Aspirin for prophylactic use in the primary prevention of cardiovascular disease and cancer: a systematic review and overview of reviews

Description of health problem (primary prevention of cardiovascular disease and cancer)

Cancer and CVD exert a heavy burden on the UK population in terms of morbidity, mortality and cost. Primary prevention measures have a large potential impact on these burdens. Some guidelines and investigators have proposed that regular use of aspirin might be effective in this regard. However, some individuals who take aspirin experience unpleasant side effects that occasionally may be life-threatening. This short report aims to review and examine the relevant evidence.

In this report we interpret primary prevention as defined for CVD by the National Institute for Health and Care Excellence (NICE) as follows ‘. . . interventions that aim to prevent CV [cardiovascular] events in people who have no clinical evidence of CVD’. 1 A similar definition may be used for primary prevention of cancer; we employ a corresponding definition by substituting ‘cancer’ for ‘CVD’ in the above statement.

Epidemiology

Cancer in the UK

According to Cancer Research UK, more than one in three people in the UK will develop some form of cancer during their lifetime. In 2009, approximately 320,000 people were diagnosed with cancer in the UK. Cancer is predominantly a disease of older people; > 63% of cancer diagnoses are in people aged ≥ 65 years, and 36% in those ≥ 75 years (Figure 1). The European age-standardised rate is higher in males than in females (429 per 100,000 vs. 372 per 100,000, respectively). 8 In the UK, around 26% of all deaths are caused by cancer. Lung cancer causes the greatest proportion of deaths in the UK (22%) followed by colorectal cancer (CRC) (10%) and breast cancer (8%). 8

FIGURE 1.

Age-specific average incidence of cancers/year (excludes non-melanoma skin cancer). Redrawn from Cancer Research UK.

Current service provision

Aspirin and cancer

Studies have suggested that aspirin reduces cancer risk, especially CRC; however, the mechanism is unknown. 30,31 Dovizio et al. 32 hypothesised that the role is also likely to involve platelets. It is thought that activated platelets may enhance the metastatic potential of cancer cells. This may occur through a direct interaction and/or the release of soluble mediators, seemingly due to the overexpression of cyclo-oxygenase 2 (COX-2), largely found in inflammatory cells and inducible with mitogens, growth factors and tumour promoters. 33 Therefore, aspirin, as an inhibitor of this enzyme, could consequently reduce metastasis. COX-independent mechanisms of aspirin, such as the inhibition of signalling and the acetylation of extra-COX proteins, have also been suggested to play a role in its putative effect against cancer.

Circulating aspirin is rapidly de-acetylated to release salicylate. Recently, the released salicylate group has been considered to have its own independent anti-inflammatory effects via accelerated polymorphonuclear apoptosis, resulting in an anti-inflammatory effect. 34

Adverse events

There is a well-documented increased risk of major and minor bleeds and gastric discomfort associated with aspirin use. COX-1 produces prostaglandins that are involved in physiological protection of the gastric mucosa. 35 Aspirin's inhibition of COX-1 suggests a mechanism for unwanted side effects. Efforts to avoid gastric problems have included development of coated or buffered tablets and of NO (nitric oxide)-aspirin, which releases nitric oxide that could counteract the undesired influence of inhibited generation of protective agents. Interestingly, the incidence of gastrointestinal (GI) bleeding from taking low-dose aspirin for a long time appears not to be influenced by the use of enteric-coated compared with buffered aspirin, although these preparations may decrease side effects of gastric irritation and the slow release may be helpful for night pain. 36

Inclusion of relevant studies

Titles and abstracts of retrieved studies were examined for inclusion by two reviewers independently. Disagreement was resolved by retrieval of the full publication and consensus agreement, with further discussions with a third reviewer if agreement was not obtained. The following inclusion criteria were used:

Exclusion of studies

All designs other than RCTs, systematic reviews or meta-analyses were excluded. Also excluded were systematic reviews or meta-analyses that included only secondary prevention or those in which primary prevention could not be separately identified. Reviews that included only observational studies, and studies not in the English language, were also excluded.

Quantifying absolute benefits and harms

The number of unwanted events (e.g. all-cause mortality) averted by taking aspirin, and the number of adverse events (e.g. bleeding) incurred from aspirin use, are best calculated using IPD, taking into account the person-years of exposure to aspirin. However, IPD is not available from study-level meta-analyses. There are various ways of calculating the rate of averted or of incurred events from study-level data. We used two methods, described below, and have compared the results across systematic reviews according to outcome.

In the ‘aggregated’ method the aggregated number of events (i.e. sum) across all included trials is divided by the aggregated number of persons. This is done separately for each arm (aspirin and control) to calculate ‘events/person’ (E/p). The weighted average follow-up time across all included trials [mean follow-up (MFU)] (for the intervention arm often equivalent to years of exposure) was calculated as:

where MTFU = mean follow-up in each trial and PT = total participants in each trial.

The difference between arms then generates the ‘events averted/person-year of follow-up or the extra events incurred/person-year of follow-up.

Because these numbers are small, we normalised the results to (1) patients-years’ exposure required for one fewer event or for one extra event and (2) number of events averted or extra events incurred should 10,000 patients be followed up for 10 years.

For the ‘pooled’ method we used the random-effects pooled risk of event for the control arm (CR). If the systematic review reported pooled odds ratio (ORp) for the outcome then the calculation proceeds as:

Again, because this number is small, we normalised the results to (1) ‘patient-years' follow-up required for one fewer event’ or for one extra event and (2) number of events averted or extra events incurred should 10,000 patients be followed up for 10 years.

It has been suggested that the risk observed in the largest available trial may offer a suitable control risk estimate for the number needed to treat (NNT) calculations that are based on study-level meta-analyses; in the face of considerable heterogeneity in control rates, this method was not adopted here because the largest trial for many outcomes was the Women's Health Study (WHS),46 which was atypical in having an alternate-day dose regimen, a 100% female population and the longest follow-up period.

Number of studies identified

The flow chart outlining the process of identifying relevant literature can be found in Figure 2. Following the removal of duplicates, the searches identified 2572 potentially relevant papers. A total of 2425 papers did not meet our inclusion criteria and were removed at title and abstract sift, leaving a total of 147 papers to be further investigated. Of these, 120 were removed at full-paper sift, resulting in 27 papers that met the inclusion criteria.

FIGURE 2.

Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) flow diagram. (a) Of the 27 included publications: CVD, systematic reviews = 9, RCTs = 3; cancer, systematic reviews = 6; diabetes, systematic reviews = 7, RCTs = 2; (b) one paper was identified from assessment of reference lists of excluded papers; this had been excluded at abstract sift but was not considered relevant until reading the paper in full.

A search of the UKCRN Portfolio and ClinicalTrials.gov databases retrieved 824 potential trials. The search strategies used can be viewed in Appendix 1. After screening by title, 12 trials were identified, two of which had already been identified via the database searches. Appendix 2 describes these 10 included trials; all have either recently finished or are ongoing.

Number of studies excluded

A list of the 121 papers that were excluded at full-paper sift is provided in Appendix 3, with reasons for exclusion. The main reason for excluding a paper at full-paper sift was because it was considered to be a non-systematic review (n = 52).

Quality assessment

The 27 included studies were assessed using standardised or modified quality assessment tools. Systematic reviews of the prophylactic use of aspirin in the primary prevention of CVD (n = 9), cancer (n = 6) and CVD in patients with diabetes (n = 7) were assessed using a modified tool developed by NHS CRD. 50 RCTs concerning CVD (n = 3) and CVD in patients with diabetes (n = 2) were quality assessed using the Cochrane Risk of Bias tool. 51 The systematic reviews and RCTs were in general very highly rated, representing the high quality of work that has previously been undertaken. Summaries of quality assessment ratings in relation to study design and disease area are provided in Tables 2–6. For further details on the quality assessment of included papers, see Appendix 5.

The three ‘new’ RCTs identified in our searches relating to CVD (see Table 7) and the two relating to CVD in diabetes added no new evidence to that already included in the systematic reviews identified. The CVD RCTs included (1) a post hoc analysis of the WHS46 to model treatment effect for individual patients; (2) a core RCT from the nine previously included in systematic reviews; and (3) a pilot RCT. Both RCTs relating to diabetes were from the core nine trials40–48 previously included in systematic reviews. The searches relating to cancer revealed no new RCTs. Detailed data extraction tables of systematic reviews/meta-analyses involving patients with CVD, cancer and CVD in patients with diabetes are provided in Appendix 6.

