Quantitative estimation of rare adverse events which follow a biological progression: a new model applied to chronic NSAID use

Transcription

1 Pain 85 (2000) 169±182 Quantitative estimation of rare adverse events which follow a biological progression: a new model applied to chronic NSAID use Martin R. TrameÁr a, *, R. Andrew Moore b, D. John M. Reynolds c, Henry J. McQuay b a Division d'anestheâsiologie, DeÂpartement APSIC, HoÃpitaux Universitaires, CH-1211 Geneva 14, Switzerland b Pain Research, Nuf eld Department of Anaesthetics, The Churchill, Oxford Radcliffe Hospital, Oxford OX3 7LJ, UK c Department of Clinical Pharmacology, Radcliffe In rmary, Woodstock Road, Oxford OX2 6HE, UK Received 6 January 1999; received in revised form 6 September 1999; accepted 10 November 1999 Abstract Randomised controlled trials (RCTs) alone are unlikely to provide reliable estimates of the incidence of rare events because of their limited size. Cohort, case control, and other observational studies have large numbers but are vulnerable to various kinds of bias. Wanting to estimate the risk of death from bleeding or perforated gastroduodenal ulcers with chronic usage of non-steroidal anti-in ammatory drugs (NSAIDs) with greater precision, we developed a model to quantify the frequency of rare adverse events which follow a biological progression. The model combined data from both RCTs and observational studies. We searched systematically for any report of chronic ($2 months) use of NSAIDs which gave information on gastroduodenal ulcer, bleed or perforation, death due to these complications, or progression from one level of harm to the next. Fifteen RCTs ( patients exposed to NSAIDs for 2±60 months), three cohort studies ( patients redeeming a NSAID prescription over a 3±12 month period), six case-control studies (2957 cases) and 20 case series (7406), and case reports (4447) were analysed. In RCTs the incidence of bleeding or perforation in 6822 patients exposed to NSAIDs was 0.69%; two deaths occurred. Of patients with bleeding or perforation with or without NSAID exposure across all reports, 6±16% (average 12%) died; the risk was lowest in RCTs and highest in case reports. Death from bleeding or perforation in all controls not exposed to NSAIDs occurred in 18 out of (0.002%). From these numbers we calculated the number-needed-to-treat for one patient to die due to gastroduodenal complications with chronic ($2 months) NSAIDs as 1= 0:69 {6±16%, average 12%} 2 0:002% ˆ 909±2500 (average 1220). On average 1 in 1200 patients taking NSAIDs for at least 2 months will die from gastroduodenal complications who would not have died had they not taken NSAIDs. This extrapolates to about 2000 deaths each year in the UK. q 2000 International Association for the Study of Pain. Published by Elsevier Science B.V. All rights reserved. Keywords: Meta-analysis; Adverse event; Adverse drug reaction; Non-steroidal anti-in ammatory drug; Aspirin; Systematic review; Complications 1. Introduction How can a rare but serious adverse effect be proven to be associated with a particular intervention? One way might be the examination of drug monitoring centres ± although this has been shown to provide highly variable results (TrameÁr et al., 1997). Cohort and case-control studies (for instance, using retrospective data from regional death certi cates) are unreliable in determining exposure, so may underestimate the problem. Postmarketing surveillance trials are usually uncontrolled series. Published case reports may be informative for rare adverse events, but lack a denominator, the number of people exposed to the drug. These methods all suffer, compared with randomised controlled trials * Corresponding author. Tel.: ; fax: address: martin.tramer@hcuge.ch (M.R. TrameÁr) (RCTs), from an increased likelihood of selection bias, observer bias, and confounding factors. Case-control and cohort studies are usually conducted many years after introduction of the intervention. For instance, an excellent large case-control study on gastrointestinal bleeding with nonsteroidal anti-in ammatory drugs (NSAIDs) (Blower et al., 1997) was published about 30 years after the introduction of NSAIDs. The information in the Blower study can be used to estimate an annual death rate from NSAID-related gastrointestinal bleeding at about 1 in 2070 (Moore and Phillips, 1999). Although adverse drug reactions happen rarely, they can still pose major health burdens. Clinical decisions should be based on the best available scienti c evidence (Davidoff et al., 1995), and RCTs are accepted as the gold-standard for quantifying ef cacy (Sacks et al., 1982). Not everything, however, can be investigated by RCTs (Black, 1996) /00/$20.00 q 2000 International Association for the Study of Pain. Published by Elsevier Science B.V. All rights reserved. PII: S (99)

2 170 M.R. TrameÁr et al. / Pain 85 (2000) 169±182 They are, because of their limited size, less useful for estimating the incidence of rare serious adverse events, which are often incompletely reported (Edwards et al., 1999). It has been calculated that the sample size needed to detect a difference between an incidence rate of 1/ and 2/ at the 90% power level is over for both treatment and control group (Wardell et al., 1979). What is needed is a method that can allow different types of evidence to be used to calculate a risk of an adverse effect occurring in absolute terms. In this paper we present a innovative model to estimate the incidence of a rare adverse event, using drug-related death as the example. The model synthesises systematically searched data from reports with different study architectures; all these different types of reports contain potentially valuable data on adverse events but with different strengths of evidence and power for causation and estimation. There is no current strategy for the synthesis of this heterogeneous information. The model is applicable to any adverse event that follows a biological progression (TrameÁr et al., 1997). As a clinical example we have chosen death from gastroduodenal complications with chronic usage of NSAIDs for three reasons. 1. There is solid evidence that NSAIDs cause peptic and duodenal ulceration, and for a biological progression from symptomatic ulcers to bleed or perforation, and even death (Hawkey, 1990; Soll et al., 1991). Estimates of risk from the model can be compared with estimates from the case-control literature. 2. NSAIDs are widely used in the western world. For each patient prescribed an NSAID there is an estimated adverse effect burden of 48 per year (Moore and Phillips, 1999). Knowing the risk of major complications and death with long-term NSAID is therefore important. 3. The uncertainty about the frequency of death due to gastroduodenal complications with chronic NSAID needs to be emphasised. Using a retrospective casecontrol study it has been estimated recently that NSAIDs caused 2560 deaths due to gastrointestinal complications in the UK per year (Blower et al., 1997). We set out to verify these numbers using a completely different approach. 