Types of studies

Randomised controlled trials (RCTs).

Types of participants

Anyone with a clinical diagnosis of IHD or CHF, excluding people with acute myocardial infarction. We included studies evaluating both ischaemic and non‐ischaemic disease only if data for the participants with ischaemic disease could be extracted separately.

Types of interventions

Studies involving the administration of autologous adult bone marrow‐derived stem/progenitor cells on their own or in combination with co‐interventions, such as cardiac surgery, as treatment for IHD or CHF.

Participants in the comparator treatment arm of the trial received either no intervention or a placebo (e.g. the medium in which the cells were suspended or plasma). Trials where co‐interventions (e.g. CABG, PCI, G‐CSF, extracorporal shockwave therapy) were additionally administered were eligible as long as the co‐interventions were equal in both arms and administered to an equivalent proportion of participants.

In summary:

1. any autologous human adult bone marrow‐derived stem/progenitor cells

2. any single dose

3. any method of stem/progenitor cell isolation

4. any route of administration

5. any co‐intervention

6. repeated intervention or multiple doses

Types of outcome measures

Electronic searches

We updated and expanded the electronic database searches, originally run in March 2013 (see Appendix 1 for details), in June 2014, March 2015, and December 2015 (Appendix 2). We identified relevant studies from searching the following:

• Cochrane Central Register of Controlled Trials (CENTRAL) (the Cochrane Library, 2015 Issue 11);

• MEDLINE (OvidSP, 1948 to 14 December 2015);

• Embase (OvidSP, 1974 to 14 December 2015);

• CINAHL (EBSCOHost, 1982 to 14 December 2015);

• PubMed (in process and epublications ahead of print only, on 14 December 2015);

• LILACS (1982 to 14 December 2015);

• IndMED (1986 to 14 December 2015);

• KoreaMed (1997 to 14 December 2015);

• PakMediNet (1995 to 14 December 2015);

• Web of Science: Conference Proceedings Citation Index ‐ Science (CPCI‐S) (1990 to 14 December 2015);

• four databases of ongoing trials on 14 December 2015:

  • ClinicalTrials.gov (clinicaltrials.gov/);

  • ISRCTN Register (www.isrctn.com/);

  • World Health Organization International Clinical Trials Registry Platform (WHO ICTRP) (apps.who.int/trialsearch/);

  • HKU Clinical Trials Registry (www.hkuctr.com).

Searching other resources

We checked the reference lists of all identified eligible papers and relevant systematic reviews. We applied no language or date restrictions.

Selection of studies

The Information Specialist (CD) conducted the electronic search for potentially relevant papers and removed references that were duplicates, clearly irrelevant, and/or included in previous search results. Two review authors (SF, EMR) independently screened all titles and abstracts identified by the review search strategy for relevance to the review question. We excluded studies that clearly did not meet the eligibility criteria at this stage. Two review authors (SF, EMR) independently assessed all other studies based on their full text for inclusion/exclusion using the criteria indicated above (type of studies, participants, interventions, and outcome measures). Disagreements were resolved through discussion.

Data extraction and management

Two review authors (SF, EMR) extracted data onto customised data extraction forms that were created and piloted specifically for this review and independently undertook data extraction for all eligible studies. Aside from details relating to the quality of the included studies, we extracted the following two groups of data.

1. Trial characteristics: place of publication, date of publication, population characteristics, setting, detailed nature of intervention, detailed nature of comparator, detailed nature of outcomes. A key purpose of these data was to explain clinical heterogeneity between included studies independently from analysis of the results.

2. Results of included studies for each of the main outcomes indicated in the review question. For dichotomous outcomes, we recorded the numbers of outcomes in treatment and control groups. For continuous outcomes, we recorded the mean and standard deviation. Where standard deviations of mean change from baseline values were not explicitly reported, where possible we calculated the standard deviation based on reported confidence intervals or P values as described in the Cochrane Handbook for Systematic Reviews of Interventions (Higgins 2011), and we used these values in the analysis.