Cardiovascular disease

Table 7 provides a summary of characteristics of the included systematic reviews and RCTs of aspirin for the primary prevention of CVD. All the papers provided a clear aim. The level of detail in the methods varied across the papers, with a number of studies not reporting (1) the search strategy (n = 3), (2) inclusion criteria (n = 3) and (3) quality assessment (n = 5). A broad range of outcome measures were reported across the included papers. The majority of the included systematic reviews did not clearly distinguish between primary and secondary outcomes and there was some lack of clarity about what was considered an adverse event (e.g. haemorrhagic stroke, GI bleed, major bleed). See Appendix 6 for further details.

The nine systematic reviews reported in Table 7 consistently report on nine (or a subset of the nine) RCTs, depending on the year that their meta-analysis was undertaken. The RCTs are POPADAD (Prevention of Progression of Arterial Disease And Diabetes),40 BDT (British Doctors Trial),45 JPAD (Japanese Primary Prevention of Atherosclerosis with Aspirin for Diabetes),44 AAA (Aspirin for Asymptomatic Atherosclerosis),42 HOT (Hypertension Optimal Treatment),43 TPT (Thrombosis Prevention Trial),48 PPP (Primary Prevention Project),41 PHS (Physician's Health Study)47 and WHS. 46 These RCTs have been repeatedly subject to meta-analysis of outcomes (Table 8).

The majority of these nine RCTs, with the exception of three,40,42,44 were outside the search dates for the current short report. Table 9 provides a summary of the aspirin dose and participant characteristics of these nine RCTs.

The latest meta-analysis, published by Seshasai et al. ,56 was a study-level meta-analysis that carefully assessed the risk of bleeding (total bleeds and major bleeds). The review team also identified an IPD meta-analysis by Baigent et al. 53 examining outcomes and patient subgroups (according to age, sex, diabetes, smoking, mean BP, blood cholesterol, body mass index) in more detail than in a study-level meta-analysis.

In the following section we evaluate the four most recent meta-analyses further,37,38,53,56 including the systematic reviews by Baigent et al. 53 and Seshasai et al. 56 We will also refer to the recent high-quality review of reviews by Raju et al. ,38 who examined the findings from these meta-analyses. 54 See Appendix 6 for detailed data extraction of these two studies and a short-form extraction of the remaining studies included.

The paper by Seshasai et al. 56 was the most recent and highest-quality study-level meta-analysis concerned with aspirin for the primary prevention of CVD. Seshasai et al. reported meta-analyses for nine outcomes based on nine RCTs (see Table 9) published between 1988 and 2010 encompassing 102,621 individuals. 40–48 Seshasai et al. 56 stated that ‘. . . primary efficacy endpoints were total CHD and total cancer mortality’. Adverse events were classified as non-trivial bleeds and all bleeds. Haemorrhagic stroke was not selected as an outcome.

Baigent et al. 53 used IPD to analyse the effects of aspirin compared with placebo according to baseline risk of CVD. The meta-analyses by Bartolucci et al. ,37 Raju et al. 38 and Seshasai et al. 56 pooled data from the same nine RCTs (see Table 9) but included different overall numbers of participants: 100,038,37 100,07638 and 102,621. 56

Raju et al. 54 considered the differences among recent meta-analyses in terms of reported samples. First, it was recognised that Seshasai et al. 56 included 2545 warfarin-treated patients, whereas Bartolucci et al. 37 and Raju et al. 38 excluded these patients. Second, Bartolucci et al. 37 excluded 60 patients for reasons that were unclear.

In terms of methods of reporting the pooled data, Seshasai et al. 56 and Bartolucci et al. 37 both reported pooled treatment effects using the odds ratio (OR). In contrast, Raju et al. 38 reported relative risk (RR). Raju et al. 54 considered that these differences would not make substantial changes to the interpretation of the findings because the OR approximates the RR when event rates are low. Baigent et al. 53 performed IPD analysis using rate ratio (RaR) (events per unit time aspirin/events per unit time control) as the major outcomes statistic.

Cancer

We identified six systematic reviews assessing the effect of aspirin on cancer mortality and cancer incidence with publication dates ranging from 2010 to 2012 (Table 10). All these reviews used RCTs in which the primary outcome was not cancer. Instead the reviews considered trials in which the primary outcomes were primary or secondary prevention of CVD, which were retrospectively used to follow-up cancer deaths. Five of the six reviews we identified were analyses by a group led by Rothwell. 22,31,49,61,62 The most recent highest-quality systematic review from these six publications was fully data extracted and summarised (see Appendix 6).

Using the NHS CRD50 assessment criteria, the quality of the papers by Rothwell and colleagues was generally rated as high,22,31,49,62 (see Quality assessment, above).

Rothwell and colleagues49 suggest that some trials had reported deaths due to cancer which was diagnosed prior to randomisation. In this case conclusions on the primary prevention of cancer should be drawn with caution.

Cardiovascular events in patients with diabetes

Table 11 provides details and publication dates of the seven identified systematic reviews meta-analysing the effect of aspirin in the primary prevention of CVD events in patients with diabetes.

Seshasai et al. 2012

Primary prevention of cardiovascular disease

The authors stated that the ‘. . . primary efficacy endpoints were total CHD, and total cancer mortality’. 56 Total CHD comprised major cardiovascular events (MCEs) defined as the composite of non-fatal MI, non-fatal stroke or CV death. The reported pooled random-effects OR was 0.86 [95% confidence interval (CI) 0.74 to 1.01] (Figure 3).

FIGURE 3.

Meta-analysis of total CHD. 56

Only one trial (PHS)47 reached statistical significance at p = 0.05. This trial was the most influential for the pooled estimate (Figure 4); when omitted from the analysis the effect size diminishes considerably and the pooled upper confidence interval (UCI) encompasses a worse result for aspirin than for the comparator.

FIGURE 4.

(a) Influence of individual studies on pooled OR for total CHD. (b) L'Abbé plot showing total CHD event rates [dashed line = OR of 1; solid line = pooled OR (random effects)].

The event rate across studies varied considerably (see Figure 4a). The risk in control groups ranged from 1% (WHS)46 to 12.8% (POPADAD)40 (Figure 5).

FIGURE 5.

Total CHD in control arm; trials arranged according to follow-up (years). LCI, lower confidence interval.

Repeated test meta-analysis according to recruitment year indicated that statistical significance in the pooled RR was reached with the inclusion of the PHS study,47 after which with addition of further studies the pooled estimate tended to diminish (Figure 6).

FIGURE 6.

Cumulative meta-analysis: OR for total CVD (data from Seshasai et al. 56); studies arranged according to recruitment period.

The NNT based on a control group risk of 3.249% (random-effects pooled estimate) and OR of 0.86 calculates to 226. Taking the mean follow-up as 6.9 years, this indicates about 64 fewer events among 10,000 persons followed for 10 years. Seshasai et al. 56 reported a reduced event rate of 100/100,000 person-years.

Primary prevention of cancer

Seshasai et al. 56 also identified total cancer mortality as a primary outcome (Figure 7). The pooled OR from eight RCTs was 0.93 (95% CI 0.84 to 1.03) in favour of aspirin.

FIGURE 7.

Odds ratio for cancer mortality (Seshasai et al. 56).

No study alone reached statistical significance and no single study was greatly influential for the pooled estimate (Figure 8a). The event rate varied across an approximate sixfold range across studies (see Figure 8b) in a manner partly explained by differing length of follow-up as indicated for the control group in Figure 9.

FIGURE 8.

(a) Influence of individual studies on pooled OR for cancer mortality. (b) L'Abbé plot showing total cancer mortality event rates [dashed line = OR of 1; solid line = pooled OR (random effects)].

FIGURE 9.

Cancer mortality in control arm; trials arranged by length of follow-up (years). LCI, lower confidence interval.

Based on pooled estimates (OR 0.93 and control risk of 2.155%) and the mean follow-up of 7.1 years, the calculated NNT of 677 suggests that very large numbers of people would need to be treated to prevent one event, equivalent to 21 cancer deaths averted among 10,000 persons followed up for 10 years.

