Statistics from Altmetric.com
In this issue of Heart, Beauchamp and colleagues paper, entitled ‘Attendance at cardiac rehabilitation is associated with lower all-cause mortality after 14 years of follow-up’, sets high expectations for cardiac rehabilitation (CR) in contributing to a long-term mortality effect.1 The study of 544 patients suggests that the relative increase in mortality risk for non-attenders was 58% compared to attenders (Hazard ratio (HR) 1.58, p=0.004) and that the dose of CR may actually determine the extent of mortality. These findings are substantially higher than previous Cochrane review data, where the mortality effect, for 47 Randomized controlled trials (RCTs), involving 10 794 patients, was (Relative Risk (RR) 0.87) and (RR 0.74) for all-cause mortality and cardiac mortality, respectively.2 When compared with the very low, non-significant, mortality effect (RR 0.98) from the rehabilitation after myocardial infarction trial (RAMIT) study,3 which continues to be debated,4 the 1.58 HR, from Beauchamp and colleagues, is as a polar opposite in terms of mortality benefit.
The obvious question is, ‘why is there so much variation in mortality estimates involving CR?’ It could be that the Melbourne CR programmes are really that good, but equally the high mortality estimate might be explained by other factors. This editorial will try to clarify the situation while highlighting wider issues and challenges with CR mortality estimates.
CR and all-cause mortality
Beauchamp and colleagues1 used retrospective data, from their previous study,5 which originally aimed to ‘identify sociodemographic and clinical predictors of non-attendance and dropout separately for men and women automatically referred to CR programmes’. As the purpose of their first study was to identify differences between attenders and non-attenders,5 it becomes obvious that differences also exist between attendees and dropouts, which question the rationale for combining them and make it difficult to picture the CR context in the follow-up paper.1 Pure death rates, as reported in the recent manuscript,1 are therefore limited in terms of interpretation. Although the authors tried to take account of this important limitation, the small sample size remains insufficient for a survival model adjusted for six covariates.
Non-attenders, as defined in the recent paper,1 were in the oldest age group with concomitant higher frequency of diabetes. To counter this variation, Beauchamp and colleagues ran different models, with and without the older age group, concluding that the 46 extra deaths, in the above 70-year-old non-attenders, had no real impact on the outcome. Although a well-known determinant of the outcome, in patient with cardiovascular disease, the number of deaths associated with diabetes was not reported. Hammill et al6 found the risk of death, in 70–78-year-old cardiac patients with diabetes mellitus, to be significant with a HR of 1.21 (CI 1.13 to 1.30). When in older age, diabetes and propensity for death are combined, in respect to the Beauchamp et al's study, it hypothetically wipes away any reported differences in all-cause mortality between attenders and non-attenders.
As with many retrospective studies, the unknown effect size for mortality, lack of prespecified study endpoints, small number of participants and limited covariate adjustment of important comorbidities may have influenced the scale of these results.
CR and mortality: is there a dose–response relationship?
One of the potentially attractive contributions of the Beauchamp et al's study was the dose–response analysis, suggesting that the extent of CR may determine mortality. They found that patients who attended fewer than 25% of sessions had twice the mortality risk than those who attended over 75% of sessions. Nevertheless, by having too few participants in each CR category and with their own results showing that smoking behaviour eroded the assumed CR mortality benefit, the findings add little to the CR dose–response evidence.
Generally, sample size requirements for multivariable Cox regression, accounting for diagnosis, comorbidities, intervention type and dose of CR, are in the order of thousands of patients. Using four of the most recent registry studies, sample sizes between 4000 and 70 000 participants were associated with a CR mortality benefit of around 14–34%.6–9 Multivariate analysis, using matched cohorts with adjustment for propensity scores, is seen as vital in determining the association between CR and mortality.10–12 One of the challenges with attributing a mortality benefit to CR, since the 1990s, is that cardiology has undergone huge changes with many meta-analyses showing, that in the run-up to the millennium, evidence-based cardiac care and medications became routine practice.2 ,13 ,14
As financial austerity takes hold, the dose of CR delivered is likely to become an important clinical question. So how far, from the evidence-base, can CR stretch yet retain its effectiveness?
West and colleagues3 suggest that CR, in the UK, has already gone too far (eg, too few multidisciplinary staff, reduced CR programme length and suboptimal intervention provision), whereas others argue that the 14 programmes in the RAMIT study were not representative of the UK and that CR remains effective.2 ,4 ,6 ,11 ,12 Nonetheless, CR programmes are being asked, often without extra resources, to increase patient throughput. The National Audit of Cardiac Rehabilitation (NACR) confirms that programmes also include patients with heart failure, stable and unstable angina, those with defibrillators and other assisted devices plus valve patients.15 Such service pressures, from funders and commissioners, could lead to CR programme delivery changes that may or may not dilute the effectiveness of CR. Clinical audit remains essential in capturing these adaptations and future research should prioritise evaluation of real-world models of CR.
