diclofenac, 1.35 (95% CI, 1.08-1.69), or ibuprofen, 1.08 (95% CI, 0.97-1.19).

The results of this systematic review of controlled observational studies allow conclusions to be drawn about the risks of cardiovascular events during treatment with selective and nonselective NSAIDs. The data confirm the elevated risk with rofecoxib and indicate that it is dose-related. In doses of around 200 mg/d, celecoxib was not associated with an increased risk, but the data did not exclude an increased risk with higher doses. Use of naproxen was not associated with any reduction in risk, as was suggested by the authors of a report of a large trial comparing it with rofecoxib.⁸ Of the other nonselective NSAIDs, the highest risk was seen with diclofenac. The increased cardiovascular risk with rofecoxib could be observed during the first 30 days of treatment. This conclusion is consistent with a recent re-analysis of the APPROVe trial of refecoxib, which contradicts the original suggestion that the vascular risk was only seen after 18 months of treatment.⁴⁰ The data on meloxicam did not allow any definite conclusions and the elevated summary RR was largely due to the results of a single study.

It is instructive to compare these data with the results of a recently published meta-analysis of randomized controlled clinical trials.⁶ Between them, these systematic reviews include most of the clinical epidemiological evidence on cardiovascular risk with anti-inflammatory drugs. In reviewing data from 138 randomized trials, Kearney et al⁶ estimated a summary RR for COX-2 inhibitors of 1.42 (95% CI, 1.13-1.78), close to the estimate we made for rofecoxib, 1.31 (95% CI, 1.18-1.46). The observational data gave a summary RR of 2.19 (95% CI, 1.64-2.91) with rofecoxib doses above 25 mg/d; but there were insufficient data from the randomized trials to make this calculation.^6,41 In contrast to the evidence from the randomized trials, we found no increase in risk with commonly used doses of celecoxib. However, the randomized data only showed an increased risk with daily celecoxib doses of 400 mg and above.^5- 6 There is agreement between the randomized and nonrandomized data that naproxen did not alter the risk of cardiovascular events.⁶ Of more concern is the evidence from both randomized and nonrandomized studies that diclofenac increases the risk of cardiovascular events: summary RR, 1.63 (95% CI, 1.12-2.37) in the case of randomized trials,⁶ and 1.40 (95% CI, 1.16-1.70) with the observational studies. We did not find an elevated RR with ibuprofen and the summary estimate from the randomized trials was not significantly different from 1.⁶ It has been suggested that ibuprofen may interact with low-dose aspirin,^19,42 but we found the RRs to be close to 1 in users and nonusers of aspirin. We found inconsistent effects of aspirin on the risks with other drugs. Significantly, Kearney et al⁶ found no evidence from the randomized trials that aspirin modifies the risk of cardiovascular events with COX-2 inhibitors.

The differences between rofecoxib and celeoxib appear important from both a clinical and regulatory standpoint. The data do not point to a safe dose level with rofecoxib, which justifies the decision taken to withdraw the drug from sale. At doses of 200 mg or less there is no convincing evidence of an increased risk of cardiovascular events with celecoxib, which remains on international markets. However, based on the randomized data celecoxib appears unsafe in doses of 400 mg or more.^5- 6 These results seem to point to different dose-effect gradients in the vascular compartment across the ranges of doses of celecoxib and rofecoxib that were used in clinical practice. This review does not allow a judgment about whether any claimed advantages of celecoxib outweigh the elevated cardiovascular risk seen with high doses.

Diclofenac is another relatively COX-2 selective drug that has been much less studied than either rofecoxib or celecoxib. Both the randomized and observational data point to an RR that is similar to what was seen with rofecoxib. Diclofenac is reported to have a similar degree of COX-2 selectivity to celecoxib.⁴³ The increased cardiovascular risk may indicate that diclofenac is commonly ingested in relatively high doses in relation to the drug's effects on COX-2 in the vascular compartment.

We found very few data on cardiovascular risk with meloxicam. This is significant because in some markets (eg, Australia), it has replaced rofecoxib and celecoxib following the publicity given to their adverse effects. The summary RR was 1.25 (upper 95% confidence limit 1.56), meaning that it may be no different from other relatively COX-2–selective drugs. Finally, we found an elevated risk of cardiovascular events with indomethacin. This is not easily explained from its pharmacology as it is not a selective COX-2 inhibitor.^7,43 The data reviewed here were sparse and indomethacin is seldom recommended because of gastrointestinal and central nervous system toxicity. This review provides an additional reason not to use indomethacin.

