Comparison to 2009 USPSTF Findings

The evidence findings informing our analysis diverge from those informing the 2009 USPSTF aspirin recommendation in several important ways. The most apparent difference is that the updated evidence review no longer finds a difference in benefits between men and women. Previously, aspirin was found to reduce the relative risk of MI in men by 32% and stroke in women by 17%; the updated review finds that aspirin reduces the relative risk of MI and stroke in both men and women by 15% and 18%, respectively. That difference means larger expected benefits for women, but the contrast is less clear for men, because MIs are more prevalent than strokes but strokes tend to have greater impact on quality of life and risk of death. Previous findings applied to men aged 45-79 and women aged 55-79, but the updated reviews considered evidence for men and women aged 40 and older. Another major difference is the new finding of lower CRC incidence risk after 10 years of aspirin use. This added benefit can account for more than half of the lifetime net benefit, in terms of life years and QALYs, from routine aspirin use (Appendix A Table 13, Case 3). Finally, the findings on harms associated with routine aspirin use also have been updated. The prior estimated rate of excess GI bleeds due to aspirin reflected a relative risk of 2.00, compared to 1.59 in the updated review. This results in approximately 40 percent fewer estimated excess GI bleeding events in our analysis. The relative risk of hemorrhagic stroke with aspirin was previously found to be 1.69, but the current best estimate is substantially lower at a relative risk of 1.14.

There are also numerous methodological differences in our approach to estimating net benefit compared to the calculations that informed the 2009 USPSTF aspirin recommendation. The prior net benefit calculations were restricted to first non-fatal events over 10 years. Baseline events were linearly projected from the estimate of baseline risk (Appendix A Table 17), such that, for example, out of 1,000 men with 10% 10-year coronary heart disease risk, 100 were predicted to have MIs (of which 32 cases could be prevented by using aspirin). In contrast, our approach derives a distribution of fatal and non-fatal preventable events, as predicted by the model, based on the risk factors representative of persons in each age, sex, and baseline CVD risk threshold. In this way, we predict far fewer MIs and strokes that may be prevented for each CVD risk group than the 2009 recommendation—even though more than one non-fatal event can be prevented for each person—as these events are subcomponents of the composite outcome for which CVD risk is selected upon (i.e., the combination of coronary death and fatal and non-fatal MI and stroke).³⁷ In addition, Appendix A Table 17 reveals that the ratio of non-fatal to fatal events generally decreases with age, as first or subsequent events become more likely to be fatal. This distinction is important because the CVD prevention benefits to aspirin are found to be realized through the direct reduction in risk of non-fatal MI and ischemic stroke events—meaning, at a given level of baseline CVD risk, there may be relatively fewer of these events that can be prevented by aspirin among persons in older age groups.

The baseline population rate of GI bleeding used in the 2009 USPSTF aspirin recommendation came from an analysis of population-based databases in the United Kingdom and Spain;⁵⁶ in this study, we derived estimates from a population-based study conducted in Italy,³⁴ with age and sex adjustments made for the U.S. population. Although there are some differences—the largest of which are among men and women aged 50-59—estimated GI bleeding rates in the baseline population are generally similar in this study (Appendix A Table 18). What is notably different, however, is that our analysis incorporates estimates of age-adjusted case-fatality associated with GI bleeding events. Accounting for fatal GI bleeds can have a meaningful effect on net benefit calculations—particularly, for men and women aged 70-79 who might have otherwise expected positive lifetime net benefits without accounting for this harm (Appendix A Table 13, Case 10).

The approach to hemorrhagic stroke is also notably different. In the 2009 recommendation, the baseline rate of hemorrhagic stroke did not vary by age for men (for women, ischemic and hemorrhagic stroke were combined). In our analysis, hemorrhagic stroke incidence is determined by a risk equation derived from Framingham Heart Study data specifically for use in our model. Hemorrhagic stroke risk predictors include age, BMI, SBP, and smoking status. Therefore, hemorrhagic stroke rates in our analysis vary by age group and by baseline CVD risk. This also means that both benefits and harms scale with baseline CVD risk in our analysis, in contrast to benefits alone. The resulting baseline population rates of hemorrhagic stroke generated by our model compare well with those found in large U.S.-based cohort studies (Appendix A Table 18) and are generally much higher than assumed by the 2009 recommendation, in part offsetting the difference in the assumed increased relative risk of hemorrhagic stroke with aspirin use.

