Risk of serious NSAID-related gastrointestinal events during long-term exposure: a systematic review

A high-sensitivity search for randomised trials was conducted by searching the Third Quarter 2002 issue of the Cochrane Controlled Trials Register from its start date. The criteria for study selection were:

• study sample included patients exposed to prescription non-selective NSAIDs for a duration greater than 6 months;

• “serious GI event” incidence or risk was reported, by duration of NSAID exposure;

• events occurred during current use of medication; and

• the study was published in full in English, French or German.

Using the same search criteria, a high-specificity search of Current Contents, EMBASE, MEDLINE and Derwent Drug File databases was conducted in July 2002 (limitations: humans, 1999–2002). In this manner, relevant observational studies and further randomised trials were identified either directly (back to 1999) or indirectly (by manual search of systematic reviews from 1999–2002, representing the NSAID literature back to the 1960s). Abstracts of 479 search results were reviewed; 360 records were excluded for not fulfilling the basic criteria, and 119 records plus an additional 102 records identified from manual searches were retrieved for more detailed evaluation (total, 221) (Box 1).

When retrospective cohort studies using the same health care databases (eg, Tennessee Medicaid, UK General Practice Research Database, Tayside, or ARAMIS) were identified, only one study per database was selected; the rest5,6,19-30 were excluded to avoid potentially redundant patient populations. Meta-analyses were excluded; however, any otherwise unidentified original published studies cited in meta-analyses were retrieved for evaluation against our search criteria. Exclusions were logged with a reason (log available from authors as Appendix 1).

Fourteen studies (represented by 15 publications) were identified as potentially relevant.

Individual studies were reviewed to evaluate the presence of a valid statistical association. Key parameters affecting internal and external validity were assessed according to standard epidemiological principles and Australian government guidelines.31-34 Details are available from the authors as Appendix 2. Based on the assessments, seven studies were included in the quantitative data synthesis. These were stratified into two groups according to study design:

• Group 1. Four randomised trials presenting time-to-event data.35-38 These investigated patients who were using NSAIDs on a regular daily basis and who were subject to protocol-driven constraints (ie, patients were not allowed to switch NSAIDs or to take concomitant anti-ulcer medication; patients were withdrawn after a serious GI event or other monitored adverse events).

• Group 2. Three cohort studies presenting time-to-event data: one prospective study2 and two registry studies.4,39 It is likely that these studies included some patients who used NSAIDs on an “as needed” (prn) basis — detailed compliance information was unavailable.

Box 2 and Box 3 summarise the validity weaknesses identified for the two groups.

Six observational studies,40-45 classified as Group 3, were excluded for lack of time-to-event data for data synthesis. These reported relative risk or excess risk for NSAID exposure versus non-exposure by duration-of-exposure categories, and were reviewed in a supplemental analysis. Details are available from the authors as Appendix 3.

An additional case–control study15 nested within a cohort of patients exposed to NSAIDs for up to 1 year was excluded. Odds ratios were not reported, but serious GI bleed rates over time were, so it potentially qualified for quantitative data synthesis inclusion. However, unlike the included studies, that study based rates on estimated exposure duration (derived from a selected subgroup of patients who did not have a serious GI event), rather than actual exposure data from the entire cohort of exposed patients. This led to concerns about bias, as selected patients without bleeds were unlikely to be representative of patients with bleeds. Furthermore, information on numbers at risk over time was not available for that study.

Data for validity assessment criteria and patient characteristics were extracted from the articles (variables available on request). Outcomes data for serious GI event incidence at specific times were extracted where available, or otherwise were taken from an electronic citation,46 email correspondence (Dr M Mamdani, Scientist, Institute for Clinical Evaluative Sciences, Toronto, Ontario, Canada) and unpublished reports (G Singh, Analysis of a cohort of long-term diclofenac users, 2003; S Morant, Statistician, Cygnus Biostatistics Ltd, Analysis of NSAID exposure duration and GI haemorrhage rates using the UK GP Database, 2002; GD Searle & Co, Final Report for the CLASS trial, 2000). If multiple serious GI event outcome definitions were reported, data were extracted for the strictest definition. Data were only extracted for patients taking non-selective NSAIDs (ie, data from placebo, COX-2 specific inhibitor, or diclofenac/misoprostol treatment arms found in some studies were outside the scope of our review).

For Group 1 and 2 studies separately, a combined analysis across studies and time points using a generalised estimating equations linear regression model47 (a form of meta-regression) was conducted using SAS software version 8 (SAS Institute Inc, Cary, NC, USA). The outcome variable was the estimated annualised probability of having an event, for each time point in each study. This was estimated from the conditional probability of having a serious ulcer complication in a specified time interval, given that no event was experienced before the start of that interval, and was derived from the published cumulative probability estimates for each study. The model included time (in months), study and treatment group as explanatory variables. Change in risk was evaluated by estimating and testing the significance of the regression line slope. Weighted analyses were conducted using the inverse variance method, with variance estimated from Greenwood’s formula.48 A sensitivity analysis using inverse standard error weighting was conducted to verify robustness of the estimates. As data were too limited to perform valid heterogeneity testing within the regression model, pre-assessment of heterogeneity was done (details available from the authors as Appendix 4). Lastly, adequacy of the linear model was assessed by including higher-order terms in the model.