2. Methods 2.1. Systematic search A systematic search was done for any type of report of gastric or duodenal ulcer, ulcer haemorrhage or perforation, and death attributable to such events with or without oral NSAID (including aspirin) exposure. MEDLINE (KnowledgeFinder 4.0 and SilverPlatter 3.25) and EMBASE databases were searched (up to December 1996), not restricted to English language, with the free text terms non-steroidal anti-in ammatory drug, aspirin, ulcer, bleeding, haemorrhage, perforation, and death, and combinations of these words. Reference lists from retrieved reports, review articles on NSAIDs, relevant meta-analyses (Gabriel et al., 1991; Roderick et al., 1993; Eisenberg et al., 1994; Koch et al., 1995, 1996), and gastroenterology textbooks were handsearched. Abstracts from scienti c meetings and reports of NSAID-induced complications other than upper gastrointestinal were not considered. Pharmaceutical companies and authors were not contacted. RCTs were considered only when drug exposure was at least two months, when NSAID was compared with no treatment, a placebo, or an NSAID plus a mucosal protective agent, and when dichotomous data on gastroduodenal complications were reported. Data from different doses of a protective agent within one RCT were combined for analysis Data extraction Information about type of reports, patient characteristics, dose of NSAID and exposure time, de nition of, and dichotomous data from different levels of harm was taken from each report. Data on peptic, gastric, or duodenal ulcer, as de ned in the original reports, were extracted. Other signs of mucosal damage (gastritis, for instance) were not considered because they were reported inconsistently. Data on haemorrhage or perforation of a gastroduodenal ulcer, as described in the original reports, were extracted. A patient with a known gastric or duodenal ulcer and melaena stool would fall into this category. Other sources of bleeding, such as oesophageal varices, and other causes of perforation, such as carcinomatous perforations, were not considered. Data on death attributable to a bleeding or perforated gastroduodenal ulcer, including death during or after surgery for such complications, were extracted Qualitative analysis A scattergram (L'Abbe et al., 1987) was created using the dichotomous data from comparative trials (i.e. RCTs and cohort studies). Rates of events with NSAIDs were plotted against the respective events with controls. A scatter lying predominantly between the line of equality and the NSAID axis would qualitatively suggest a consistently increased risk with NSAIDs and relative homogeneity of the data. There was no intention to analyse the scatter quantitatively Quantitative analyses The primary aim of this study was to estimate quantitatively the number of patients who need to be treated chronically ($2 months) with oral NSAIDs for one to die due to gastroduodenal complications who would not have died had they not received NSAIDs (McQuay and Moore, 1997). A rst attempt to estimate this additional risk was done with dichotomous data from RCTs and cohort studies only. A

3 M.R. TrameÁr et al. / Pain 85 (2000) 169± second estimate of NSAID-related risk of death was calculated taking into account ratios between different levels of harm from all reports independent of study architecture (i.e. the model of biological progression). The ratio `ulcer: bleed/perforation' indicated how many patients with an ulcer went on to have a bleed or a perforation. The ratio `bleed/perforation: death' indicated how many bleeding or perforated ulcers resulted in death. These ratios were calculated for all reports (i.e. independent of study architecture). Case-control studies, case series, and case reports had to document absolute numbers of patients progressing from one level of harm to at least the next level of harm. For instance, in an uncontrolled series of patients with bleeding ulcers we sought the number of patients who eventually died due to their ulcer. This example would include both patients exposed to NSAIDs and those who did not receive NSAIDs (i.e. not limited to NSAID patients). Combined ratios weighted by sample size were calculated. 3. Results The search strategy identi ed 150 reports. Forty-nine reports contained data potentially relevant to this metaanalysis (Table 1): fteen RCTs with patients receiving different NSAIDs and controls receiving a placebo, or NSAID plus misoprostol, ranitidine, famotidine, omeprazole, or nizatidine; three cohort studies with patients ( of them exposed to NSAIDs); six case-control studies with 2957 `cases' with a diagnosis of gastroduodenal ulcer or bleeding/perforation; 20 uncontrolled series with a total of 7406 patients with a diagnosis of an ulcer or bleeding/perforation; and 4447 case reports or spontaneous reports to drug monitoring centres on patients with upper gastrointestinal complications. These 49 reports included 2081 patients with an uncomplicated ulcer, patients with ulcer bleed or perforation, and 1329 deaths due to ulcer bleed or perforation (Table 1). All drugs were given orally. In RCTs drug exposure was 2±60 months. In cohort studies exposure was de ned as redeeming an NSAID prescription in the previous 3±12 months. Extraction of data for individual NSAIDs was not possible because no appropriate distinction between drugs was made in the original reports. Data on NSAIDs and aspirin were pooled for analysis. Because reports were inconsistent in reporting intention-to-treat and ef cacy data, only ef cacy data were analysed Risk in RCTs and cohort studies Event rate scatter with data from RCTs and cohort studies In RCTs and cohort studies there was a consistently increased risk for all three levels of harm (ulcer, bleed/ perforation, and death) with NSAIDs (Fig. 1). For an uncomplicated ulcer there were two clusters, i.e. trials reported two different ranges of risk. These were identi ed respectively as endoscopically diagnosed ulcers and clinically diagnosed (i.e. symptomatic) ulcers. There was overlap between the data for clinically diagnosed ulcers and bleed/perforation Endoscopically diagnosed ulcer in RCTs In 12 RCTs involving a total of 4111 patients with osteoarthritis or rheumatoid arthritis (Ehsanullah et al., 1988; Graham et al., 1988, 1993; Robinson et al., 1989, 1991; Roth et al., 1989; Verdickt et al., 1992; Levine et al., 1993; Elliott et al., 1994; Agrawal et al., 1995; Raskin et al., 1995; Ekstrom et al., 1996; Taha et al., 1996) all patients underwent screening endoscopy to identify the incidence of new ulcers. The median trial size was 274 patients (range 70±1171). Different NSAIDs plus placebo were compared with NSAIDs plus a protective agent, follow-up was 2±12 months, and endpoints (i.e. ulcer diameter on endoscopy) were well de ned. These trials made up the cluster with the higher incidence of uncomplicated ulcers (Fig. 1). The average absolute risk for an endoscopically diagnosed ulcer with NSAID was 21%. Because these trials did not report any patient with a more serious level of harm (bleeding or perforated ulcer, for instance) they were not analysed further Clinically diagnosed (symptomatic) ulcer in RCTs and cohort studies The other three RCTs reported the incidence of ulcers which were diagnosed because they caused symptoms (as opposed to detection by screening endoscopy) (Steering Committee, 1989; Kurata and Abbey, 1990; Silverstein et Fig. 1. Rate of gastroduodenal complications (`events') with NSAIDs (includes aspirin) compared with control (no NSAID, or placebo, or NSAID 1 mucosal protection therapy) in 15 RCTs (squares) and 3 cohort studies (circles). White symbols: uncomplicated peptic, gastric or duodenal ulcer; grey symbols: ulcer bleed or perforation; black symbols: death attributable to a bleeding or perforated ulcer. Several trials reported several levels of harm, i.e. several `events'. Death outcomes in two RCTs had a control event rate of 0%; these were set a control event rate of 0.001% for graphical purposes. The dotted line indicates the line of equality.