Disagreements between the review authors over data extraction were resolved by consensus. When disagreements regarding any of the above could not be resolved through discussion, we attempted to contact authors of the original trials to provide further details. One review author (SF) then transcribed the data into the systematic review computer software Review Manager 5 (Review Manager 2014).

Assessment of risk of bias in included studies

The two review authors (SF, EMR) independently undertaking the data extraction assessed the risk of bias for each trial using the criteria outlined in the Cochrane Handbook for Systematic Reviews of Interventions (Higgins 2011). For trials included in the previous version of this review, we re‐evaluated the risk of bias in the context of the revised outcomes and long‐term follow‐up studies, and updated accordingly. Disagreements were resolved through discussion.

A study of trials published in Chinese medical journals that were described as randomised found that a high proportion of these trials did not adhere to accepted methodology for randomisation, and hence could not be deemed authentic RCTs (Wu 2009). It is now widely accepted that trials carried out in China may lack appropriate randomisation; we therefore deemed any Chinese studies for which methods of randomisation were not described and could not be clarified with trial authors to have a high risk of selection bias, and evaluated sensitivity to these trials through sensitivity analyses (see Sensitivity analysis section below).

Measures of treatment effect

We carried out separate analyses according to the duration of follow‐up after treatment: short term (less than 12 months) and long term (equal to or greater than 12 months). We expressed dichotomous data for each arm in a particular study as a proportion or risk and the treatment effect as a risk ratio (RR) with 95% confidence intervals (CIs), calculated using Mantel‐Haenszel methods. We expressed continuous data for each arm in a particular study as a mean and standard deviation, and the treatment effect as the mean difference (MD) if outcomes were measured in the same way across trials. For outcomes measured using different methods, we combined the treatment effect data and analysed them using the standardised mean difference (SMD).

Although we intended to analyse continuous outcomes as mean change from baseline, several studies only reported baseline and endpoint data. Where possible, we calculated the standard deviation of the mean change from baseline based on reported confidence intervals or P values, and used these values in the analysis. However, for several studies, insufficient information was reported to calculate the standard deviation. Since the mean difference based on the change from baseline can be assumed to address the same underlying intervention effects as an analysis based on final measures (i.e. the differences in mean final values will on average be the same as the differences in mean change scores), we combined studies reporting mean change from baseline values with those reporting endpoint values, but have presented mean change and endpoint values separately as well as in combined analyses for clarity, as suggested in the Cochrane Handbook (Higgins 2011). We did not conduct this pooling of studies by method of reporting of continuous measures for analyses of exercise capacity, since the assumption of consistent underlying effects does not hold for standardised mean differences.

Unit of analysis issues

Three published reports of trials randomised participants to one of two treatment arms, each with a comparator control group (Hamshere 2015_IC; Hamshere 2015_IM; Heldman 2014_BMMNC; Heldman 2014_BM‐MSC; Mozid 2014_IC; Mozid 2014_IM); we have considered each of these studies as reporting two separate trials within one publication and treat them as such throughout this review. In the first trial (Heldman 2014_BMMNC; Heldman 2014_BM‐MSC), exercise capacity, quality of life, and LVEF measures were reported pooled for both control groups; for these outcomes the pooled control data are used as the comparator for both intervention arms. In other studies in which there were multiple interventions in the same trial compared with a single control group, we combined the intervention trial arms for a single comparison with the comparator (control) arm to avoid double counting of participants and potential correlation of results. We thus pooled data across different methods of administration (intramyocardial/intracoronary) (Ang 2008), cell types (Assmus 2006), and cell doses (Losordo 2007; Losordo 2011). However, for subgroup and sensitivity analyses, where the two intervention arms were classified into different categories (e.g. type of cell, cell dose, route of administration of cells), we included results for each treatment arm in the corresponding group, with the control group included in both groups. In order to avoid unit of analysis issues, we treated cross‐over trials as parallel trials and included them in the review up to the point of cross‐over, i.e. first‐phase data only.

