A collaborative group was established, coordinated from the Heart Failure Unit of the Imperial College School of Medicine, London. A prospective protocol was written and agreed by the collaborative group before data collection, specifying the methods to be used, the main prespecified analyses, and a common dataset of collected variables.
We searched Medline for randomised controlled trials since 1990 of exercise training in patients with chronic congestive heart failure or left ventricular dysfunction. We cross checked our findings to identify any other published or unpublished trials by consulting researchers and colleagues in exercise physiology and heart failure and by scrutinising reference lists from review articles, and abstracts presented at scientific sessions and published in Circulation, the Journal of the American College of Cardiology, and the European Heart Journal. A subsequent search of the Cochrane Reviews database yielded no additional studies.
Selection and validity assessment
The characteristics of trials to be included were that they should be randomised parallel group controlled trials and should evaluate exercise training without any other simultaneous intervention that could confound the results, should study patients with stable heart failure (three months or more of stability) due to left systolic ventricular dysfunction (left ventricular ejection fraction less than 50%), should have an exercise programme lasting eight weeks or more, should utilise training involving at least both legs (trials of arm or single leg training were excluded), and should have survival follow up of three months or more.
Data abstraction and study characteristics
Initial screening identified 101 potential reports. Several groups had reported information from the same trial in more than one publication or had conducted more than one similar trial. Individual patients were therefore combined into one dataset for that centre, yielding 41 non-overlapping datasets: 27 had been published as peer reviewed original articles, 12 had been presented in abstract form only, and two had not been published in either form. Nine met the eligibility criteria.
After formal agreement, all principal investigators were asked to provide datasets in the form of anonymised individual patient data for each patient originally randomised. Datasets included age, sex, blood pressure, ischaemic versus non-ischaemic causes, functional class according to the New York Heart Association, left ventricular ejection fraction, exercise peak oxygen consumption, dates of randomisation, follow up, death, and, if available, data on any events that had occurred during hospital stay. If patients experienced more than one such event, only the first was recorded. Data on drugs and renal function were requested, but non-availability of this data was not an exclusion criterion for the meta-analysis.
Datasets from each trial were transferred in electronic format (email or disk) to the coordinating centre at the Heart Failure Unit. They were checked for completeness and consistency to ensure that no errors had occurred in reformatting of the data and to check for consistency with the original publications. Queries were resolved by communication with the principal investigators. The number of events in this meta-analysis may differ slightly from those reported by the trials because follow up is now more complete. The datasets were then incorporated into the master database, which was used for analysis. Results of analyses were discussed at several meetings.
Analysis of potential publication bias
The potential for publication bias was first examined visually by constructing a funnel plot displaying precision of the estimate of the effect size (the reciprocal of its standard error) against the estimate of the effect size (odds ratio, on a logarithmic scale).6,7
Asymmetry was assessed formally by the Egger test.7
Finally, the association between variance and effect size was analysed by rank correlation using the Kendall tau method.
Analysis of time to end point
The prespecified event information collected was time to death (from any cause) and, in those studies that recorded it, time to first admission to hospital (for any reason). In all cases an intention to treat approach was used, so events occurring after allocation of treatment were included regardless of the duration of participation in an exercise programme. We considered a P value of less than 0.05 as significant.
The primary end point was time to death. A secondary end point was death or time to admission to hospital. Time to death was available for all studies and time to death or admission to hospital was available for eight of the nine studies. The treatment arms were combined into one arm as were the placebo arms. Statistical packages used were SAS Statview 5.0 and R (R Foundation for Statistical Computing, www.r-project.org
). The Kaplan-Meier method was applied to display the time to end point in the exercise and control arms during two years of follow up. After testing the Cox proportional hazards assumption, we calculated the hazard ratio for events in the exercise group compared with the non-exercise group.8
Statistical significance was tested with the Mantel-Cox log rank method.
Subgroup meta-analysis of dichotomous end points
The effect of exercise was also assessed in prespecified subgroups—males versus females, New York Heart Association functional class I-II versus III-IV, ischaemic versus non-ischaemic causes, age, peak oxygen uptake (< 15 ml/kg/min v ≥ 15ml/kg/min), left ventricular ejection fraction (< 27% v ≥ 27%), and duration of training programme (< 28 weeks v at least 28 weeks). The continuous variables were each dichotomised at their corresponding median values over the whole dataset. For each subgroup we determined whether the interaction term (subgroup×treatment arm) was significantly different from zero.