Another important distinction is that the microsimulation model used in our analysis accounts for the dynamics of competing risks among fatality by CVD, CRC, and GI bleeding and other causes of death in quantifying net benefits. The model also accounts for background use of secondary prevention following a CVD event. When the first non-fatal CVD event of a simulated person is prevented or delayed by aspirin use, their use of aspirin, statins, and anti-hypertensive medications for secondary prevention also may be prevented or delayed. The prevention or delay of an initial non-fatal event also changes the risk of subsequent non-fatal and fatal events. This provides more realistic estimates of the marginal value of aspirin in primary prevention relative to secondary prevention. We also assume that aspirin therapy will be stopped immediately if adverse events are encountered.

Another difference and important strength in our approach is that the baseline 10-year CVD risk for each simulated individual is calculated using the ACC/AHA risk equation, which is separate from the model's risk engine. When assessing CVD risk, clinicians are likely to use the ACA/AHA or a similar 10-year risk calculator in daily practice. Because the risk calculator is separate from the model's risk engine, there is imperfect correlation between a simulated person's baseline line risk categorization and their CVD events as determined by the model. This parallels the imperfect correlation between baseline risk as predicted by a calculator in clinical decision-making and the realized patient experience with CVD over time, similar to that encountered in daily practice. Appendix A Figures 2 and 3 illustrate this imperfect correlation and reflect patterns similar to those shown in other comparisons of the difference between observed outcomes and those predicted by the ACC/AHA risk calculator.^57-60 Despite the imperfect correlation, baseline event rates predicted by our model validate reasonably well to U.S. population event rates observed in NHANES data (Appendix A Table 1).

Another important strength to this study is in providing both short- and long-term outcomes. Standard guidance for the time horizon in health policy evaluations is to ensure a sufficient analytic window such that all important harms and benefits are captured.⁵² Our results reveal that the lifetime horizon is needed to meet this standard and that the benefit-to-harm ratio generally increases over time—such that the largest average net benefit is realized with long-term aspirin use. There are several reasons for this. First, following the findings from the systematic evidence review, benefits to reducing CRC incidence are not realized until after 10 years of starting aspirin use. Second, the absolute risk of CVD and CRC generally increase at a greater rate with age than the risk of bleeding events. Third, and most important, the direct and indirect benefits from preventing non-fatal events can take time to accrue. For example, the direct benefits of preventing or delaying an ischemic stroke accumulate over time, due to the ongoing reduction in quality of life that a person would have otherwise endured after the serious event. However, even if an event has no discernible long-term effect on quality of life—such as is often the case with a non-fatal MI—the risk of future events or death is still increased. In this way, a prevented non-fatal CVD event often can confer lasting indirect benefits in averting or delaying future events or death that would have occurred counterfactually. These indirect downstream benefits are a major factor in explaining the large differences in net benefits often seen between the 20-year and lifetime horizons in our analysis. We recognize, however, that lifetime benefits and indirect outcomes may be extensively discounted or too abstract in the context of individual decision-making. For these reasons, we also provide 10- and 20-year outcomes for more comprehensive clinical context to informing shared decision-making.

The relationship between and valuation among benefits and harms from long-term aspirin use is complex. Another important contribution of this study is in providing life years and QALYs as outcome measures, in addition to specific fatal and non-fatal benefit and harm events. Life years are an important measure because they incorporate differences in the expected length of life that may come from increased prevalence of fatal hemorrhage episodes, balanced against indirect reductions in CVD or CRC mortality. QALYs are an important measure because they incorporate both expected length of life and quality of life effects, balanced among all fatal and non-fatal benefit and harm events. Together, we believe these two measures are the best overall summary of—and are the most useful when assessing—the balance of benefits and harms from routine aspirin use. Still, we recognize that these composite measures may be abstract to interpret—or that patients may have personal preferences with respect to how to weigh the importance of specific benefit or harm events. For these reasons, the detailed outcomes presented in Appendix A Tables 5-12 also provide important context for decision makers.