4 172 M.R. TrameÁr et al. / Pain 85 (2000) 169±182 Table 1 Oral NSAIDS/aspirin and gastroduodenal complications: analysed reports a Reference Type Patients, source population, inclusion criteria, follow-up Design (number of patients per group) Agrawal et al., 1995 RCT RA and OA patients taking diclofenac, mean age 57 years, without GI lesions (score #3 on baseline endoscopy) Misoprostol 400±600 mg 1 diclofenac (102) Placebo 1 diclofenac (102) Ehsanullah et al., 1988 RCT RA patients taking NSAIDs, 22±87 years (mean 60), without GI lesions (baseline endoscopy) Graham et al., 1988 RCT OA patients, with abdominal pain and endoscopically no ulcer. Age 22±90 years (mean 59 years). Treatment for 3 months Ekstrom et al., 1996 RCT OA and RA patients, 25±78 years, history of uncomplicated peptic ulcers or dyspepsia. Screening endoscopy after 1 and 3 months Elliott et al., 1994 RCT Arthritis patients,.18 years (average 66 years), 12 months follow-up (endoscopy) Graham et al., 1993 RCT OA, RA etc patients, ^ abdominal pain, endoscopically no ulcer. Age n/a. Treatment for 3 months Kurata and Abbey, 1990 RCT Adverse event pro le from AMIS (JAMA 1980;243:661±9). MI patients, mean age 55 years (30±69). Excluded: ulcer. Follow-up minimum 3 years Levine et al., 1993 RCT 496 OA patients with NSAIDassociated dyspeptic symptoms..18 years (average 57 years). Exclusion: ulcer at rst endoscopy. 12 weeks exposure and follow-up Raskin et al., 1995 RCT 1623 OA and RA patients (and others) with gastric and/or duodenal lesions, age 18±89 years. Endoscopic examination in all patients. 12 weeks follow-up Robinson et al., 1989 RCT Arthritis patients without gastric and/or duodenal lesions, mean age 43±50 years. Endoscopic examination in all patients. 8 weeks follow-up Roth et al., 1989 RCT RA patients with gastric and/or duodenal lesions, mean age 53 years. Endoscopic examination in all patients. 8 weeks follow-up Roth et al., 1989 RCT RA patients with gastric and/or duodenal lesions, mean age 53 years. Endoscopic examination in all patients. 8 weeks follow-up Silverstein et al., 1995 RCT 8843 RA patients without an active peptic ulcer, mean age 68 years. 6 months follow-up Ranitidine 150 mg b.i.d. 1 NSAID (137) Placebo 1 NSAID (126) Misoprostol 400 mg 1 NSAID (143) Misoprostol 800 mg 1 NSAID (139) Placebo 1 NSAID (138) Omeprazole 20 mg 1 NSAID (85) Placebo 1 NSAID (90) NSAID 1 misoprostol 600±800 mg/day (32) NSAID 1 placebo (38) Misoprostol 800 mg 1 NSAID (215) Placebo 1 NSAID (240) Aspirin 1 g/day (2267) Placebo (2257) NSAID 1 nizatidine 150 mg b.i.d. (248) NSAID 1 placebo (248) NSAID 1 misoprostol 400±800 mg (846) NSAID 1 placebo (325) NSAID 1 ranitidine 300 mg/day (61) NSAID 1 placebo (54) Aspirin 2.5±5.2 g/day 1 misoprostol (123) Aspirin 2.5±5.2 g/day 1 placebo (116) Aspirin 2.5±5.2 g/day 1 misoprostol (123) Aspirin 2.5±5.2 g/day 1 placebo (116) NSAID 1 misoprostol (4404) NSAID 1 placebo (4439) No. of analysed patients Ulcer Bleed or perforation Death De nition No. De nition No. De nition No. 102 Endoscopic gastric or duodenal ulcer (`ulceration with or without visible vessel') at 52 weeks with diclofenac 1 placebo 126 Endoscopic gastric or duodenal ulcer (.5 mm) at 8 weeks with NSAID 1 placebo 138 Endoscopic gastric ulcer (.3 mm) at 3 months with NSAID 1 placebo Endoscopic gastric ulcer (.3 mm) at 3 months with NSAID 1 placebo Endoscopic gastric ulcer (2±12 mm) at 12 months with NSAID 1 placebo 240 Endoscopic gastric or duodenal ulcer (.5 mm) at 3 months with NSAID 1 placebo 2267 Hospitalisation for duodenal or gastric ulcer with aspirin GI bleeding with aspirin Gastric- or duodenal ulcer ( 3 3 mm) endoscopically at 12 weeks Gastric ulcer ( 3 3 mm) endoscopically at 12 weeks Gastric or duodenal ulcer (.5 mm) endoscopically at 8 weeks Hospitalisation because of severe upper GI bleeding with aspirin 1 placebo 1 n/a `New' ulcers (endoscopically - grading, crater) at 8 weeks with aspirin 1 placebo 25/81 (31%); with misoprostol 16/86 (19%) 4439 Uncomplicated duodenal or gastric ulcer with NSAID 1 placebo Ulcer bleeding (categories 3±6), perforation or obstruction (categories 1 and 2) with NSAID 1 placebo 33 Death from haemorrhage with NSAID 1 placebo 1