Repeated test meta-analysis according to recruitment year (Figure 10) indicated that after the early BDT45 RCT, the inclusion of subsequent studies pulls the pooled OR towards a null effect for aspirin. The pooled OR failed to reach statistical significance (p < 0.05) at any time.

FIGURE 10.

Repeated test with accumulating studies arranged by recruitment year (cancer mortality).

Berger et al. (2011)

The primary outcome in this meta-analysis was a composite major cardiovascular event (MCE) end point, which included non-fatal MI, non-fatal stroke or CV death. The reported pooled RR was 0.90 (95% CI 0.85 to 0.96) (Figure 11), indicating a 10% risk reduction from aspirin.

FIGURE 11.

Meta-analysis of MCEs (Berger et al. ). 39

Only the PHS47 reached statistical significance (p < 0.05), and this was the single most influential study for the pooled estimate (Figure 12b). The event rates varied considerably across trials (see Figure 12a). In the control arms the risk of MCE varied between 2.6% and 10.2% (Figure 13).

FIGURE 12.

(a) Influence of individual studies on pooled RR for total bleed events. (b) L'Abbé plot showing bleed event rates [dashed line = RR of 1; solid line = pooled OR (random effects)].

FIGURE 13.

Major CV events in control arm [studies arranged according to follow-up (years)]. LCI, lower confidence interval.

Repeated test meta-analysis according to recruitment year, indicated that statistical significance in the pooled RR was reached with the inclusion of the PHS47 study, after which with addition of further studies the pooled estimate tended to slightly decrease but remained statistically significant at p < 0.05 level (Figure 14).

FIGURE 14.

Repeated test with accumulating studies arranged by recruitment year (MCE).

Berger et al. 39 reported an NNT of 253 and a mean follow-up of 6.9 years; this yields 57 fewer events for 10,000 persons followed for 10 years. Using a pooled estimate of risk for the control group of 5.84%, the pooled RR of 0.9 and mean follow-up of 6.9 years, a higher value of 84 events avoided is estimated for 10,000 persons followed for 10 years.

Raju et al. (2011)

In this meta-analysis, the primary outcome was specified as part of the objective of the study which was stated as: ‘to obtain best estimates of the effect of aspirin on mortality in primary prevention’. The pooled RR of death was 0.94 (95% CI 0.88 to 1.00) indicating a modest 6% reduced risk with aspirin use (Figure 15).

FIGURE 15.

Relative risk for all-cause mortality (data from Raju et al. 38).

No study alone reached statistical significance. The 95% UCI for the pooled RR encompassed no effect. The range in event rate varied considerably across both arms of the trials (Figure 16a). No single study was particularly influential for the pooled estimate; however, omitting any one of BDT,45 PPP,41 HOT43 or WHS46 from the analysis moved the 95% UCI beyond RR of 1.0 (see Figure 16b). The risk of death in the control arms varied from 2% in PHS47 to 16% in the POPADAD40 study (Figure 17). Cleland71 commented on the Raju meta-analysis,38 stating that, despite 100,000 years of follow-up, aspirin prevented only 21 deaths. In fact, in aggregating across studies, 21 more deaths are recorded in the aspirin group (1859/50,868 vs. 1838/49,208). However, aggregation by group breaks randomisation and, in this instance, underestimates the contribution of the influential BDT45 study, in which participants were randomised 2 : 1 aspirin–comparator.

Using the ‘aggregated’ and ‘pooled’ methods for calculating absolute benefit yielded estimates of 36 and 46 events avoided for 10,000 persons followed up for 10 years. Estimates of 33 and 46 events can be calculated using data from Berger et al. ,39 who also meta-analysed all-cause mortality from the same set of nine trials.

FIGURE 16.

(a) L'Abbé plot showing all-cause mortality event rates [dashed line = RR of 1; solid line = pooled RR (random effects)]; (b) Influence of individual studies on pooled RR for all-cause mortality.

FIGURE 17.

All-cause mortality event rate in control arms (data from Raju et al. ). 38 LCI, lower confidence interval.

Repeated test meta-analysis according to recruitment year indicated that statistical significance in the pooled RR was not convincingly reached at any time; the pooled point estimate remained stable with the addition of further studies (Figure 18).

FIGURE 18.

Cumulative meta-analysis of RR for all-cause mortality; studies arranged according to recruitment period. [NB-PPP (2001)41 and HOT (1998)43 had comparable estimated mid-point recruitment periods (i.e. 1993).]

Bartolucci et al. (2011)

The primary outcome was not explicitly defined but ‘total coronary heart disease’ was categorised as prespecified ‘outcome 1’ and was defined as ‘non fatal and fatal MI and death due to CHD’; it was not clear if non-fatal stroke was omitted from this outcome because event and patients numbers were not included in the publication. 37 Pooled OR for this outcome was 0.854 (95% CI 0.688 to 1.061) (Figure 19).

FIGURE 19.

Odds ratio for total CHD (data from Bartolucci et al. 37).

Repeated test meta-analysis according to recruitment year indicated that statistical significance in the pooled OR was achieved with inclusion of the four earliest studies; the pooled estimate tended to lower effect size with the addition of subsequent later studies and statistical significance was lost (Figure 20).

Because of the lack of event numbers and patient numbers reported, further exploration of this meta-analysis has not been undertaken.

FIGURE 20.

Cumulative meta-analysis of OR for total CHD; studies arranged according to recruitment period.

Antithrombic Trialists' Collaboration (individual patient data)

The implicit primary outcome in the ATT Collaboration53 (IPD) meta-analysis was the risk of any serious vascular event (a composite of MI, stroke or CV death). The pooled RaR (yearly event rate) for the six included RCTs was 0.88 (95% CI 0.82 to 0.94). Data for individual trials for this outcome were not reported. The absolute difference in rates (aspirin minus control) was reported as 0.07%/person-year (70 events avoided for 10,000 persons followed up for 10 years). The RaR for ‘any major coronary event’ (non-fatal MI or CHD death) was 0.82 (95% CI 0.75 to 0.90) (Figure 21) and the corresponding absolute rate difference 0.055% (55 events avoided for 10,000 persons followed up for 10 years). The RaRs for any vascular death and for CHD death were 0.97 (95% CI 0.87 to 1.09) and 0.95 (95% CI 0.82 to 1.10), respectively, indicating a lack of statistically significant benefit from aspirin for these outcomes; the absolute rate difference for CHD death was 0.01%/person-year, equivalent to only 10 events averted among 10,000 persons followed up over 10 years. Overall stroke mortality (including both haemorrhagic and ischaemic stroke) was worse in the aspirin than the comparator group (RaR 1.23, 95% CI 0.84 to 1.74). The event rates for all-cause mortality (any death) were reported as 0.5% (aspirin) and 0.53%/person-year (control); a difference of 0.03% represents 30 deaths avoided should 10,000 persons be followed up for 10 years.

FIGURE 21.

Meta-analysis of IPD for any major coronary event reported by the ATT Collaboration. LCI, lower confidence interval.

Selak et al. (2010)

Selak et al. 55 undertook a modelling study using ATT53 IPD data so as to assess the balance of harm and benefit in primary prevention of CVD. The expressed aim was to ‘interpret results in the light of current New Zealand CVD risk assessment and management guidelines’. The ATT outcome chosen to represent benefit was ‘any serious vascular event’ (a composite of MI or stroke or CV death) for which the Trialists' IPD analysis estimated an annual RaR of 0.88 (95% CI 0.82 to 0.94) assumed by Selak et al. 55 to equate to a 12% reduction in events over 5 years (table 1 in Selak et al. ,55 reproduced below in Table 12). Selak et al. 55 applied this 12% reduction across 5 years to hypothetical populations of 1000 men or women in different 10-year age bands (50–59 years to 80–89 years) for whom the number of expected events per 5 years without aspirin was based on the 5-year risk of a serious CV event predicted using the Framingham equation. For example, for 1000 individuals (any age group or sex) whose 5-year risk of a CV event is 10%, the expected number of events is 100 in 5 years; applying the assumed 12% reduction delivers 12 events avoided in 5 years (i.e. over 5000 person-years; table 1 in Selak et al. 55). This absolute reduction in events of 12/5000 person-years is considerably larger than the absolute reduction reported by the ATT (0.07%/person-year, or 3.5 events averted over 5000 person-years). The only cohorts of 1000 persons for whom the predicted number of events averted was not greater than the overall rate reported in the ATT meta-analysis were those with the lowest 5-year risks of 1% or 2%.

