This meta-analysis followed the standards set by the guidelines of the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 and was registered in the PROSPERO database (Registration number CRD42023491805) [12]. Ethical clearance was not obtained as this study collected secondary data from already published studies.
Eligibility criteria
The screening process involved evaluating titles and abstracts of gathered studies, adhering to these selection criteria: (1) research including participants aged 18 years or older diagnosed with CVD, including arrhythmia, aortic dissection, cardiomyopathy, congenital heart disease, cor pulmonale, heart failure, ischemic and hemorrhagic stroke, coronary heart disease, peripheral artery disease, pulmonary embolism, pulmonary hypertension, rheumatic heart disease, valvular heart disease, venous thrombosis; (2) RAR value was measured in ml/g; (3) reporting either all-cause or cardiovascular mortality as the primary outcome; (4) study was done under interventional (randomized or nonrandomized trial) or observational studies (case-control, prospective and retrospective cohort studies) study design, and (5) written in English. There were no restrictions on the year of publication. Studies were excluded if they had unavailable full texts, involved non-human subjects, or had overlapping populations/outcomes.
Search strategy and study selection
IF, CCZ, WW, and PP, executed a comprehensive search for studies available up to February 1, 2024, across multiple databases (PubMed, Web of Science, Scopus, ProQuest), complemented by manual and bibliographic searches for further data. Table 1 describes the keyword employed in the search strategy.
Table 1. Keywords used in the literature searching
No
|
Keywords
|
1
|
((Indicator) OR (Predictor) OR (Prognostic) OR (Diagnostic) OR ("Predictive Value") OR (Mortality) OR (Death) OR (Mortalities) OR (Severity))
|
2
|
(("RDW-Albumin Ratio") OR ("RDW to serum albumin value") OR ("Red Blood Cell Distribution Width-Albumin ratio") OR ("Red Cell Distribution Width-to-Albumin ratio") OR ("Red Cell Distribution Width-Albumin ratio") OR ("Red Blood Cell Distribution Width-to-Albumin ratio") OR ("Red Blood Cell Distribution Width") OR ("Red Cell Distribution Width"))
|
3
|
(("Angina pectoris") OR (Aorta) OR ("Aortic Dissection") OR (Arrhythmia) OR (Arterial) OR (Arteriosclerosis) OR (Atherosclerosis) OR (Atrial) OR (Atrial Fibrillation) OR ("Cardiac arrest") OR (Cardiomyopathy) OR ("Cardiovascular Disease") OR ("Congenital heart disease") OR ("Cor Pulmonale") OR ("Coronary Artery Disease") OR (CVD) OR ("Heart Disease") OR ("Heart failure") OR ("Hemorrhagic Stroke") OR ("Ischemic Heart Disease") OR ("Ischemic Stroke") OR ("Myocardial Infarction") OR ("Peripheral Artery Disease") OR ("Pulmonary Embolism") OR ("Pulmonary Hypertension") OR ("Rheumatic Heart Disease") OR (Stroke) OR (Thromboembolism) OR ("Valvular heart disease") OR (Vasculitis) OR ("Venous Thrombosis") OR (Ventricular))
|
Final
|
1 AND 2 AND 3
|
Following this, duplicates removal and abstracts screening were conducted by CCZ and PP. IF, CCZ, WW, and PP independently reviewed the full texts of studies that initially met the criteria. Differences of opinion among the authors were settled through collective discussion.
Data extraction
Included studies were extracted independently by authors, using a structured table based on the outcome of interest of this study. Recorded data include the following: author's name, published year, number of participants, demographic characteristics of participants, type of cardiovascular diseases, follow-up period, and patients' outcome (all-cause mortality (ACM), cardiovascular mortality).
Quality assessment
The authors independently performed a risk of bias assessment on individual studies. Newcastle-Ottawa Scale (NOS) is utilized to assess the quality of observational studies [13]. Randomized trials would be assessed using the Cochrane RoB-2 tool [14]. Results of the assessment were presented in total scores, which classify studies into poor, fair, and good quality.
Primary and secondary endpoints
The primary endpoints observed in this study were overall ACM, 30-day ACM, 90-day ACM, 1-year ACM, 3-year ACM, and in-hospital mortality. The secondary endpoints included hospital length of stay and ICU length of stay.