5 M.R. TrameÁr et al. / Pain 85 (2000) 169± Table 1 (Continued) Reference Type Patients, source population, inclusion criteria, follow-up Steering Committee, 1989 RCT Physicians, follow-up 60 months (average) Taha et al., 1996 RCT 285 patients with RA or OA, without peptic ulcers at endoscopy. 18±88 years (average 53±57). Standard doses of different NSAIDs. 24 weeks follow-up Verdickt et al., 1992 RCT 339 RA patients, average age 53 years, 3 months' exposure and follow-up Bloom, 1989 cohort Data compiled from Medicaid Management Information System. Pennsylvania 1984/5. Excluded: diagnosis of GI illness, ulcer medication. Adjusted population on NSAID ˆ , controls ˆ Median 61 years Carson et al., 1987 cohort Retrospective (historical) cohort study. Data from computerised on-line Medicaid system. 1980, Michigan and Minnesota. NSAID-exposed patients n ˆ Unexposed n ˆ (96 bleedings). Excluded: prior history of GI bleeding, aspirin users Guess et al., 1988 cohort Records linkage, review of hospital charts and autopsy reports of fatal GI bleeding and/or perforation among all members of the Saskatchewan Health Care Plan NSAID users , nonusers Bartle et al., 1986 CC Patients hospitalised with nonvariceal acute upper GI tract haemorrhages (endoscopically diagnosed), age 27±99 years Henry et al., 1987 CC Patients admitted to hospital with complicated peptic ulcer. Average age 69 years (40±80) Holvoet et al., 1991 CC MC, patients hospitalised because of upper GI bleeding, mean age 68 years (19±92) Holvoet et al., 1991 CC MC, patients hospitalised because of upper GI bleeding, mean age 68 years (19±92) Jick et al., 1987 CC Clinical charts and automated data study in Seattle 1977±83. Age.10 years. All patients with diagnosis of perforated peptic ulcer on discharge from hospital McIntosh et al., 1988 CC New cases of peptic ulcer in different centres 1982±85 Design (number of patients per group) Aspirin 325 mg/2 days (11 037) Placebo (11 034) NSAID 1 famotidine 20 mg b.i.d. (95) NSAID 1 famotidine 40 mg b.i.d. (97) NSAID 1 placebo (93) Diclofenac 1 misoprostol 400± 600 mg (137) Diclofenac 1 placebo (153) No. of analysed patients Ulcer Bleed or perforation Death De nition No. De nition No. De nition No Ulcer (de nition n/a) with aspirin 138 Upper GI haemorrhage among those with an ulcer with aspirin 93 Gastric and duodenal ulceration at 24 weeks (endoscopy 3 3 mm): 24/ 93 on NSAID 1 placebo Death from GI haemorrhage Gastric or duodenal ulcer of any size (endoscopy) at 3 months Gastric ulcer, duodenal ulcer, all other peptic ulcers with NSAID 131 GI bleeding (vomiting, blood, blood in stool, GI haemorrhage) with NSAID Upper GI bleeding with NSAID Hospitalisations of NSAID-users with fatal outcome in which GI bleeding and/or perforation was a con rmed diagnosis Hematemesis and/or black stool on admission; endoscopic con rmation 57 Death Admitted to hospital with diagnosis of haemorrhage from, or perforation of, a peptic ulcer. 161 Patients needing surgery for `erosive' upper GI lesions (peptic oesophagitis, gastric erosions, gastric ulcer, duodenal ulcer) n/a Symptoms from peptic ulcer disease 140 Bleeding gastric or duodenal ulcer Death after admission Death People admitted to hospital with `well-documented' perforated peptic ulcer (any aetiology) 54 Death New peptic ulcer 225 Bleeding 63

6 174 M.R. TrameÁr et al. / Pain 85 (2000) 169±182 Table 1 (Continued) Reference Type Patients, source population, inclusion criteria, follow-up Design (number of patients per group) No. of analysed patients Ulcer Bleed or perforation Death De nition No. De nition No. De nition No. Rodriguez and Jick, 1994 CC Data from computerised medical information system. Source population: people, aged 25±77 years, registered with GPs in 1990 (UK). Follow-up until rst hospital admission for upper GI bleeding, death, or end of study period Abbasakoor et al., 1995 CS Hospital record study. Consecutive patients with perforated peptic ulcer undergoing laparotomy. Mean age 49 years (15±89) Agrez et al., 1992 CS All cases of perforated ulcer admitted to hospitals 1979±86. Dg: surgery, autopsy, radiography. Age 11±94 years (mean 57±65 years) Armstrong and Blower, 1987 CS Consecutive series of elderly patients (most.60 years) admitted with melena or haematemesis. 1983±85. Retrospective case note examination and prospective evaluation. Population about Beard et al., 1987 CS NSAID prescriptions and hospitalisations for upper GI bleeding of all members of a consumer-owned organisation (Seattle, 1977±83). Age.65 years. n ˆ members in 1983 Booth and Williams, 1971 CS 10- years experience in a general hospital 1958±67: patients (mean age 45±52 years). Only data on subgroup with duodenal ulcer reported Cohen, 1971 CS Hospital record study (Vancouver area: 1959±68, 12 hospitals, mean age 45 years): all admissions with proven (laparotomy or autopsy) perforated ulcer (n ˆ 852). Population mean n ˆ (1961), (1968) Doberneck, 1993 CS Hospital chart review of patients undergoing urgent surgery for bleeding or perforated gastric ulcer. 1 centre, 1977±88. Average age 51 years (21±78) Fisher et al., 1966 CS Patients operated for benign ulcer disease, 1 hospital, 1952± Episode of bleeding or perforation involving the stomach, the duodenum, or a peptic ulcer. 