It is worth noting that most individuals included in the ATT analysis had an estimated 5-year risk of < 2.5%.

Selak et al. 55 selected the ATT outcome ‘non-fatal extracranial bleeds’ to represent harm. For this outcome, the ATT IPD analysis reported 554 events over 660,000 person-years (0.08%/person-year or 4.2 events per 5000 person-years), a risk ratio of 1.54 (95% CI 1.30 to 1.82) and a relative rate of 1.98 greater for men than women, and 2.15 greater for each decade beyond 50–59 years. The absolute rates (rounded to two decimals) for aspirin and control groups were 0.10%/person-year and 0.07%/person-year, respectively (equivalent to 0.03 extra events per 100 person-years or 1.5 events per 5000 person-years). Selak et al. 55 allocated one additional (aspirin dependent) event for 1000 women (age band 50–59 years) over 5 years (see Table 12). This was multiplied by 1.98 for men and by 2.15 for each sex according to ATT data (see Table 12, bottom row).

Assuming equivalence of desirability for beneficial and harmful events the balance between benefit and harm favoured benefit for most risk groups under the age of 80 years (represented by the unshaded cells in Table 12).

After considering additional factors, including the availability and effectiveness of other primary prevention measures (especially lipid-lowering with statins), the authors concluded that, in New Zealand, aspirin should be considered for primary prevention of CVD in those persons with five-year CVD risk of ≥ 15%, up to the age of 80 years. This represents only 13% of the New Zealand primary prevention population. 72 A recent UK study of a primary prevention population from England and Wales (encompassing 750,232 individuals, and 2,969,311 person-years of observation, mean age 48 years) estimated that about 10% had a 10-year risk of CVD of ≥ 20%. 3 This similarly implies that aspirin use for those with a 5-year risk of ≥ 15% and up to 80 years of age (suggested by Selak et al. 55) would in practice involve only a small proportion of the primary prevention population in England and Wales.

The crude incidence rate of CVD events [a diagnosis of CVD including angina, MI, stroke or transient ischaemic attacks (TIAs), but not PVD] in the Hippisley-Cox et al. study3 was reported to be 0.73%/person-year for men and 1.05%/person-year for women; remarkably, this is about 10 times greater than that observed in the ATT study. 53

Further relevant systematic reviews

Adelman et al. 52 undertook a study-level subgroup analysis by sex. The primary outcome for this meta-analysis was stroke. The reported OR was 0.83 (95% CI 0.70 to 0.97) for women (based on the HOT,43 PPP41 and WHS46 trials) and 1.13 (95% CI 0.96 to 1.33) for men based on BDT,45 HOT,43 PHS,47 PPP41 and TPT48 studies. These results were reproduced from Berger's 2006 meta-analysis. 70

Rothwell et al. (2010)

Rothwell et al. 31 used IPD to look at the incidence and mortality of CRC over 20 years of follow-up in RCTs that were aimed at investigating effectiveness of aspirin in CVD prevention. Eligible trials needed to have recruited at least 1000 participants and a median scheduled treatment period of at least 2.5 years. Two CVD primary prevention trials were included: TPT48 and BDT,45 and two CVD secondary prevention trials (SALT, Swedish Aspirin Low Dose Trial;73 UK-TIA, Transient Ischaemic Attack trial74) in which patients had experienced a TIA or mild stroke before recruitment. Those patients in these trials who were free of cancer at the start of treatment could be construed as suitable individuals in whom to test a hypothetical benefit of aspirin in the primary prevention of CRC. However, it should be borne in mind that post hoc investigation of a post hoc outcome for a subgroup selected post hoc may compromise the continuing integrity of the randomisation. The study's design relies on correct ascertainment, and recording, of cause of death (COD) in the post-trial period.

The pooled OR for death from CRC was 0.66 (95% CI 0.52 to 0.85) (Figure 22), indicating a one-third reduction in risk from aspirin use. However, the large WHS46 and PHS47 trials were not included in these analyses because alternate-day dose regimens were excluded (WHS46 and PHS47 100 mg and 325 mg aspirin, respectively, on alternate days). Subgroup analysis according to aspirin dose hinted at greater effectiveness for aspirin at lower dosages (75–300 mg daily compared with 500–1200 mg daily). Further subgroup analysis indicated that aspirin was most effective in preventing deaths from colorectal tumours in the proximal colon, and was relatively less effective for rectal tumours. Taking the reported event rates as 119/8282 aspirin users and 121/5751 non-users and a follow-up of 20 years the number of colorectal deaths avoided should 10,000 persons be followed for 10 years is 34. Using the pooled OR of 0.66, taking follow-up as 20 years and the random-effects pooled control risk of 0.0217, the estimated number of deaths avoided, should 10,000 persons be followed for 10 years, is 36.

FIGURE 22.

Representation of results as reported by Rothwell et al. ,31 stratified by dose regimen. Note: The authors applied a correction to allow for the fact that the control group from UK-TIA74 was used twice in the analysis.

The hazard ratio (HR) for CRC incidence (irrespective of scheduled duration of aspirin use) was reported as 0.76 (95% CI 0.63 to 0.94) (397 events). For low-dose aspirin (196 events out of 8073 participants) the HR was reported to be 0.75 (95% CI 0.56 to 0.97). It is worth noting that the OR for CRC incidence reported over considerable follow-up in the WHS46 and PHS47 trials was 1.07 (95% CI 0.75 to 1.53) and 0.97 (95% CI 0.77 to 1.24), respectively. If HR is equated to OR then inclusion of these trials generates an overall pooled estimate (OR) of 0.91 (95% CI 0.74 to 1.11) (Figure 23).

FIGURE 23.

Odds ratio/hazard ratio for incidence of CRC including WHS46 and PHS47 trials.

Rothwell et al. (2011)

Rothwell et al. 22 analysed cancer deaths in eight trials40,42,44,45,48,74–76 of aspirin for primary or secondary prevention of CVD; these were BDT,45 UK-TIA,74 ETDRS (Early Treatment Diabetic Retinopathy Study),75 SAPAT (Swedish Angina Pectoralis Aspirin Trial),76 TPT,48 JPAD,44 POPADAD40 and AAA42 trials. Seven of these were available for IPD analysis. 40,42,44,48,74–76 Eligible trials had a median or mean treatment period of at least 4 years and a range extending beyond 5 years. For inclusion, aspirin had to be given without a second agent or, if given with another agent, both needed to be used in the same way in both trial arms; trials of aspirin at any dose for primary or secondary CVD prevention could be included.

The within-trial pooled OR for death from cancer (eight trials40,42,44,45,48,74–76) was 0.79, 95% CI 0.68 to 0.92 (main paper, slightly different in supplementary appendix because of reclassification of some deaths). All of the trials but one75 had reduced risk for the aspirin group; however, only the UK-TIA study74 reached statistical significance (Figure 24).

FIGURE 24.

Odds ratio of death from cancer in eight RCTs40,42,44,45,48,74–76 of primary or secondary prevention of CVD. Note: Based on data provided in the supplementary web appendix.

There was little statistical heterogeneity among the studies. Using the aggregated data from the eight studies40,42,44,45,48,74–76 to estimate the number of cancer deaths avoided requires a value for mean follow-up; taking 10 years as mean follow-up then 66 cancer deaths would be averted for every 10,000 persons receiving aspirin. If the pooled random-effects control rate (Figure 25) is used for the calculation then 54 deaths are averted.

FIGURE 25.

Risk of cancer death in the control arm in eight RCTs40,42,44,45,48,74–76 CVD primary or secondary prevention. LCI, lower confidence interval.

The OR for all-cause mortality for the eight trials40,42,44,45,48,74–76 was 0.92 (95% CI 0.85 to 1.00) (Figure 26).

FIGURE 26.

Odds ratio of death (any cause) in eight RCTs40,42,44,45,48,74–76 of primary or secondary prevention of CVD.

Using the aggregated data, assuming 10 years' mean follow-up, the number of deaths (any cause) averted is estimated to be 85 among 10,000 persons who received aspirin; using the pooled control risk of 10.4% (Figure 27) the averted deaths amount to 75. According to these estimates, approximately three-quarters of all of the deaths averted because of aspirin use are attributable to avoidance of cancer.