ModelHealth: CVD also incorporates race/ethnicity-specific CVD and CRC risk factors. The relationship between behaviors and CVD events over time is estimated using the strength of the Framingham Heart Study's considerable longitudinal data, but this comes with well-recognized limitations with respect to generalizability. By incorporating disparities in risk factors by race/ethnicity, the model provides estimates that are more generalizable to the U.S. population. However, it must be recognized that not all differences are necessarily accounted for, including any disparities in environmental risk exposure such as air pollution, utilization of other CVD preventive measures, and utilization of effective CVD treatments. Differences in predicted outcomes by race/ethnicity are not reported, but corresponding differences in CVD risk factors, CRC incidence, and CRC case-fatality rates may affect the relative net benefits that may be expected for specific persons or population groups.

Appendix A Table 19 compares the 10-year risk thresholds (of coronary heart disease for men and stroke for women) identified by the USPSTF in 2009 to the corresponding results from this study for which the benefits of using aspirin are predicted to exceed the harms in terms of net events over 10 years. Despite the numerous differences in the informing evidence and methodology—and the tools to estimate CVD risk thresholds themselves—positive net event thresholds (which exclude CRC in both cases) are of similar magnitude for men aged 40-59 and women aged 50-69. For men aged 60-69, we find positive net events are expected for men with 10-year CVD risk of 19% and higher (compared to 9% previously), and in contrast to prior findings, positive net event thresholds over 10 years were not identified in our analysis for men and women aged 70-79. Differences in findings for these age groups—and to a lesser extent, younger age groups—are primarily explained by the large differences in estimated baseline preventable CVD event rates for each risk threshold (Appendix A Table 17) and the approximately 40% difference in estimated rates of excess bleeding with aspirin use.

Comparison to Other Recent Analyses

A recent study by van Kruijsdijk et al⁶¹ used long-term follow-up results from the Women's Health Study (WHS) to develop competing risk prediction models for the estimation of absolute risk reduction among CVD, cancer, and GI bleeding. Findings from WHS are included among our parameter estimates, and outcomes from this study concord with the average across other studies and others diverge from the broader evidence base. Specifically, over 10.1 years of average follow-up, WHS found no effect on non-fatal MI (RR = 1.01, 95% confidence interval [CI]: 0.83, 1.24), concordant effects on non-fatal stroke (RR = 0.81, 95% CI: 0.67, 0.97), below average effects on serious MI bleeding (RR = 1.40, 95% CI: 1.07, 1.83), and above average effects on hemorrhagic stroke (RR = 1.24, 95% CI: 0.82, 1.87) [44]. Notably, a statistically significant effect on non-fatal MI was found among women aged 65 and older (RR = 0.66, 95% CI: 0.44-0.97). No statistically significant effect was found on total cancer (RR = 1.01, 95% CI: (0.94-1.08) or CRC (RR = 0.97, 95% CI: 0.77-1.24) during the trial period;⁶² however, a highly concordant relative risk of 0.58 (95% CI: 0.41, 0.81) was found in this population during the 8 years post-trial.⁴⁵ Proportional hazard models estimated using trial-period data were used to predict changes in absolute risk for major CVD events, CRC, non-colorectal cancer, major GI bleeding, and death by another cause for 10- and 15-year periods. Overall, they found that harms generally exceed benefits for women younger than age 65, but that benefits modestly exceed harms for women aged 65 and older. Appendix A Table 20 shows a comparison of the major CVD, CRC, and GI bleeding incidence findings with those in our study. Despite differences in underlying evidence and methods, our results over 10 and approximately 15 years are generally quite comparable. The only exception is the higher rate of prevented major CVD events in women aged 65 and older predicted by van Kruijsdijk et al., but this is readily explainable by the large reduction in non-fatal MI seen among this group in WHS.

Another recent study by Cuzick et al.⁶³ used a population-based incidence model to estimate the net difference in event rates over 15 years with prophylactic aspirin use in the general population of the United Kingdom (U.K.). In comparison to our model parameters, their literature review found relative risks for the incidence of non-fatal events to be 0.82 for MI, 0.95 for stroke, 0.65 for CRC, and 1.54 for major extracranial bleeding. They also found relative risks for mortality to be 0.95 for MI, 1.21 for stroke, 0.60 for CRC, and 1.60 for GI bleeding. Effects on both incidence and mortality were applied in their analysis. Another major difference was in finding relative risk reductions for incidence and mortality of esophageal, gastric, lung, prostrate, and breast cancers—although, three coauthors felt the evidence was still inconclusive with respect to lung, prostrate, and breast cancers. Net event rates were calculated using population incidence rates by age and sex in the U.K. over a 15 year period, with aspirin used actively for the first 10 years. CVD benefits were assumed only during the 10 years of active use, and cancer benefits are assumed for years 4-15. Overall, they found that net benefit events generally exceed net harms; however, cancers accounted for 61-80% of the net benefits they found and 30-36% of that was attributed to CRC reduction. Appendix A Table 21 shows a comparison of the non-fatal MI, stroke, and GI bleeding incidence findings to those in our study. For that which can be compared, findings are generally consistent between studies, with differences in net events explainable by differential baseline event rates between the U.S. and U.K. populations and by the combined versus separated approach to aspirin's effect on stroke type.