100 Perforated gastric (7), pyloric (2), or duodenal ulcer (91) (Dg surgical) Survey 174 Perforated gastric or duodenal ulcer 235 Perforated or bleeding gastric or duodenal ulcer (Dg surgical or post mortem) 201 Patients hospitalised for gastroesophageal bleeding 1457 Death Operative mortality 6% Death Death Death at least in part as a result of bleeding 340 Duodenal perforation 340 Death following perforation Proven (laparotomy or autopsy) perforated peptic ulcer 24 Urgent operation for bleeding or perforated gastric ulcer Survey 449 Surgery for chronic or acute bleeding, or perforation Fries, 1992 CS Follow-up of RA patients (n ˆ 2747). Data base study Hospitalisations for GI problems (mostly GI haemorrhages). Gerber et al., 1981 CS Hospital record study. Period 23 years. Patients with peptic ulcer and RA. Mean age 41.2 years Horowitz et al., 1989 CS Retrospective analysis of all patients having diagnoses of perforated ulcer at 2 hospitals in Pennsylvania 1975±85. Excluded: malignancy, stress ulcer, gastrinoma. Mean age 54±63 years 27 Active ulcer (presence of symptoms and radiographic or endoscopic demonstration of an ulcer crater) in patients on NSAID 27 Haemorrhage or perforation in patients on NSAID 80 Perforated ulcers (duodenal, peptic, gastric, prepyloric) 852 Death (overall) Death Death Death attributed to NSAID-related gastropathy 5 Death due to bleeding in patient on NSAID 80 Death

7 M.R. TrameÁr et al. / Pain 85 (2000) 169± Table 1 (Continued) Reference Type Patients, source population, inclusion criteria, follow-up Design (number of patients per group) No. of analysed patients Ulcer Bleed or perforation Death De nition No. De nition No. De nition No. Irvin, 1989 CS Review of case records of patients admitted to one hospital with a perforated peptic ulcer (1979±86). Mean age 70 years (19±98) Jatzko et al., 1993 CS Survey on all surgically treated (acute and elective) ulcer patients in one hospital 1984±92. Average age women: 72 years; men: 58 years) Jordan et al., 1966 CS Patients operated for perforated gastroduodenal ulcers. 3 hospitals, 16 year period ending 1964 Madsen et al., 1995 CS Long-term recurrence rate and survival in patients with bleeding gastric ulcer. Median age 69 (23±92) O'Riordain et al., 1990 CS Elderly patients (77 (70±93) years), admitted with perforated duodenal ulcer. 1978±87 (10 year period) Peoples, 1985 CS Retrospective review of hospital records in 1 hospital 1981±83. Patients with diagnosis of peptic ulcer. Mean age 60 years (14±92). Excluded: stress-, malignant ulcers Sanders, 1967 CS Survey (case notes, record search, necropsies) in the Oxford area between 1957 and All perforated gastric and duodenal ulcers (n ˆ 217). Mean age 51±61 years. Assumed population: Suter et al., 1992 CS Retrospective hospital record study. All patients admitted for perforated or bleeding ulcer. Mean age 62 years. Period 1982±92. Surgical or conservative treatment Werbin et al., 1990 CS Retrospective survey, 10 year period 1978±87, elderly patients (70±94 years, mean age 77.5 years), operated for free perforation of a duodenal ulcer. Cina et al., 1994 CR 4 misdiagnosed cases of perforated peptic ulcers Figueras et al., 1994 CR Spontaneous reporting of adverse drug reactions to NSAIDs in Spain (1983± 1991) reports on NSAID adverse effects: 1294 analysed. Subgroup (n ˆ 998) on 4 different NSAIDs Leonard and O'Briain, 1995 CR Simultaneous perforation in an elderly couple CSM, 1986a a YC Spontaneous reporting of NSAID adverse events to the CSM (1964 to mid-1985) CSM, 1986b b YC Spontaneous reporting of NSAID adverse events to the CSM (1964 to mid-1985) Case record review 284 Perforated peptic ulcer, con rmed by operation (n ˆ 265), radiology (n ˆ 3) or at autopsy (n ˆ 16) 320 Surgery for bleeding or perforated ulcer Survey 651 Surgery because of perforated gastric or duodenal ulcer Prospective follow-up for 5±8 years (median 6.5 years) 90 Bleeding gastric ulcer discharged after non-operative treatment Retrospective survey 33 Perforated duodenal ulcer, needing surgery 284 Death Postoperative death Death Death 3 33 Death Peptic ulcer disease 265 Haemorrhage or perforation 80 Death due to haemorrhage or perforation 216 Perforated peptic ulcers (Dg: x- ray, surgery, post mortem) 216 Death Perforated ulcer or haemorrhage 283 Death Perforated duodenal ulcer, needing surgery 35 Death 8 4 Perforation 4 Death due to peritonitis GI adverse events 750 GI bleeding 220 Death due to GI bleeding 3 2 Perforation of duodenum ulcer 2 Death due to perforation Upper GI bleeding or perforation 3443 Death 576 n/a Upper GI bleeding or perforation n/a Death n/a Sum a CC, case control; CR, case report; CS, case series; RCT, randomised controlled trial; YC, yellow cards (ˆspontaneous reports to drug monitoring centres); n/a, not applicable; GI, gastrointestinal; NSAID, non-steroidal anti-in ammatory drug; OA, osteoarthritis; RA, rheumatoid arthritis.