FIGURE 27.

Risk of death (all-cause) in the control arm in eight RCTs of CVD primary or secondary prevention. LCI, lower confidence interval.

In IPD analysis of seven trials,40,42,44,45,48,74,75 the risk of death from cancer during trial treatment was reduced in the aspirin group (HR 0.82, 95% CI 0.70 to 0.95; p = 0.01); benefit became apparent after about 5 years, after which the percentage risk in the control group increased more rapidly than for the aspirin group. These results are represented in Figure 28.

FIGURE 28.

Risk of cancer death during treatment period (IPD analysis) (data taken from Rothwell et al. 22).

Further analyses were stratified according to type/site of tumour (for details see Appendix 6). Three trials45,48,74 provided IPD for analysis up to 20 years; HRs (aspirin vs. control) indicated reduced risk for small-cell lung cancer, adenocarcinoma of the lung and of the oesophagus but not for squamous cell lung cancer (for details, see Appendix 6). In three trials45,48,74 with prolonged follow-up the HR for all-cause mortality over 20 years was 0.96 (95% CI 0.90 to 1.02; p = 0.37).

Rothwell et al. (2012)

In this IPD study,49 the risk of cancer death and cancer incidence, and their time courses, were examined in 51 RCTs in which daily aspirin was investigated for the prevention of CVD (vascular events). Eligible trials randomised participants to daily aspirin (any dose) compared with no aspirin (studies of aspirin in association with anticoagulation were also eligible). Secondary cancer prevention trials and studies of duration < 90 days were excluded. Twelve of the included trials were classified as primary prevention studies and 39 as secondary prevention. A primary outcome was not specified.

The risk of cancer death in the 51 RCTs is shown in Figure 29. The pooled OR for cancer death was 0.84 (95% CI 0.74 to 0.94; p = 0.002). Figure 30 shows the risk of cancer death in the control arms.

FIGURE 29.

Risk of cancer death in 51 RCTs of CVD prevention. ST_CDs, cancer deaths in 21 small studies; ST_NVDs, non-vascular deaths in 17 small studies. Meta-analysis adjusted for 2 : 1 randomisation ratio used in some trials. 49 EAFT, European Atrial Fibrillation Trial.

FIGURE 30.

Risk of cancer death in control arms (based on Rothwell et al. 49). LCI, lower confidence interval. EAFT, European Atrial Fibrillation Trial.

The two large primary prevention studies PHS47 and WHS46 in which aspirin was administered on alternate days and where outcomes were likely less favourable for aspirin did not satisfy the inclusion criteria.

The number of cancer deaths averted should 10,000 persons be followed for 10 years is estimated to be 25 (assuming mean follow-up for the 51 RCTs was 10 years) and 36 (assuming mean follow-up was 7 years) according to the ‘aggregated’ method, and 31 and 44 (follow-up assumed to be 10 and 7 years) according to the ‘pooled’ method.

Cancer incidence

Rothwell et al. 49 also reported the OR for cancer incidence beyond 3 years in six CVD primary prevention trials;40–44,48 this was found to be 0.76 (95% CI 0.66 to 0.88). However, if data from the WHS trial46 are included then the effectiveness of aspirin appears to be reduced and the OR is increased to 0.819 (95% CI 0.690 to 0.970; Figure 31). Data for incidence of all cancer types from the PHS trial47 beyond 3 years are not available. It should be noted that Seshasai et al. 56 reported greater cancer mortality in the aspirin group for the PHS trial. 47

FIGURE 31.

Odds ratio for cancer incidence when data from WHS46 is included.

Rothwell et al. (2012)

This study by Rothwell et al. 62 generally repeated previous analyses on cancer incidence and mortality apart from analysis of the impact of aspirin on cancer metastasis. This element is excluded from this report, which has primary (metastasis is secondary prevention) prevention as its focus.

Algra et al. (2012)

The primary outcome in this systematic review61 derived from observational studies that provided evidence about the effect of aspirin or NSAIDs on the incidence and outcome of cancer. The authors included the meta-analysis of RCTs, which was also presented in other systematic reviews published by these authors (details in Appendix 7). The conclusion from the review was that data from observational studies support the results derived from analysis of RCTs.

Mills et al. (2012)

The authors60 undertook a study-level meta-analysis of RCTs that investigated the effect of low-dose aspirin on cancer mortality or non-CV death. Trials were included if daily aspirin was administered alone (75–325 mg aspirin and no other anticoagulants) in any population and if the required outcomes were reported. Of 24 included trials, 11 reported cancer mortality. The pooled RR of death from cancer was 0.77 (95% CI 0.63 to 0.95). The numbers of events and participants in each study were not provided. The forest plot listed 11 studies but RR was reported for only eight. The reason why three trials were excluded was unclear.

Adverse events

Table 14 summarises the adverse events reported in the four selected study-level systematic reviews37–39,56 and the ATT IPD review of CVD prevention and Table 9 summarises the RCTs from which the systematic reviews were drawn.

Bartolucci et al. 37 did not meta-analyse adverse events but produced a table of percentage of participants with GI bleed in the nine primary studies; the remainder of this section concentrates on the other three recent meta-analyses. 38,39,56

Non-trivial bleeds and major bleeds; study-level meta-analyses

A difficulty faced by systematic reviewers was that the primary studies spanned three decades of investigation, during which classification and reporting of clinically important bleeding evolved. Seshasai et al. 56 meta-analysed ‘non-trivial bleeds’ using ORs, whereas Berger et al. 39 and Raju et al. 38 meta-analysed ‘major bleeds’ using RRs. Definitions provided in the systematic reviews for these outcomes are summarised in Table 15, whereas Table 16 is based on the systematic review by Seshasai et al. ,56 which provided the definitions of bleeds given in the nine studies40–48 of aspirin for primary prevention of CVD.

FIGURE 32.

Odds ratio of non-trivial and RR of major bleeds in study-level meta-analyses. (a) Seshasai et al. ;56 (b) Berger et al. ;39 (c) Raju et al. 38 Note that the apparent precision for the WHS48 study is greater in the Seshasai et al. 56 analysis.

All of the three reviews used a random-effects model. The random-effects pooled OR for non-trivial bleeds was 1.31 (95% CI 1.14 to 1.50);56 the pooled RRs for major bleeds were, respectively, 1.62 (95% CI 1.31 to 2.00)39 and 1.66 (95% CI 1.41 to 1.95)38 (Figure 32). All three analyses therefore found that aspirin was associated with a statistically significant increase in non-trivial or major bleeds, considered clinically important by the reviewers.

Berger et al. 39 and Seshasai et al. 56 included data on important bleeds from all nine core studies; Raju et al. 38 excluded both the JPAD44 and POPADAD40 studies from the analysis. In the Seshasai et al. 56 analysis, those studies with the largest number of participants (HOT,43 PHS47 and WHS46) were the most influential in determining the non-trivial bleed pooled OR, whereas in the Berger et al. 39 analysis WHS,46 PPP41 and POPADAD40 studies were most influential for the RR of major bleeds. In the Raju et al. review,38 the most influential studies were WHS46 and PPP41 (since Raju et al. 41 excluded POPADAD40 from the meta-analysis).

FIGURE 33.

Influence of individual studies on pooled estimates of (a) odds ratio for non-trivial bleeds and (b) risk ratio for major bleeds.

Figure 33 shows our re-analysis for the OR for non-trivial bleeds (see Figure 33a) and the RR for major bleeds (see Figure 33b). Each line indicates the effect on the outcome measure of excluding the listed study. As can be seen in Figure 33a, removal of the HOT trial43 reduces the OR considerably, and removal of the PHS trial47 increases the OR. In Figure 33b the RR is reduced by exclusion of the PPP trial41 and increased by exclusion of the WHS. 46

Non-trivial bleeds and major bleeds: study-level meta-analyses concerns

As might be expected, the event rates in aspirin and control groups for non-trivial bleeds56 varied over a much wider range than those for major bleeds39 (Figure 34). In Berger et al. ,39 the POPADAD study40 appears to be an outlier, although this study was not included in the Raju et al. review. 38,39

The event rates in the Seshasai et al. 56 meta-analysis for several studies are so much larger than for the other meta-analyses that it is suspected that double-counting of individuals may have occurred so that ‘count data’ (see Cochrane Handbook for Systematic Reviews of Interventions, section 9.2.579) from these trials has been treated as dichotomous data. The Cochrane handbook79 terms this ‘a unit-of-analysis’ error (p. 261, section 9.3.5).