The studies investigated a wide selection of NSAIDs taken in usual clinical doses. One study included aspirin at 1000 mg/day,37 a dose that is standard for pain management but higher than commonly used in Australia for antithrombotic treatment in patients with myocardial infarction. The studies measured the occurrence of upper GI ulcer bleeds and perforations. In addition to ulcer bleeds, two trials37,38 and one observational study43 included hospitalised symptomatic ulcers in the outcome definition.

All studies sampled adult community populations of NSAID users. They included patients with osteoarthritis36 and rheumatoid arthritis,2,35,36,38 unspecified NSAID users,41,42,44,45 or subgroups (> 65 years;4 > 40 years;39,40 and low-income elderly43). NSAID use was for pain treatment, or, if unspecified, was assumed most likely for pain. One study enrolled only patients with myocardial infarction.37 Most patients had taken NSAIDs before study commencement, but one cohort study specifically included NSAID-naïve patients.4

Of the randomised trials, the AMIS trial37 included 100% long-term users (> 6 months) and the VIGOR,35 Neustadt38 and CLASS36 trials included more than 50% long-term users. The prospective cohort study by Singh and Triadafilopoulos2 included more than 50% long-term users. Although the registry cohort and case–control studies were larger, they included a lower proportion of long-term users (< 12% in Mamdani et al,4 12% in UKGPRD,39 and < 18% in Mellemkjaer et al42). Information about the proportion of long-term users was not available for the other three studies.

The results of the individual studies are provided in Box 2 and in the Kaplan–Meier curves in the original articles. Four randomised trials provided Kaplan–Meier time-to-event data from seven non-selective NSAID groups (aspirin 1000 mg/day, na-proxen 1000 mg/day, diclofenac 150 mg/day, two arms with ibuprofen 2400 mg/day, and two arms with etodolac 300 mg/day and 1000 mg/day). Five groups were included in the pooled regression model; the etodolac arms38 were excluded as time-dependent analysis was not considered valid with only two serious GI events per arm.

When randomised trial data were pooled, a small significant reduction in risk of − 0.005% (95% CI, − 0.008% to − 0.001%) per month was seen (Box 4). Three sensitivity analyses were performed. The first addressed the finding that the slope of the CLASS trial data was significantly different to that of the other studies, mainly driven by the ibuprofen group. The analysis was re-run allowing different slopes for the CLASS study versus the other studies. The latter studies showed a small decline in annualised risk of serious GI events of − 0.004% (95% CI, − 0.005% to − 0.003%) per month. The disparate CLASS study showed a significant decline in annualised risk of serious GI events of − 0.16% (95% CI, − 0.223% to − 0.096%) per month of exposure. The second sensitivity analysis again allowed different slopes for the CLASS study versus the others, and excluded the AMIS study (because the patient population was very different to those in the other trials). This showed a non-significant decline in annualised risk of − 0.018% (95% CI, − 0.04% to 0.001%) per month for the VIGOR and Neustadt studies combined. Results were similar when sensitivity analysis was conducted using the alternate weighting method.

The results of the individual studies are summarised in Box 3. For the pooled analysis, the NSAID cohort from the study by Singh and Triadafilopoulos2 was excluded because information on patient numbers over time was not retrievable. Time-to-event data from the diclofenac cohort of the study by Singh and Triadafilopoulos and data from the other two cohort studies were pooled. The slope of the combined regression line shows no significant decline in the annualised probability of serious GI events over time (Box 5). Results were similar when sensitivity analysis was conducted using the alternate weighting method.

Four of six observational studies showed a slight decline in relative risk and excess risk over time, but the risk of serious GI events in NSAID-treated patients did not decline to control-group levels in any of the studies at any time over 1–2 years.

The primary limitation of our study was the small number of long-term studies. Our analysis is a form of meta-regression, which should generally not be considered when there are fewer than 10 trials.55 However, because our analysis had multiple time points per study, this guidance is not as relevant. Detailed discussion of other limitations (eg, the studies, except two, were not designed a priori to evaluate this question; use of high-specificity literature search methods) is available on request. The small number of studies and other limitations mainly affected the precision of the results, but overall, loss of precision would not change the conclusions regarding the nature, direction (slight downward slope) and marginal clinical significance of the risk-over-time relationship.

Our review provides an overview of the evidence to 2002. Although the trends observed in this analysis may suffice for the purposes of clinical practice, further large trials are needed to address this question with optimal statistical rigour. We are aware of one randomised trial that was completed after the 2002 data cut-off for our project. The trial included very large samples of naproxen and ibuprofen users, and reported the cumulative occurrence of serious GI events over 1 year.17 Our analysis could not be updated to include the trial because necessary patient numbers at various time points are not publicly available. However, the published Kaplan–Meier results support our findings of consistent risk over time, leading to an accumulation of serious risk over the long term.17

If future studies are conducted, these should, if feasible, minimise the effect of rare-event bias56,57 (which can be amplified over time for pain agents with tolerability issues) and depletion of susceptibles, maximise the numbers of patients remaining in the study long term, and capture potentially important confounding factors such as anticoagulant use, intermittent versus regular therapy, and whether patients continued NSAID treatment after a serious GI event. These were sources of uncertainty in this review and raised new research questions.