Limitations

The results reflect average aspirin effectiveness as determined by the systematic reviews. As such, the results reflect cross-contamination of intervention and control groups that occurred in the abstracted clinical trials. Cross-contamination has two sources: participants assigned to the control group may choose to use aspirin, and participants assigned to the aspirin group may choose not to use aspirin (non-adherence). During an aspirin trial, there may be very few who stop taking the placebo and start daily aspirin, but after the trial ends, those assigned to the placebo may begin daily aspirin use. This may impact effect sizes calculated in long-term follow-up studies. Adherence among persons volunteering for and selected into efficacy trials is likely to be higher than in the general population. It is not known by how much cross-contamination reduced the effect sizes used in the model. Although the effectiveness estimates used in the model do reflect some non-adherence, it is not known how the model's effect sizes compared to what would be observed with typical adherence levels and with a pure control group. We do expect, however, that the effectiveness of aspirin reflected in the model and these results should correspond with “good” adherence, insofar as they mirror a population willing to participate in an extended randomized controlled trial. Patients and providers may wish to consider the value and appropriateness of aspirin use for patients with lower expected adherence patterns differently.

Other possibly important limitations of the provided estimates include using the same effect size for all age groups. The systematic reviews did not find compelling evidence of differential effects by age group; therefore, we used the same relative aspirin impact for all age groups. It is not clear how robust homogenous relative risk effects are for all population groups—particularly, for those with low event rates or those not well-represented in the trial populations, such as those in their early 40s. Some aspirin trials included persons who enrolled or aged into their 80s, and we extended aspirin effects at the same levels for persons over the age of 80; however, we did not evaluate aspirin initiation for persons at these ages (nor for those younger than 40), due to their limited representation in the enrollment of the aspirin trials. Population data suggest that age-based inference may matter in some cases. For instance, excess GI bleeding risk from aspirin may be higher among persons younger than age 50.³⁴

By design, both CVD and CRC mortality risk may be affected indirectly by aspirin use in our analysis. With respect to the relative risk of CVD mortality, the low-dose aspirin trials indicate that there may be a small reduction in risk, but this finding is not statistically significant (relative risk = 0.96, 95% confidence interval = 0.84 to 1.11).¹³ Although we assumed no direct benefit in our base case analysis, risk of death from CVD may be reduced in our model as an indirect downstream effect proceeding the prevention of a non-fatal MI or ischemic stroke. Appendix A Table 22 shows that although our analysis also finds small average reductions in CVD fatality with aspirin over 10 years through this indirect pathway (with relative risks ranging from 0.995 to 0.999), these results are still consistent with the non-statistically significant findings among the aspirin trials. With respect to CRC, evidence indicates that the relative risks of CRC incidence and mortality are both reduced with aspirin use. To avoid double-counting, we chose to directly model aspirin's effect on CRC incidence only. That is, when there are fewer cases of CRC, there also will be fewer cases of CRC mortality—all else held equal. Appendix A Table 22 shows that the relative risk reductions in CRC mortality in our modeled populations—due to reduced CRC incidence alone and ranging from 0.73 to 0.86—are within the upper end of the confidence range observed among the trials (relative risk = 0.67, 95% confidence interval = 0.52 to 0.86).¹⁴

By simulating individuals, the model accounts for correlation between risk for CVD and CRC due to tobacco use. Hemorrhagic stroke risk also correlates with overall CVD risk. We did not, however, establish and incorporate into the model GI bleeding risk equations that would account for correlation between GI bleed risk factors and CVD risk factors, such as tobacco use and diabetes, among others. The scope of the project did not allow for this advancement of methods. Careful determination of independent bleed risk factors in the general population and development of GI bleeding risk equations are important priorities for creating more precise estimates of net benefit from low-dose aspirin use.