FIGURE 34.

L'Abbé plot of event rates. (a) Non-trivial bleeds (Seshasai et al. 56); (b) and (c) major bleeds (Berger et al. ,39 Raju et al. ,38 respectively). The dashed line represents null effect and the solid line the pooled effect size. Note the difference in axis scales.

Figure 35 illustrates the control group risk for non-trivial bleeds and major bleeds with studies arranged according to length of follow-up; the risks for non-trivial bleeds can be seen to be more variable and the pooled effect is much higher.

FIGURE 35.

Baseline risk. (a) Non-trivial bleeds (Seshasai et al. 56); (b) and (c) major bleeds (Berger et al. ,39 Raju et al. ,38 respectively). LCI, lower confidence interval.

The assessment of risk by Berger et al. 39 and Raju et al. 38 for the PHS,47 TPT48 and WHS46 studies was clearly different from the Seshasai et al. 56 assessment of risk. For example, in Berger et al. ,39 the control group risk of a major bleed in the WHS46 trial was ≈0.5%, but in Seshasai et al. 56 the risk of a non-trivial bleed in WHS46 was considered to be 18.4%. In Seshasai et al. , ‘non-trivial bleeds’ was a composite end point defined as “clinically ‘nontrivial’ bleeding (fatal bleeding from any site; cerebrovascular or retinal bleeding; bleeding from hollow viscus; bleeding requiring hospitalisation and/or transfusion; or study-defined major bleeding regardless of source).”56 In the report of WHS,46 we consider that it is likely that this definition identifies events rather than individuals and may have been meta-analysed as ‘counts’. However, mixing count data with dichotomous data in the meta-analysis is unlikely to be a sound procedure. 79

Cumulative meta-analyses

We undertook cumulative meta-analysis to examine how the pooled estimate for risk of important bleeds changed with time; studies were arranged according to estimated mid-period of recruitment. The pooled estimates became statistically significant after the accumulation of only the first two studies45,47 and remained fairly constant thereafter (Figure 36).

FIGURE 36.

Repeated test with accumulating studies arranged by recruitment year: non-trivial bleeds. (a) Major bleed; (b) Berger et al. ;39 and (c) Raju et al. 38 [NB-PPP (2001)41 and HOT (1998)43 had comparable estimated mid-point recruitment periods (i.e. 1993).]

All major bleeds; individual patient data-level meta-analyses

The ATT53 Collaboration performed an IPD meta-analysis using data from six41,43,45–48 of the core nine RCTs40–48 of aspirin for the primary prevention of CVD. The three RCTs not included were JPAD,44 POPADAD40 and AAA,42 which were published in 2008–10. Major bleeds were defined as major GI and extracranial bleeds that were fatal or required blood transfusion. Trials with fewer than 1000 non-diabetic patients (e.g. JPAD44 and POPADAD40) were excluded. Major bleeds were meta-analysed as RaRs (ratio of events/person-years at risk) determined using log-rank statistics from analysis of IPD. The report quoted a 99% CI for RaRs for individual studies and a 95% CI for the pooled estimate. The pooled RaR was 1.54 (95% CI 1.30 to 1.82) (Figure 37), representing a 54% increase in event rate from aspirin use. The reported rate of events (rounded) across six studies was 0.1% per year for the aspirin group and 0.07% per year for the comparator group; this generated a reported yearly absolute difference of 0.03%/year (i.e. three events per 100 patient-years of exposure). The numbers of major bleed events were 335 and 219 for aspirin and control groups, respectively and the total person-years 330,000. This generates an absolute rate difference (not rounded) of 0.03515%/person-year, equivalent to 35 extra events incurred should 10,000 persons be followed up for 10 years.

FIGURE 37.

Meta-analysis of IPD for major bleeds reported by ATT. 53 LCI, lower confidence interval.

Gastrointestinal bleeds

Raju et al. 38 presented a random-effects meta-analysis for the RR of a GI bleed (RR 1.37, 95% CI 1.15 to 1.62) (Figure 38).

FIGURE 38.

Relative risk of GI bleed (Raju et al. 38).

The risk of a GI bleed was higher in the aspirin group in all studies except POPADAD. 40 The two largest studies (WHS46 and PHS47) were highly influential in determining the pooled estimate (Figure 39) and had high risk in both arms of the trials. Bartolucci et al. 37 presented a table with percentage of patients with GI bleeds in the nine core RCTs. 40–48 These are shown in Table 19, together with percentages calculated from the data in Raju et al. 38

TABLE 19 - Percentage of patients with GI bleeding

Trials	Data from Bartolucci et al.37		Data from Raju et al.38
	Aspirin	Control	Aspirin	Control
WHS46	4.50	3.80	4.57	3.77
BDT45	0.30	0.40
PHS47	4.00	3.80	7.63	6.31
HOT43	0.80	0.40	1.14	0.59
PPP41	0.80	0.20	0.76	0.22
TPT48	1.40	0.90	1.74	0.79
AAA42	0.50	0.50	0.54	0.48
JPAD44	0.80	0.30	0.95	0.31
POPADAD40	4.40	4.90	4.39	4.86

FIGURE 39.

(a) Influence of individual studies on pooled estimate of RR of GI bleeds; (b) L'Abbé plot showing GI bleed event rates (the dashed line corresponds to RR of 1; the solid line represents the pooled RR (random effects).

Across the eight RCTs40–44,46–48 the risk of a GI bleed in the control arms was heterogeneous, varying between 0.22% and 6.30% (Figure 40); the highest risk was exhibited by the two largest studies (PHS47 and WHS46) and the POPADAD study. 40

FIGURE 40.

Risk of GI bleed in control groups; studies arranged according to follow-up. LCI, lower confidence interval.

Raju et al. 54 reported an NNH of 109 for a GI bleed based on the Raju et al. 38 meta-analysis. It is unclear what control risk was used in this calculation. Using number of events and total number of patients aggregated across the control arms (1560/47,498) and a RR of 1.37 the absolute risk difference (aspirin minus control) becomes 0.012 and the NNH 83. If the pooled risk for the control group is used (random effects 0.01281) with the random-effects pooled RR (1.37), the NNH becomes 211. The number of extra GI bleeds if 10,000 persons were followed for 10 years was estimated to be 117 according to the ‘aggregated’ method and 68 according to the ‘pooled’ method.

Haemorrhagic stroke

Both Raju et al. 38 and Berger et al. 39 reported study-level meta-analyses of the RR of haemorrhagic stroke; data came from eight40–42,44–48 of the nine40–48 core CVD RCTs. Data input and results for each trial were similar (small differences existed for TPT and JPAD); no data were available from HOT. 43 No study alone reached statistical significance; the pooled estimates [random-effects model RR 1.36 (95% CI 1.01 to 1.82) in Raju et al. ,38 and 1.35 (95% CI 1.01 to1.81) in Berger et al. 39] indicated a statistically significant increased risk of 35% or 36% with use of aspirin (Figure 41). As pooled outputs were almost identical, only the results from Raju et al. 38 are illustrated in Figure 41.

FIGURE 41.

Relative risk of haemorrhagic stroke (Raju et al. 38).

The most influential studies were PPP,41 PHS47 and WHS;46 if any one of these were omitted from the analysis (Figure 42) the pooled estimate lost statistical significance (at p < 0.05). Repeated test meta-analysis (Figure 43) according to recruitment year indicated that statistical significance in the pooled RR was reached with the inclusion of the PPP41 study, after which, with addition of further studies, the pooled estimate tended to decrease but remained statistically significant at the p < 0.05 level.

FIGURE 42.

(a) Influence of individual studies on pooled estimate of RR of haemorrhagic stroke; (b) L'Abbé plot showing haemorrhagic stroke event rates [the dashed line corresponds to RR of 1; the solid line represents the pooled RR (random effects)].

FIGURE 43.

Haemorrhagic stroke: repeated test with accumulating studies arranged by recruitment year.

FIGURE 44.

Risk of haemorrhagic event in comparator groups. LCI, lower confidence interval.