As with any systematic review, the limitations reflect those of the individual studies. Most relied on information from databases. While minimizing selection and recall biases, the definition of exposure relies on the recording of a drug being prescribed or dispensed rather than actually consumed, so misclassification is possible. An additional weakness was the inability in many studies to measure directly consumption of nonprescription aspirin and NSAIDs. Several authors acknowledged the latter but pointed out that the generally elderly populations studied would have had access to subsidized medicines and therefore, little incentive to purchase their own. Notwithstanding, it remains a concern that self-prescription with agents that might confound the results cannot be quantified. It is possible that unrecorded exposure to aspirin or anti-inflammatory drugs might account for some of the observed heterogeneity. Because of the reliance on stored data, which were not collected primarily for research, information on other potential confounders such as smoking, hypertension, and elevated cholesterol was incomplete. Use of relevant drugs will have acted as a proxy for these risk factors but it is possible that adjustments for confounding were inadequate. Other possible causes of between-study heterogeneity include the different ages and baseline risks of the study populations and varying ingested doses of drugs. From a statistical standpoint, the degree of heterogeneity was impressive mandating the use of a random-effects meta-analysis. But inspection of Figure 2 and Figure 3 shows that the range of individual study estimates of RRs was not massive. The large size of many of the studies led to precise RR estimates, meaning that relatively small differences between them were statistically highly significant.

In our view, the other problem with this review lies in the interpretation of pooled RR estimates that are precise but in many cases close to the null. Typically, in pharmacoepidemiological studies there is a reluctance to accept as causal RR estimates much below 2. These studies are subject to a range of biases including confounding by indication and by disease severity. Channeling of certain drugs to patients with high levels of morbidity may lead to serious adverse events being wrongly attributed to the drug rather than a condition from which the patient was already experiencing. This has been noted in the case of selective COX-2 inhibitors.⁴⁴

However, in the face of this and the heterogeneity noted above, a number of factors lead us to believe that the associations are real. The first is that while selection factors may bias estimates of cardiovascular risk made for the class of drugs, they are less likely to affect between-drug comparisons (eg, celecoxib vs rofecoxib or diclofenac vs naproxen). To account for potential confounding at study level, we performed sensitivity analyses where we used only within-study comparisons of drug risks, and our conclusions were unaffected. Additional factors supporting the validity of our observations are the concordance between the results of the randomized and nonrandomized studies and the biological plausibility of the proposed mechanisms.

In conclusion, controlled data from observational and randomized studies confirm a dose-related risk of cardiovascular events with selective COX-2 inhibitors. The observational data indicate that the risk increases early in treatment. An older NSAID, diclofenac, seems to share this risk and, unlike celecoxib, it appears to be harmful at commonly used doses. We believe that there are grounds for reviewing its regulatory status.

Since this article was submitted, Helin-Salmivaara and colleagues in Finland have published a population-based study of the risks of hospitalization with myocardial infarction and use of nonsteroidal anti-inflammatory drugs (Eur Heart J 2006;27:1657-1663; doi 10.1093/eurheartj/ehl053). We have not updated our systematic review but in view of the size of this study we believed it was important to determine if inclusion of the new data would change the results or conclusions of our study. Accordingly, we reran the analyses with inclusion of the new data. The revised summary RR estimates (95% CI) using a random-effects model are as follows: refecoxib, 1.36 (1.18-1.58; 12 studies); celecoxib, 1.06 (0.92-1.22; 12 studies); diclofenac, 1.40 (1.19-1.65; 10 studies); meloxicam, 1.24 (1.06-1.45; 4 studies); naproxen, 0.99 (0.89-1.09; 16 studies); ibuprofen, 1.09 (0.99-1.20; 17 studies); piroxicam, 1.16 (0.86-1.56; 5 studies); indomethacin, 1.36 (1.15-1.61; 7 studies). Our conclusions are unchanged: there appear to be clinically significant differences in summary RR estimates between individual drugs in the doses that are used in the community.