In the analysis of quality of life outcomes, we converted Minnesota Living with Heart Failure Questionnaire (MLHFQ) scores to negative values in order to include these in a meta‐analysis with other measures on different scales using the standardised mean difference.

Dealing with missing data

We attempted to contact the authors of 27 studies (describing 30 independent trials) by email for clarification of methods (randomisation, allocation concealment, and blinding), potential overlapping of studies, and/or requests for additional data. We failed to establish contact with the authors of 16 studies (17 independent trials) by email (Ang 2008; Bartunek 2012; Erbs 2005; Heldman 2014_BMMNC; Heldman 2014_BM‐MSC; Mathiasen 2015; Nasseri 2012; Patel 2015; Perin 2011; Perin 2012a; Perin 2012b; Pokushalov 2010; Santoso 2014; Tse 2007; Wang 2010; Yao 2008; Zhao 2008), and the authors of one study initially responded but did not reply to subsequent emails (Jimenez‐Quevedo 2011).

We are grateful to the authors of 10 studies (12 independent trials) who responded to our emails as follows:

• Assmus 2006: results were reported for a pooled randomised cohort and a non‐randomised pilot study cohort; the authors provided full clinical and surrogate outcome data for the randomised cohort alone, as well as details of the method of randomisation used;

• Assmus 2013: we received clarification of analysis sample sizes and confidence intervals for mean change in NYHA;

• Hendrikx 2006: we received left ventricular end‐systolic volume (LVESV) and end‐diastolic volume (LVEDV) data (as only LVESV/LVEDV index values were reported) (see previous version of this review);

• Hu 2011: the authors confirmed overlap of multiple publications and provided mean change from baseline data for exercise capacity, LVEF, and other surrogate outcome measures (see previous version of this review);

• Mozid 2014_IC; Mozid 2014_IM: results were reported pooled across intervention arms; the authors provided mortality, MI, rehospitalisation and arrhythmia rates, and mean NYHA and CCS baseline, follow‐up, and change from baseline values separately for each randomised arm of the trial;

• Hamshere 2015_IC; Hamshere 2015_IM: this study was published in abstract form only with limited presentation of results. The authors kindly provided data for mortality, morbidity, NYHA class, and CCS class;

• Patel 2005: we received clarification of randomisation methods;

• Patila 2014: we received mean (rather than reported median) values for LVEF and NYHA class;

• Trifunovic 2015: LVEF data were reported graphically; the authors provided the actual data used to generate the graphs;

• Turan 2011: a discrepancy in brain natriuretic peptide data between papers was resolved; overlap of multiple publications was confirmed.

Assessment of reporting biases

Although we made every effort to identify unpublished studies, we assessed publication bias for the primary outcome of mortality using a funnel plot and with a formal test for publication bias using Egger's test for asymmetry (Egger 1987), implemented with the statistical software programme R v2.14.1 (R Core Team 2013). We accept that asymmetry, one cause of which may be publication bias, is difficult to detect with the small numbers of studies (i.e. fewer than 10) often encountered in systematic reviews.

Data synthesis

We undertook meta‐analyses using Review Manager 5, employing random‐effects models throughout due to the anticipated heterogeneity arising from differences in participant characteristics, interventions, and duration of follow‐up (Review Manager 2014). This differs from the previous version of the review, in which fixed‐effect models were used for meta‐analyses in the first instance.

Although quantitative synthesis was the main method of analysis, we incorporated insights from a qualitative evaluation of studies for an overall interpretation of the data. We based conclusions on patterns of results identified across clearly tabulated results of included studies as well as summary measures, taking both direction and magnitude of any mean effect sizes from random‐effects models into account.

We included all studies in the main analyses irrespective of risk of bias and performed sensitivity analyses for risk of selection, performance, and attrition bias as described in the Sensitivity analysis section below. Periprocedural adverse events were summarised for each trial in tabular form and evaluated descriptively. We made no formal evaluation of the frequency of periprocedural adverse events in each treatment group due to the differences in definition and reporting of periprocedural adverse events between studies.