Raju et al. 38 did not compute a NNH for this outcome. Using the pooled random-effects risk (1.36) and random-effects pooled estimate for risk in the control group (0.219%) (Figure 44), the absolute risk difference is 0.00079 and NNH 1270. With a mean (weighted) follow-up across the eight studies40–42,44–48 of 7.64 years both ‘aggregated’ and ‘pooled’ methods estimate 10 extra haemorrhagic strokes should 10,000 persons be followed up for 10 years. Using the corresponding data from the Berger et al. review39 generates nine extra haemorrhagic strokes (by both ‘aggregated’ and ‘pooled’ methods).

Haemorrhagic stroke: individual patient data-level meta-analyses

The ATT Collaboration53 reported a statistically significant increased risk of first haemorrhagic stroke for individuals receiving aspirin (pooled RaR of 1.32, 95% CI 1.00 to 1.75; Figure 45). 53 The estimates of absolute rates (rounded) were 0.04%/patient-year and 0.03%/patient-year for aspirin and comparator group, respectively, providing an absolute difference of 0.01%/patient-year (10 extra first haemorrhagic strokes if 10,000 persons are followed up for 10 years; this corresponds well with Raju et al. 's38 study-level analysis of haemorrhagic stroke38). The numbers of haemorrhagic stroke events were reported as 116 and 89 for aspirin and control groups, respectively, and the total person-years 330,000. This generates an absolute rate (not rounded) of 0.03515%/person-year (aspirin) and 0.02697%/person-year (control) equivalent and a rate difference equivalent to eight extra events incurred should 10,000 persons be followed up for 10 years.

FIGURE 45.

Meta-analysis of IPD for first haemorrhagic stroke reported by ATT. 53 LCI, lower confidence interval.

The number of haemorrhagic stroke events used in meta-analyses differs somewhat between the different meta-analyses (Table 20); the reason for this is not clear. Only ATT specified ‘first haemorrhagic stroke’;53 if second strokes were counted in the other meta-analyses then it is possible that double-counting occurred. Equally, if a first stroke occurred before randomisation then such patients should not have been included in a primary prevention trial. Transcription errors or variation in interpretation of the primary study data represent other potential reasons for discrepancies.

TABLE 20 - Number of RCT haemorrhagic stroke events used in systematic review meta-analyses

The ATT Collaboration doubled the events because of the 2 : 1 randomisation ratio.

Trial	ATT 201153		Berger 201139		Raju 201138
	Aspirin events	Control events	Aspirin events	Control events	Aspirin events	Control events
BDT45	13	14a	13	6	13	6
PHS47	24	12	23	12	23	12
TPT48	14	6	12	6	2	0
HOT43	12	14
PPP41	2	2	2	3	10	3
WHS46	51	41	51	41	51	41

Haemorrhagic stroke: study-level subgroup analysis by sex

Adelman et al. 52 reviewed meta-analyses of RCTs for primary prevention of CVD so as to investigate a possible difference in outcome between men and women. The particular focus was total stroke and subcategories of stroke (ischaemic and haemorrhagic). For these outcomes, the authors reproduced the study-level sex-specific subgroup meta-analysis published by Berger et al. 70 For this meta-analysis, six41,43,45–48 of the core nine40–48 primary studies, which were also analysed in Berger et al. ,39 Raju et al. 38 and Seshasai et al. ,56 were included.

Three of the six primary studies45,47,48 enrolled only male patients, one enrolled only women46 and two41,43 enrolled men and women. These last two studies41,43 were subdivided by sex and contributed data to the separate analyses of men and women. In the PPP study,41 randomisation was stratified by physician but not by sex; the HOT trial43 stratified randomisation by baseline variables including sex. HOT43 did not provide usable outcome data for haemorrhagic stroke, so for haemorrhagic stroke only PPP41 and WHS46 provided data for women, whereas BDT,45 PHS,47 PPP41 and TPT48 provided data for men.

FIGURE 46.

Meta-analysis of haemorrhagic stroke by sex using erratum data from Berger et al. 80

The reported pooled ORs for haemorrhagic stroke (aspirin compared with comparator) were 1.07 (95% CI 0.42 to 2.69) for women and 1.69 (95% CI 1.04 to 2.73) for men. This apparent difference between sexes led to a discussion of possible reasons for different response to aspirin. The meta-analysis figure in Adelman et al. 52 was taken from the original Berger et al. 72 paper; however, using the data therein does not generate the reported results. A published erratum corrected the data and these have been used in Figure 46.

Aspirin increases the risk of haemorrhagic stroke in men, but for women, although data indicate a 20% increased risk with aspirin, this does not reach statistical significance. The pooled random-effects OR is about 1 and depends on the continuity correction applied for the PPP study. 41 With the continuity correction set at 0.1 the pooled random-effects estimate becomes 1.180 (95% CI 0.518 to 2.686). In the PPP study41 aspirin appeared protective but there were only two events (none in the aspirin arm) and using random-effects weighting to the studies may provide an underestimate of the haemorrhagic stroke risk for women.

Total bleeds

Seshasai et al. 56 reported the pooled random-effects OR for total bleeds (1.70, 95% CI 1.17 to 2.46). For ease of comparison with Berger et al. 39 and Raju et al. 38 meta-analyses of bleeds, Figure 47 shows RR for total bleeds (1.33, 95% CI 1.19 to 1.49). In all studies other than POPADAD40 there were more bleeds in the aspirin group than the comparator group.

FIGURE 47.

Meta-analysis of total bleeds (Seshasai et al. 56).

In this analysis the WHS46 was the most influential trial (Figure 48a); in the WHS46 a very high proportion of participants experienced bleeds in both aspirin and comparator groups (see Figure 48b).

FIGURE 48.

(a) Influence of individual studies on pooled OR for all bleeds; (b) L'Abbé plot showing bleed event rates [dashed line = OR of 1; solid line = pooled OR (random effects)].

As shown in Figure 49, the proportion with events in the control arm of the trials varied from < 1% to nearly 80%. From the number of events (15,542) attributed to the control group in the WHS trial,46 it appears that the numbers of any bleed were combined from multiple categories of event (haematuria, easy bruising and epistaxis) to provide a grand total. As many of these events are likely to have occurred in the same individuals, the results for the WHS46 are almost certainly counts of events rather than proportion of individuals with a bleed event. Combining these data as dichotomous measures represents a ‘unit-of-analysis’ error (as discussed in the Cochrane handbook, section 9.2.5.179).

FIGURE 49.

Percentage with bleed event in comparator arm; trials arranged by length of follow-up (years). LCI, lower confidence level.

Relative effects; benefits

In CVD primary prevention meta-analyses, the potential relative benefits of aspirin appear modest: reduced risk ranges from 6% for all-cause mortality, RR 0.94 [95% CI 0.88 to 1.00 (Raju et al. )]38 to 10% for MCEs RR 0.90 [95% CI 0.85 to 0.96 (Berger et al. )],39 whereas the OR 95% CI for total CHD included a null effect or harm from aspirin (Bartolucci et al. 37 95% CI 0.69 to 1.06, and Seshasai et al. 56 95% CI 0.74 to 1.01). In cumulative meta-analysis the effect sizes appear to have diminished in recent years with the accumulation of later studies. Early studies tended to be more favourable and this may be ascribed to improving treatments for CVD over the years or changes in underlying risk and lifestyle factors as suggested by Seshasai et al. 56 and others.

Apparent cancer benefits appear after about 5 years from start of treatment. The reported pooled ORs for total cancer mortality was 0.93 (95% CI 0.84 to 1.03) when mean follow-up was only about 6 or 7 years. 56 With longer follow-up, a HR of 0.76 (95% CI 0.66 to 0.88) has been reported. 20 The within-trial OR for cancer death in eight trials (25,570 persons) was 0.80 (95% CI 0.69 to 0.93); the large WHS46 and PHS47 (together representing nearly 62,000 individuals) were omitted. The analyses may be considered at risk of selective inclusion of relevant studies (the very large alternate-day dosing studies were excluded and these provide little or no evidence of cancer benefit from aspirin). 81,82 Relative beneficial effects have been found to be most striking for CRC mortality, where an OR of 0.66 (95% CI 0.9 to 1.02) has been reported. However, again, this study omitted the two largest studies where aspirin was given every other day. Including these studies in the estimate of CRC incidence produced 95% CIs suggesting the possibility that aspirin might increase as well as reduce risk. In the longest follow-up analysis of cancer benefit22 at 20 years, the HR for all-cause mortality for three long-term studies was 0.96 (95% CI 0.90 to 1.02); authors hypothesised that the small magnitude of long-term mortality benefit might be due to a rebound effect subsequent to withdrawal from continuous aspirin use.