8 176 M.R. TrameÁr et al. / Pain 85 (2000) 169±182 al., 1995) (Table 2.1). These three trials made up the cluster with the lower incidence of uncomplicated ulcers (Fig. 1). The median trial size was 8843 patients (range 4524± ), and the exposure time 6±60 months. The average incidence of a symptomatic ulcer in patients exposed to NSAID was 1.48% (Table 2.1.). One of the three RCTs compared a low dose of aspirin (325 mg every second day) with placebo for 60 months (Steering Committee, 1989). The incidence of a symptomatic ulcer with drug exposure in this trial was 1.53% compared with 1.39% in the other two RCTs. This difference was not statistically signi cant (relative risk 1.1 (95% CI 0.9±1.4)).One cohort study (Bloom, 1989) reported an incidence of a symptomatic ulcer of 0.39% only in patients (Table 2.2.). This incidence was signi cantly lower than the incidence reported in the RCTs (relative risk 2.3 (95% CI 1.7±3.1)) Ulcer bleeding or perforation in RCTs and cohort studies In the four RCTs that reported it (Roth et al., 1989; Steering Committee, 1989; Kurata and Abbey, 1990; Silverstein et al., 1995), the average incidence of a bleed or perforation with NSAID was 0.48% in patients (Table 2.3.). There were three trials with three different aspirin regimens: 325 mg every second day for 60 months (Steering Committee, 1989), 1 g per day for 36 months (Kurata and Abbey, 1990), and 2.5±5.2 g per day for 2 months (Roth et al., 1989). The incidence of an ulcer bleed or perforation was 0.34% with the lowest dose of aspirin, 0.57% with the medium dose, and 0.86% with the highest dose (Table 2.3). The difference between the low dose aspirin trial (0.34%) and the other three trials (combined incidence of an ulcer bleeding or perforation with aspirin or any NSAID was 0.69%) was statistically signi cant (relative risk 2.8 (95% CI 1.2±6.2)).There were two relevant cohort studies which reported ulcer bleeding or perforation (Bloom, 1989; Carson et al., 1987) and in these the incidence of a bleeding or perforation was 0.22% only in patients redeeming an NSAID prescription (Table 2.4.). This incidence was signi cantly lower than the incidence reported in all RCTs (relative risk 2.2 (95% CI 1.7±2.8)) Death in RCTs and cohort studies Twelve deaths due to gastrointestinal events were reported in patients exposed to NSAIDs in two RCTs and one cohort study reporting deaths (Table 2.5). Both RCTs reported one death, one with NSAID (Silverstein et al., 1995), one with low-dose aspirin (Steering Committee, 1989); the cohort study reported ten deaths on NSAID exposure. The combined death rate in exposed patients was 0.008%.Eighteen deaths due to a bleed or a perforated ulcer were reported in controls who had received either a placebo for 60 months in one RCT (Steering Committee, 1989), NSAID plus misoprostol for 6 months in another RCT (Silverstein et al., 1995), or no NSAID for 12 months in a cohort study (Guess et al., 1988). This yielded a control death rate of 0.002% (Table 2.5.). The difference between death rates in exposed patients and controls was statistically signi cant in the cohort study only (relative risk 3.5 (95% CI 1.6±7.5)) but not in the two RCTs. Based on data from the two RCTs and the cohort study which reported death, the number-needed-to-treat to treat for one patient to die due to chronic usage of NSAID was (95% CI 9478±79 284), the relative risk ( xed effect model) was 3.4 (95% CI 1.3±8.7) Redoing the estimation of the risk of NSAID-related death The number-needed-to-treat The number-needed-to-treat informs us about the additional risk or bene t of an experimental intervention compared with a control intervention. This calculation requires the absolute risk of an event without exposure to the experimental intervention (control event rate) and the absolute risk of the same event with exposure (experimental event rate) (Laupacis et al., 1988). To estimate the additional risk of death from chronic usage of NSAIDs we need to estimate valid death rates with and without drug exposure. The control death rate in patients from RCTs and cohort studies was 0.002% (Table 2.5.). There cannot be the same degree of con dence in the death rate with drug exposure as reported by RCTs because only two deaths were reported. Nor can there be much con dence in the risk of drug-related complications in a cohort study since cohort studies underestimated the absolute risk for both drug-related symptomatic ulcer (Table 2.2.) and drugrelated bleeding/perforation (Table 2.4.) compared with RCTs by a factor of 3±4. The absolute risk of death with drug exposure can, however, be estimated by combining the data from RCTs for drug-related ulcer bleeding or perforation (i.e. the most reliable endpoint apart from death) with that from all reports (i.e. including observational studies) documenting the risk of death once bleeding or perforation has occurred. This calculation is based on a biological progression (TrameÁr et al., 1997) Biological progression: ratios between different levels of harm In the different types of reports (i.e. including observational studies) which documented progression from a gastroduodenal ulcer to bleed or perforation, 1769 ulcer patients with or without NSAID exposure were described and 541 of them had a haemorrhage or perforation (Table 3). The mean likelihood of an ulcer bleeding or perforating was 1 in 3.3 (30%). In three RCTs and two case-control studies (214 patients with a bleed or perforation) the ratio was 1 in 2.8 (36%). In two case series and one report of single observations (305 patients with a bleed or perforation) the ratio

9 M.R. TrameÁr et al. / Pain 85 (2000) 169± Table 2 Oral NSAIDS/aspirin and gastroduodenal complications: raw data from randomised control studies and cohort studies a Reference Report type Experimental intervention Control intervention Exposure time (months) Endpoint Number of patients with endpoint/total number of patients Experimental intervention EER% Control intervention CER% 2.1. Symptomatic ulcer: RCTs Silverstein et al., 1995 RCT NSAID 1 placebo NSAID1misoprostol 800 mg 6 Uncomplicated gastric or duodenal ulcer 60/ Kurata and Abbey, 1990 RCT Aspirin 1g/day Placebo 36 Hospitalisation for gastric or duodenal ulcer 33/ Steering Committee, 1989 RCT Aspirin 325 mg/2 days Placebo 60 Ulcer 169/ RCTs pooled 262/ Symptomatic ulcer: cohort study Bloom, 1989 Cohort NSAID No NSAID n/a Gastric, duodenal, or peptic ulcer 131/ Ulcer bleeding or perforation: RCTs Steering Committee, 1989 RCT Aspirin 325 mg/2 days Placebo 60 Ulcer hemorrhage 38/ Kurata and Abbey, 1990 RCT Aspirin 1 g/day Placebo 36 GI bleeding inclusive tarry stools 13/ Roth et al., 1989 RCT Aspirin 2.5±5.2 g/day1placebo Aspirin1misoprostol 800 mg 2 Hospitalisation because of upper GI bleeding 1/ Silverstein et al., 1995 RCT NSAID1placebo NSAID1misoprostol 800 mg 6 Ulcer bleeding, perforation or obstruction 33/ RCTs pooled 85/ RCTs pooled (Steering Committee, / excluded) 2.4. Ulcer bleeding or perforation: cohort studies Bloom, 1989 Cohort NSAID No NSAID n/a GI bleeding 22/ Carson et al., 1987 Cohort NSAID No NSAID n/a Upper GI bleeding 155/ Cohort studies pooled 177/ Death attributable to ulcer bleeding or perforation: RCTs and cohort study Silverstein et al., 1995 RCT NSAID1placebo NSAID1misoprostol 800 mg 6 Death as a direct result of GI event 1/ / Steering Committee, 1989 RCT Aspirin 325 mg/2 days Placebo 60 Death due to haemorrhage 1/ / Guess et al., 1988 Cohort NSAID No NSAID n/a Fatal upper GI bleeding or perforation 10/ / RCTs and cohort studies pooled 12/ / a RCT, randomised controlled trial; NSAID, non-steroidal anti-in ammatory drug; CI, con dence interval; GI, gastrointestinal; n/a, not applicable (exposure time ˆ redeeming a prescription within a de ned period of time); EER, experimental event rate (absolute risk for the endpoint with NSAID); CER, control event rate (absolute risk for the endpoint with control ˆ placebo, or no treatment, or NSAID 1 mucosa protection therapy).

