2.2 Methodology
2.2.1 Criteria for study selection
The main inclusion criteria for patients will be as follows: (1) adults, age ≥18 years; and (2) diagnosed with HF. The exclusion criteria for patients will be as follows: (1) history of cardiogenic arrest or cardiogenic shock; and (2) diagnosis of depression17 or cognitive functioning disorder18. The groups will be usual care with or without education as compared to those with self-care interventions. We will include randomized controlled trials evaluating the effects of self-care intervention models and observational studies, and qualitative studies and duplicate reports will be excluded.
2.2.2 Search strategy
We will search the following databases from January, 2000 to June, 2021 (with the language restricted to English): MEDLINE/PubMed, Embase, the Cochrane Library, CINAHL, ClinicalTrials.gov, and PsycINFO. MeSH terms will be combined with text words in the literature retrieval related to self-care (as defined above) and “heart failure”. To find additional studies, we will screen the relevant reference lists of the included studies and previous literature reviews including systematic reviews and meta-analyses. We will also search for studies by searching the literature published by the first and senior authors of eligible studies. Given that many studies many be unpublished, we will also search the clinical trial registries to identify any ongoing or unpublished trials. All search results will be imported to an Endnote (version X9) tool to facilitate the screening of titles and abstracts.
2.2.3 Data collection and extraction
Two independent reviewers will work together to screen eligible studies. In the initial screening, they will read titles and abstracts to include studies containing one or more intervention models. All potentially relevant studies will be retrieved for full-text screening. A full-text review of potentially eligible studies will then be conducted, and the reasons for exclusion after full-text screening will be recorded in the Endnote database. After screening, data will be extracted independently by the reviewers using a data extraction form. The extraction form will consist of the following categories: characteristics of studies (e.g., years, time of publication, title), characteristics of participants, and characteristics of interventions and controls, outcomes, and others. Discrepancies will be resolved through discussions until reaching consensus between reviewers.
2.2.4 Risk of bias assessment
The same investigators will assess the risk of bias for all qualifying studies based on the Cochrane risk of bias assessment tool. This tool consists of seven parts, namely random sequence generation, allocation concealment, blinding of participants and personnel, blinding of outcome assessors, incomplete outcome data, selective outcome reporting for each outcome, and other potential threats to validity. The investigators also will independently assess and rank the risk of bias for every eligible study as low, medium, or high19.
2.2.5 Outcomes
The four primary outcomes of interest will be as follows: quality of life assessed by the Kansas City Cardiomyopathy Questionnaire (KCCQ), the Minnesota Living with Heart Failure questionnaire (MLHFQ), and other validated questionnaires; mortality (all-cause, HF-related); hospital admission/readmission (all-cause, HF-related); and healthcare use.
2.2.6 Data synthesis
Because of complexities in the process of statistical analysis, there may be modifications of certain aspects during the network meta-analysis.
Nonetheless, as the first step, a DerSimonian-Laird random effects model will be used for the conventional pairwise meta-analysis. Standardized mean difference (SMDs) and 95% confidence intervals (CIs) will be calculated for continuous data measures by Hedges’ g and interpreted according to Cohen’s criteria20. For dichotomous outcomes, the relative risks of each study will be calculated. We will present pooled effect results with 95% CIs and use forest plots with I2 (test based on 95% CIs)21 to investigate heterogeneity. Then, the study effect sizes will be synthesized using random effects, network meta-analysis in a Bayesian framework. We will utilize the heterogeneity variance to measure the degree of the effects of variability across and within studies on the effects of intervention models. To estimate and present the likelihood of each rank order, we will use Stata (version 16.0) to obtain the surface under the cumulative ranking curve (SUCRA). The consistency will be tested by Cochran’s Q.
To evaluate and adjust the effects of covariates (i.e., duration of intervention, provider, healthcare setting, age, gender, and HF severity during baseline period), meta-regressions will be conducted. The goodness-of-fit between models will be evaluated by comparing deviance information criterion (DIC); the lowest DIC indicates the best model fit. Based on the optimal fitting model, we will conduct a network meta-analysis after adjusting for covariates. We will fit all models in WinBUGS (version 1.4.3) and assume uninformative priors for all meta-regression coefficients. After considering the Brooks-Gelman-Rubin diagnostic and autocorrelation plots, Markov Chain Monte Carlo chains will be visually checked to ensure the convergence of the model. Publication bias will be investigated with comparison adjusted funnel plots. We will use R (version 1.3.1093) and Stata (version 16.0) to evaluate the inconsistency and produce the network graphs and result figures. The GRADE tool will be used to assess the strength of the body of evidence.