Within each included trial, all participants were analysed in the treatment groups to which they had been randomised. We undertook an available‐case analysis, including all participants who were randomised to treatment and were included in the analysis, irrespective of whether or not they had received their randomised treatment.

In two trials, no variation in NYHA class, in Trifunovic 2015, or CCS class, in Perin 2012b, between participants within the treatment group was observed (and hence the sample standard deviation was zero). For these outcomes, we estimated the standard deviation by that observed in the control group in order to incorporate these data into the meta‐analysis.

We constructed 'Summary of findings' tables using GRADEpro GDT (GRADEpro GDT). We focused our summary of findings on long‐term follow‐up of the primary outcome of mortality, morbidity (non‐fatal MI, rehospitalisation for HF, composite MACE, arrhythmias) and the surrogate outcome of LVEF measured by magnetic resonance imaging (MRI). We excluded studies with a high or unclear risk of selection bias from random sequence generation from the 'Summary of findings' tables and from summary results presented in the abstract. We made an assessment of the quality of the evidence based on study design limitations, inconsistency of results, indirectness of evidence, imprecision, and publication bias as described in the GRADE handbook (Schünemann 2013), with consideration of the optimal information size generated from trial sequential analysis (TSA).

Subgroup analysis and investigation of heterogeneity

A range of different methods (MRI, left ventricular angiography (LVA), single‐photon emission computed tomography (SPECT), echocardiography, and radionuclide ventriculography (RNV)) were used to measure LVEF across studies, with several studies reporting LVEF as an outcome using more than one method of measurement. The limitations of some of these methods are well known (Arnesen 2007). Consistent with the previous version of this review, we subgrouped analyses of LVEF according to the measurement method used.

We assessed the percentage of variability in effect estimates due to heterogeneity using the I² statistic (Higgins 2002; Higgins 2003). We performed pre‐planned subgroup analysis for mortality (primary outcome). For outcomes with substantial observed heterogeneity (I² ≥ 50%) in combined analyses (or separate analyses for outcomes reported as standardised mean difference) and a minimum of three studies in each subgroup, we investigated potential sources of heterogeneity by performing the subgroup analyses described below as exploratory analyses, and by visual inspection of forest plots with consideration of individual trial characteristics (Higgins 2003). Where possible, we based subgroup analyses on combined analyses of mean values at endpoint and mean change from baseline values, consistent with the main analyses as described in the Measures of treatment effectsection above. We performed subgroup analyses on all available trials irrespective of risk of bias.

Subgroup analysis considered the following factors:

1. mean dose of stem/progenitor cells administered (≤ 10⁷, 10⁷ to 10⁸, or > 10⁸);

2. route of cell administration (intramyocardial, intracoronary);

3. baseline cardiac function (mean baseline LVEF < 30%, 30% to 50%, or > 50%);

4. type of cell administered (mononuclear cells; circulating progenitor cells; haematopoietic progenitor cells; and mesenchymal stem cells);

5. participant diagnosis (chronic IHD; HF (secondary to IHD); intractable/refractory angina), classified in consultation with a clinical expert (AM);

6. use of co‐interventions (PCI or CABG or shockwave administered or not administered).

We regarded the last three subgroup comparisons listed above as hypothesis‐generating.

For trials with multiple active‐intervention arms, in subgroup analyses where the intervention arms were stratified across the subgrouping strata, we used the single control group as the comparator in each subgroup.

Sensitivity analysis

For the outcomes of mortality, non‐fatal MI, rehospitalisation for HF, composite major adverse clinical events, NYHA class, and LVEF measured by MRI, we assessed results for sensitivity to risk of selection bias (by excluding studies with a high or unclear risk of bias from random sequence generation). We also assessed the primary outcome of mortality for sensitivity to risk of attrition bias (by excluding studies with a high or unclear risk of attrition bias) and performance bias (by excluding studies with a high or unclear risk of performance bias due to known lack of blinding of participants and clinicians).