Relative effects: harms

Study-level meta-analyses of nine trials indicated a 62% and 66% increased risk of a major bleed from aspirin usage. IPD meta-analysis of six trials41,43,45–48 suggested a similarly increased event rate of 54%. Increased risk of a GI bleed was estimated to be 37% (study-level analysis of eight RCTs40–44,46–48). The estimated increased risk of a haemorrhagic stroke ranged from 32% (IPD analysis of six trials41,43,45–48) to 36% (study-level analysis of eight trials40–42,44,45–48).

Absolute number of events averted or incurred through use of aspirin

The number of unwanted events (e.g. all-cause mortality) averted by taking aspirin, and the number of adverse events (e.g. bleeding) incurred from aspirin use, are best calculated using IPD taking into account the person-years of exposure to aspirin or length of follow-up. This was done in the ATT meta-analysis of CVD primary prevention trials. 53 The ATT publication reported the rate of averted and of incurred events as %/person-year; thus an absolute difference (aspirin minus control) of –0.06% is equivalent to 0.06 events being avoided per 100 patient-years of exposure. However, the ATT analysis included only six of the core nine trials40–48 currently available, which have been used in recent study-level meta-analyses. 37–39,56 The inclusion of these trials might modify the results coming from the ATT analysis. We therefore estimated the aspirin dependent number of events avoided or incurred using study-level data reported in the four recent meta-analyses, and, where possible, we compared these estimates with those of the ATT analysis. There are various ways of calculating the rate of averted or of incurred events from study-level data; here we have used two methods, an ‘aggregated’ and an ‘alternative’ procedure (see Chapter 3).

The numbers of averted events should 10,000 persons be followed up for 10 years, estimated from data presented in various meta-analyses, were as follows: 33–46 deaths (any cause), 60–84 MCEs (MI or stroke or CV death), 47–64 total CHD events, 34–36 CRC deaths, and 17–85 cancer deaths. The number of incurred events should 10,000 persons be followed for 10 years were 46–48 major bleeds, 117–182 GI bleeds, and 8 or 10 haemorrhagic strokes. Overall the estimated event rates conform to a few tens of events per 100,000 person-years' follow-up, other than GI bleeds, which appear to occur at somewhat higher rates of 68–117 per 100,000 person-years. It should be borne in mind that these values represent ‘best point’ estimates and although based on the most complete available systematic review evidence they are associated with appreciable uncertainties. Table 22 summarises these findings.

Composite primary outcomes in the primary prevention of CVD in diabetes show that for all seven of the included systematic reviews and meta-analyses, all of the upper 95% CIs included the possibility of no improvement, and for some67–69 CIs implied the possibility of greater risk from aspirin.

Strengths and limitations of this review

We undertook comprehensive searches and thorough systematic review methods following recognised guidelines. We evaluated all studies and re-analysed meta-analytic findings to ensure our own interpretations were in agreement with those of the authors.

A limitation was that we date-limited searches to 2008; nevertheless, because of the intense interest that this subject has generated and the cataloguing of all primary research in so many systematic reviews, we are confident that we have not omitted any major relevant RCTs or systematic reviews.

A further limitation is our potential over-reliance on study-level systematic reviews in which the person-years of follow-up are not accurately ascertainable. However, estimates of number of events averted or incurred through aspirin use calculated from data in study-level meta-analyses did not differ substantially from estimates based on IPD-level meta-analyses for which person-years of follow-up were more accurate (although based on less-than-complete assemblies of currently available primary studies). Seshasai et al. 56 calculated similar estimates of averted and incurred events to us. Despite identifying several weaknesses in terms of double-counting, we found the analyses of previous RCTs and listing of outcome definitions across studies in this paper extremely helpful.

Overall assessment of evidence

Benefits of aspirin for primary prevention of cancer or CVD are relatively modest, remain statistically uncertain, and are an order of magnitude less than that observed in secondary prevention for CVD. In contrast, harms (especially bleeding) occur at relatively higher frequency (apparently very high frequency in some populations) and are statistically based on strong evidence. Second, investigations that use a mix of IPD and study-level analyses of RCTs now point to a possible protection against several cancers (notably colon cancer) emanating after about 5 years of aspirin use, and also protection against cancer metastasis. However, these studies should be viewed with caution because data from the two largest primary prevention trials (WHS46 and PHS47) were excluded but show no evidence of cancer protection by aspirin after ≥ 10 years' follow-up,81,82 and because these are post hoc analyses of studies aimed at a different primary outcome, for which rigorous case ascertainment cannot be verified. In practice, people who suffer GI problems may self-select to discontinue aspirin use and this effect may introduce selection bias.

Absolute benefits and risks of aspirin use, estimated using various methodologies, are relatively rare, (usually tens of events per 100,000 years of follow-up) compared with the total burden of the relevant diseases in the population and are finely balanced. It should be borne in mind that estimated values represent ‘best point’ estimates and although based on the most complete available systematic review evidence are associated with appreciable uncertainties. We recommend that policy decisions about the long-term use of aspirin for primary prevention of CVD or cancer in contemporary health care should be made on the basis of evidence becoming available from new trials. In the meanwhile, each individual doctor and patient should make their own decisions about the benefits and risk of aspirin in relation to CVD and cancer.

Author date reference: Adelman 201152

Title: Gender differences in the primary prevention of stroke with aspirin.

Author date reference: ATT (2009)53

Title: Aspirin in the primary and secondary prevention of vascular disease: collaborative meta-analysis of individual participant data from randomised trials.

Author date reference: Bartolucci et al. (2011)37

Title: Meta-analysis of multiple primary prevention trials of cardiovascular events using aspirin.

Author date reference: Berger et al. (2011)39

Title: Aspirin for the prevention of cardiovascular events in patients without clinical cardiovascular disease – a meta-analysis of randomised trials.

Author date reference: Raju et al. (2011)38

Title: Effect of aspirin on mortality in the primary prevention of cardiovascular disease.

Author date reference: Raju et al. (2012)54

Title: The aspirin controversy in primary prevention

Author date reference: Selak et al. (2010)55

Title: Aspirin for primary prevention – yes or no?

Author date reference: Seshasai et al. (2012)56

Title: Effect of aspirin on vascular and nonvascular outcomes: meta-analysis of randomized controlled trials.

Author date reference: Wolff et al. (2009)57

Title: Aspirin for the primary prevention of cardiovascular events: an update of the evidence for the US Preventive Services Task Force

Author date reference: Nelson et al. (2008)58

Title: Feasibility of conducting a primary prevention trial of low-dose aspirin for major adverse cardiovascular events in older people in Australia: results from the ASPirin in Reducing Events in the Elderly (ASPREE) pilot study.

Author date reference: Dorresteijin et al. (2011)59

Title: Aspirin for primary prevention of vascular events in women: individualized prediction of treatment effects.

Author date reference: Fowkes et al. (2010)42

Title: Aspirin for prevention of cardiovascular events in a general population screened for a low ankle–brachial index: a randomized controlled trial.

Author date reference: Mills et al. (2012)60

Title: Low-dose aspirin and cancer mortality – a meta-analysis of randomised trials.

Author date reference: Algra et al. (2012)61

Title: Effects of regular aspirin on long-term cancer incidence and metastasis: a systematic comparison of evidence from observational studies versus randomised trials.

Author date reference: Rothwell et al. (2010)31

Title: Long-term effect of aspirin on colorectal cancer incidence and mortality – 20-year follow-up of five randomised trials.

Author date reference: Rothwell et al. (2011)22

Title: Effect of daily aspirin on long-term risk of death due to cancer – analysis of individual patient data from randomised trials.

Author date reference: Rothwell et al. (2012)62

Title: Effect of daily aspirin on risk of cancer metastasis: a study of incident cancers during randomised controlled trials.

Author date reference: Rothwell et al. (2012)49

Title: Short-term effects of daily aspirin on cancer incidence, mortality, and non-vascular death: analysis of the time course of risks and benefits in 51 randomised controlled trials.