Why use retrospective observational data analysis?
A genuine alternative to RCTs that have shown promise in recent years is the use of observational studies.12 Healthcare systems routinely collect data on hundreds of thousands of patients, yet little is done to utilise this valuable source of data. There is no doubt that RCT findings rank highest in decision making about effectiveness;12 however, recent observational studies, strengthened by quality data retrieval, have enabled analysts to develop robust models, with appropriate matching of confounders, in cardiac populations.6–9 ,12
A further rationale is that the sample sizes required for modern CR effectiveness trials have been estimated at around 8000 patients for a 20% mortality benefit.3 West and colleagues, in a well-resourced CR trial, demonstrated significant difficulties with recruitment to usual care groups.3 If the RAMIT trial was unable to recruit successfully, to target, it is difficult to see how others could improve on it. Finally, numerous trials and meta-analyses have established CR as a proven intervention,2 ,4 ,14 meaning that pure RCTs of mortality are unlikely to gain ethical approval or can be seen as valid by clinicians and patients.
However, in the real-world, data quality of observational studies is often a difficult issue, for example, due to a high number of missing values or loss to follow-up cases. Generally, survival data are documented either as a time until occurrence of an event or as the time until censoring (due to study end or loss to follow-up). In observational studies, it is often difficult to determine these exact time points.10–12 The general rule, when pursuing mortality estimates from observational data, is that every effort to improve the quality of data should be made.
Hammill et al6 carried out the most current observational analysis of CR in over 30 000 patients, with a mean age of 74 years, and found important differences in mortality and health outcomes in CR attenders and non-attenders matched across important variables and comorbidities seen in older populations. Similar approaches in slightly younger cardiac populations have successfully applied a range of analytical techniques to observational data from registries.7–9 One element of good practice, in all these studies, is the need for large sample sizes, often requiring multiple thousands of patients, to accommodate the appropriate analytical model.
Basic requirements for mortality analysis using observational data
Generally, propensity score methods as well as the standard Cox model, with adjustment for covariates, or combinations of each approach represent valid methods to analyse mortality trends.12 ,16 Of course, this implies that all important confounders are known and reported. The hazard ratio analysis remains important, especially for competing risks,10 and is best interpreted as the instantaneous risk to experience an event at time ‘t’ expressed as a ratio between two different groups. The application of mortality statistical models that account for covariates is generally limited by the available sample size and significantly influenced by the presence of heterogeneity and homogeneity.
So what next for cardiac rehabilitation?
Dramatic headlines that claim CR saves lives or that CR does not save lives should be treated with caution!
CR practitioners should be confident in the evidence supporting the benefits of CR, which is based on 47 rigorous RCTs and 4 large observational studies.2 ,6–9 ,14 With the exception of one trial,3 which showed little effect for CR, the culmination of findings from meta-analyses and observational studies supports meaningful outcomes for patients in quality of life, anxiety and depression, fitness, adherence to medication, smoking cessation and reduced hospital admissions. Although the impact of the dose of CR on mortality requires further research, we should not forget that healthcare is about treating the living. The quality of life of survivors is of prime importance and CR is key to delivering sustainable outcomes that can improve peoples’ lives.
Clearly, more research is needed on the dose–response of CR and the subsequent effect on mortality and other health-related outcomes. However, this is a much bigger, longer-term challenge requiring complex analysis in very large studies using quality data associated with CR interventions. This evidence may not come, in the end, purely from meta-analyses of prospective trials but also from well-funded, high-powered, retrospective registry-based observational studies with adequate sample sizes to support multivariate mortality analysis.
In the UK, this will involve the NACR15 with further linkage with other UK national cardiac audits. This will allow high-quality analysis of very large data sets with appropriate matching for covariates and risk scores. As well as unravelling the issues around mortality benefit, this study will seek answers to important clinical and commissioning questions about the type and intensity of CR. Similar, registry-based, opportunities exist in other parts of Europe. It is envisaged that such research will not only help clarify best practice in the UK and Europe but also guide future research to enable CR to remain fit for the purpose.
Professor Bob Lewin and the NACR team at the University of York.
Contributors PD prepared the first draft and PD and GR contributed significantly to the design, analysis and interpretation of subsequent drafts and final version.
Competing interests None.
Provenance and peer review Commissioned; internally peer reviewed.
If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.