10 178 M.R. TrameÁr et al. / Pain 85 (2000) 169±182 Table 3 Oral NSAIDs/aspirin and upper gastrointestinal complications: ratios between different levels of harm a Type of report Ulcer: bleed/perforation Bleed/perforation: death Number of patients Ratio Studies Number of patients Ratio Studies Ulcer Bleed or perforation Bleed or perforation Death RCT Cohort No data Case-control Case series Case reports Total a Data are shown of patients who deteriorated from one level of harm to at least one worse level of harm. Data from case-control studies, case series, and case reports include both patients exposed to NSAIDs and those who did not receive NSAIDs. RCT, randomised controlled trial. No haemorrhage, perforation or death was described in the 12 RCTs which diagnosed ulcers on screening endoscopy. was 1 in 3.4 (29%). In one cohort study (22 patients only with a bleed or perforation) the ratio was 1 in 6 (16%). In the different types of reports (i.e. including observational studies) which documented progression from bleed or perforation to death, patients with a bleeding or perforated ulcer with or without NSAID exposure were described, of whom 1319 eventually died (Table 3). The mean likelihood of death from a bleeding or perforated ulcer was 1 in 8.3 (12%). This ratio varied from 1 in 35.5 (3%) in RCTs (2 deaths only), to 1 in 17 (6%) in casecontrol studies (161 deaths), 1 in 8 (12%) in case-series (571 deaths), and 1 in 6.3 (16%) in case reports (585 deaths). No such ratio could be calculated with data from cohort studies. We cannot say with certainty whether the range of fatality rates in observational studies (6±16%) relates to either different patient populations with different underlying risks or to different NSAIDs. In contrast to case-control studies (ratio 1 in 17), several uncontrolled series (ratio 1 in 8) reported on subgroups of patients who were at additional risk of death because of advanced age, concomitant disease, or because they had to undergo emergency surgery (Armstrong and Blower, 1987; Irvin, 1989; O'Riordain et al., 1990; Werbin et al., 1990; Doberneck, 1993). Case reports (ratio 1 in 6.3) may have exaggerated this risk even further because they concentrated on the description of disasters such as misdiagnosed cases (Cina et al., 1994) Combining the data ± The model (Fig. 2) In trials with screening endoscopy, no patient was reported with a more serious level of harm. These data could, therefore, not be used for the model. Data from patients with clinically diagnosed (symptomatic) ulcers overlapped with data from patients with bleed or perforation (Fig. 1), suggesting that the de nition of symptomatic ulcer included those patients whose ulcers only became apparent because they bled or perforated. The data from RCTs for bleed or perforation were, therefore, regarded as the clearer endpoint because they were not subject to such confounding. Thus, if the incidence of ulcer bleeding or perforation with chronic usage of NSAIDs, based on RCTs, is about 0.7% (Table 2), and between 6% and 16% (average 12%) of these patients will die (Table 3), then the NSAID death rate is 0:7 6% to 0:7 16% average 0:7 12% ˆ 0:042% to 0:122% average 0:084% Hence the number-needed-to-treat for one patient to die due to chronic usage of oral NSAIDs is 1= 0:112% 2 0:002% ˆ909 to 1= 0:042% 2 0:002% ˆ2500 average : 1= 0:084% 2 0:002% ˆ1220 This number-needed-to-treat for one patient to die may be veri ed by using the second, but less reliable starting point with drug exposure, namely symptomatic ulcer. If the incidence of a clinically diagnosed ulcer with chronic usage of NSAIDs, again based on RCTs, is about 1.5% (Table 2), and on average 30% (1 in 3.3, across all studies) of these ulcers will bleed or perforate (Table 3), and on average 12% (1 in 8.3, across all studies) of bleeding or perforated ulcers will be lethal (Table 3), then the average NSAID death rate is 1:5% 30% 12% ˆ 0:053% In this case the average number-needed-to-treat for one patient to die due to chronic usage of oral NSAIDs is 1= 0:053% 2 0:002% ˆ1960 This second calculation produces a very similar estimate to the rst and thus gives us con dence that the magnitude of the average number-needed-to-treat for one patient to die of 1220 is about right. 4. Discussion Knowing that the risk of NSAID-related gastroduodenal

11 M.R. TrameÁr et al. / Pain 85 (2000) 169± Fig. 2. Death due to chronic usage of oral NSAIDs: biological progression. complications is statistically signi cant (Chalmers et al., 1988; Gabriel et al., 1991), the clinically important question is: how often is the outcome fatal? A satisfactory answer as to how often clinically important ulcers, haemorrhage or perforation of ulcers, and death occur in patients who take oral NSAIDs regularly (Feldman, 1994) is only now emerging. A prospective cohort study in the USA estimated rates of hospital admission with gastrointestinal events, or death, in patients with osteoarthritis and rheumatoid arthritis (Singh and Triada lopoulos, 1999). Estimated annual death rates were 1 in 910 to 1 in 450 respectively, though based on fewer than 20 deaths. Large case-control studies (Blower et al., 1997) estimate 1 in Widely differing results are in part due to imprecise de nitions of drug exposure. We estimated the risk of death due to gastroduodenal complications with chronic ($2 months) NSAID exposure using information from various sources and levels of evidence Different study architecture ± different strength of evidence No clinician would doubt that NSAIDs cause gastroduodenal ulceration which may progress to bleeding, perforation and even death. However, calculating the incidences with which these complications occur is far from easy. Since RCTs reported only 2 deaths (0.2% of all deaths) there is a false sense of reassurance about the safety of NSAIDs. The reason for this is that individual RCTs are too small to estimate rare events accurately. Even one cohort study with patients exposed to NSAIDs reported only 10 deaths. Therefore, using these limited data only from RCTs and cohort studies to estimate the rate of death, and calculating a number-needed-to-treat to kill one patient of with a wide con dence interval extending from 9478± does not provide much help. In order to calculate the additional risk of death in patients with NSAID-induced ulcers it is necessary to turn to studies which document the mortality rates amongst large series of patients with bleeding or perforated ulcers. Observational studies provided valuable additional data on bleeding or perforation, and death due to NSAIDs. Eighty percent of all patients with bleeding or perforation (435 cases) and 99.8% of all deaths (1317 cases) were found in studies other than RCTs or cohort studies. Interestingly, the likelihood of having an ulcer going on to bleed or perforate was highly consistent (1 in 2.8 to 1 in 3.4) throughout all of these very different types of reports, with the exception of one cohort study which featured only 22 cases (Table 3). Hence RCTs and uncontrolled observational studies provide consistent data on the risk of developing a complicated (i.e. bleeding or perforated) gastroduodenal ulcer Potential sources of bias when combining the data When combining data from such varied studies, it is important to be aware of the potential sources of bias that may result. A number of biases exist in this study. It is important to note that all these biases are on the side of producing a more conservative estimate of harm from NSAIDs. Firstly, cohort studies generally involve large numbers of patients and enable comparisons to be drawn between patients receiving a drug and those who do not. However, unlike with an RCT, it is not possible to be certain of the extent of drug exposure. The cohort studies de ned NSAID