These results naturally raise questions about whether there is an optimal age to stop aspirin use; however, evidence is lacking on the implications of aspirin discontinuation after long-term use. For example, it could be that the relative risk of harms diminishes with extended use of aspirin, and/or the relative risk of CVD could rebound to the same or a higher level when discontinuing aspirin after long-term use. It is also not clear how long after discontinuing sustained aspirin use the benefits to preventing CRC persist—or whether, as we potentially conservatively assumed in this study, the benefits cease immediately after stopping aspirin. Using a model to inform discontinuation decisions could be misleading without better data to support such analyses.

This analysis approached the decision to use aspirin from the perspective of a person's age, sex, and 10-year risk for CVD. Given the systematic evidence review findings of substantial benefit from aspirin for the prevention of CRC incidence, persons with elevated risk for CRC may have an interest in taking aspirin for this benefit alone. Stratifying net benefits by CRC risk was outside the scope of this analysis, but we believe results for persons at low 10-year CVD risk (for which the CRC share of benefits will be generally greatest) and the detailed outcomes presented in Appendix A Tables 5-12 may be helpful for those approaching decisions to take aspirin from this perspective.

Accounting for the benefits and harms from aspirin with respect to stroke is challenging. It is widely believed that aspirin reduces the risk of ischemic stroke, but increases the risk for hemorrhagic stroke. The latter was not found to be statistically significant in the updated systematic review, but we included this harm in our decision analysis due to its biological plausibility and due to the lack of power in aspirin trials to detect statistically significant differences in this relatively rare event. Only two of the seven low-dose aspirin trials included in the updated systematic review reported non-fatal ischemic stroke independently;^13,38,40 therefore, the combined stroke relative risk observed across trials included hemorrhagic stroke events. We used this combined measure as the best estimate for the relative risk of ischemic stroke in our model. Although the rate of hemorrhagic stroke is much lower than for ischemic stroke, this approach should at least modestly—and systematically—underestimate the ischemic stroke reduction benefit conferred from routine aspirin use. This conservative approach may be appropriate, however, given the imprecision in measuring the increased risk of hemorrhagic stroke.

Finally, case-fatality rates for GI bleeding events are not well-established in the literature. Aspirin primary prevention trials do not show a difference in GI bleed mortality, but they do not have sufficient statistical power to show significant differences because deaths from GI bleeding are rare. At older ages, GI bleed and death risk are increased, and we have assumed a large jump in case-fatality rates from 3% to 19% between ages 60-79 and 80+. Better estimates of how age, sex, aspirin, and other possible risk factors interact to affect GI bleeding and case-fatality rates may modify the net benefit findings—particularly, among older age groups (Appendix A Tables 13-15, Cases 10 and 11).

Conclusions and Future Research Needs

These results indicate that several population groups may benefit from taking aspirin for the primary prevention of CVD and CRC. Specifically, lifetime benefits are predicted to exceed harms among all men and women aged 40-69 with non-elevated bleeding risk. Net benefits are generally greater for persons at higher levels of 10-year CVD risk. For men and women aged 70-79, lifetime net outcomes are mixed: net life years are negative, but net QALYs are positive. Net benefits from aspirin over 10 and 20 years of use are generally much lower and may be negative. Net benefit calculations also favor early over delayed initiation of aspirin use for all men and women aged 40-69.

Discretion should be used when interpreting these results, as deterministic and probabilistic sensitivity analyses reveal meaningful uncertainty about the magnitude of net benefit. Net benefit calculations are most sensitive to uncertainty regarding the effect of low-dose aspirin on the increased risk of hemorrhagic stroke and in the primary prevention of CVD mortality. The relative risks of CRC incidence and ischemic stroke also introduce moderate uncertainty. Moreover, parameter estimates used in this study may not be reliable for populations underrepresented in the aspirin primary prevention trials (such as persons under age 50). A better understanding of the impact of aspirin by age group and the development of comprehensive risk equations for GI bleeding would improve confidence in and precision of the simulation results. Quality of life benefits from using aspirin may be considerably diminished among persons who dislike taking routine medications. Finally, future research may identify additional benefits (such as protective effects against other cancers) or harms that may substantially alter these findings. These sources of uncertainty and patient preferences should be carefully considered in shared decision-making.