12 180 M.R. TrameÁr et al. / Pain 85 (2000) 169±182 exposure on the basis of patients redeeming prescriptions within certain times before the study period. Many of these patients may have received only single courses, incomplete courses or been non-compliant with NSAID therapy. Hence the `at risk' population is likely to be diluted by those who were in fact not receiving regular NSAIDs. Also, it is likely that general practitioners are taking account of risk factors when prescribing NSAIDs and consequently the drug cohort is a lower risk population than the control cohort (MacDonald et al., 1997). These factors reduce the apparent harm due to NSAIDs in a cohort study, and cohort studies indeed consistently reported signi cantly less risk with drug exposure compared with RCTs (Table 2). Secondly, the non-drug users in cohort studies may have purchased NSAIDs over the counter. Any non-prescription use of NSAIDs by the control group would lead to overestimation of the incidence of ulcer disease in the control population. The control death rate of 0.002% used in this study (Table 2), mainly based on data from one cohort study, is therefore likely to be an overestimation. This in turn affects the calculation of the number-needed-to-treat for one patient to die and, again, results in less harm being attributed to NSAIDs. Thirdly, observational studies produced a likelihood of dying from a complicated ulcer of between 1 in 6 to 1 in 17. Amongst these patients there will have been NSAID users and nonusers, and the conservative assumption is made that having bled or perforated, NSAID-using patients are at no greater or lesser risk than non-users (e.g. from excessive bleeding arising from platelet inhibition). There is evidence that NSAID users are, in fact, at greater risk. Among emergency admissions for upper gastrointestinal disease, NSAID users were more likely than non-users to need a blood transfusion, to need more blood (Blower et al., 1997), and 17% of those taking NSAID died compared with only 12% not taking an NSAID. Using an average gure for death rates is likely to underestimate the risk for NSAID users. Finally, we have de ned NSAID-induced damage on the basis of bleeding or perforated peptic ulcers and have excluded bleeding from other causes (e.g. tumours, gastritis, varices). This, again, has the effect of underplaying the burden of NSAID-induced harm What is the relevance of these results? The estimate of drug-related death is both surprising and alarming for two reasons. First, on average one in 1220 patients (range one in 909 to one in 2500) taking oral NSAIDs chronically will die due to gastroduodenal complications who would not have died had they not received the drug. These calculations suggest that chronic oral NSAIDs pose a substantial risk of death from gastroduodenal complications. This provokes the questions as to how this estimate of risk extrapolates to the general population, and how it compares with the results of other studies. Eight million people in the UK consult their family doctor each year with some arthritic disorder, four million of them aged more than 65 years (SCRIP, 1997). If only half of these take oral NSAIDs for more than 2 months, then 2 million people are at risk. If for this population the risk of death attributable to NSAID-induced ulcer bleeding or perforation is estimated to be 1 in 1000 (long-term treatment in elderly patients with a history of gastroduodenal lesions), then 2000 people will die each year in the UK. This number comes very close to a recently published estimate from a retrospective case control study of about 2500 deaths per annum in the UK resulting from peptic ulcer complications in NSAID users (Blower et al., 1997). The similarity of these two estimates, which were derived from two completely different mathematical approaches, increases con dence in their validity. These numbers will have profound implications for prescribing. Second, the magnitude of our estimate suggests that an RCT would need 3705 patients in both experimental and control groups to be 95% con dent that drug-related death will happen at least once (Hanley and Lippman-Hand, 1983). Only one trial ful lled this criterion (Silverstein et al., 1995). This trial reported one death in the experimental (i.e. NSAID) group (and none in the control group), a difference which was not statistically signi cant. This demonstrates the need for large numbers, and in this case the power of meta-analysis. Our model describes a way to synthesise data with different strength of evidence from reports of different study architecture to gain more insight into the frequency of rare adverse events. It is very likely that the risk of gastroduodenal complications with NSAIDs is dependent on patient related factors such as age or concomitant disease. The original trials did not allow such risk factors to be taken into account. Also, there is the likelihood that different NSAIDs and different doses may be associated with different risks (Rodriguez and Jick, 1994). Again we could not explore differences between the NSAID drugs within this data set. The effect of dose, however, was evident. With aspirin, the absolute risk of a bleed or perforation increased with increasing dose (Table 2.3.), a relationship which has been shown previously (Roderick et al., 1993). In conclusion, to estimate the risk of rare serious complications, large numbers and bias-free assessment are needed. For adverse events which follow a biological progression our model proposes to estimate the more common complication from RCTs, then to estimate the ratio of the common to rare complications from observational studies, and nally to combine these estimates to quantify the rate of the rare complication. Quantitative analyses of data from systematically searched reports with different study architecture suggest strongly that chronic oral NSAIDs pose a high risk of minor and major harm and death. Patients should receive minimum effective doses for the minimum possible time.