Statistical analysis
In this study, the effect measures used were the hazard ratio (HR) for mortality outcomes, along with the standardized mean difference (SMD) for length of stay outcomes. The SMD was pooled using the Hedges'g method. Mortality outcomes of the highest vs lowest RAR categories were pooled using the Generic Inverse Variance method, which required obtaining the HR and their respective 95% confidence intervals (CI) from each study. Anticipating a significant degree of heterogeneity, we employed a random-effects model to pool the effect sizes. Studies that reported RR as the effect measure were considered equivalent to HR, while studies that reported OR would be converted to HR using a method by Zhang and Yu [15], which was also used by previous reports [16,17]. All statistical analyses were carried out in R software (version 4.3.2) using meta [18], dmetar [19], and dosresmeta [20] packages. A two-sided p-value of less than 0.05 was deemed as statistically significant, unless specified otherwise.
High vs low RAR values
The included studies categorized the RAR values in various ways: above and below the median, per tertiles, and per quartiles. To reduce the heterogeneity of the categorization, we assumed the categorization of all included studies to using per tertile and pooled the effects observed from the highest versus the lowest tertile of RAR values, using method by Danesh et al. [21]. The HRs were log-transformed and then multiplied by the factors of 2.18/2.54 or 2.18/1.59 for studies using quartiles or comparing values above versus below the median, respectively. This method has also been widely used in subsequent studies [22,23]. Heterogeneity in this study was quantified using the Cochran Q (χ2) statistic and I2 statistic test, where I2 of 50% and greater and p < 0.10 of the Q statistic represented evidence of significant heterogeneity [24,25]. Given the negligible differences in sample sizes across the included studies, we adopted the Paule-Mandel (PM) method with a Hartung-Knapp (HK) adjustment [26] for estimating between-study variance (tau²), aligning with several current recommendations from simulation studies [27,28]. Subgroup analyses were conducted to investigate the causes of the heterogeneity according to CVD diagnosis and follow-up time (post-hoc).
Dose-response meta-analysis
Due to the variations in cut-off values for RAR categories across studies, a dose-response meta-analysis was conducted. This employed generalized least-squares regression for trend estimation of HRs across dose categories, using the REML method as suggested by Berlin, Greenland and Longnecker, and Orsini [29,30]. The analysis required the number of cases/person-years, the total number of participants, and the HR with its 95% CI for each category. We extracted only the most adjusted outcomes for this purpose. A one-stage random-effects meta-analysis for aggregated data was employed to include studies with fewer than three RAR categories while still producing results comparable to the standard two-stage method [31]. Log estimates were then exponentiated to produce the predicted HRs. The dose-response relationship was assessed using a nonlinear, restricted cubic spline model using three knots at the 10th, 50th, and 90th percentiles if the Wald test was significant at a two-sided p-value of <0.10 [25]. In cases where the Wald test was not significant or a linear model yielded the lowest Akaike Information Criterion (AIC), we reported the relationship using a linear model [32]. Studies that did not report any categorization of RAR values were excluded from this particular analysis.
A non-zero reference group was designated in this analysis. The assigned doses were the mean or median RAR within each reported category. In the absence of these measures, we used the midpoint of the range. For open-ended lower categories, we subtracted the stated lower bound by the adjacent group's interval width [33], while for open upper bounds, half of the adjacent interval width was added to the upper bound [16]. For studies reporting only a low versus high group, the doses were set at half the value for the low group and one and a half times the value for the high group, respectively. In cases where complete event data within each category were not available, estimations of missing data were made based on the total case count and the hazard ratio (HR) for every category [34,35].
Sensitivity analysis
To ensure the robustness of our meta-analysis results, we undertook sensitivity analyses by performing meta-analyses using alternative methods, such as applying a fixed-effects model and different between-study variance estimators and excluding the HK adjustment. To identify potential outliers, we conducted a 'leave-one-out' sensitivity analysis, recalculating pooled effect sizes and heterogeneity results after sequentially omitting one study at a time [36]. We then conducted another analysis excluding the outlying studies to examine the difference in the results from the main analysis [25]. Furthermore, to evaluate the influence of different covariates on RAR's prognostic significance, a meta-regression analysis was conducted employing the restricted maximum likelihood (REML) approach. To assess for publication bias, we utilised both qualitative and quantitative approaches, which were through a visual inspection of funnel plots and conducting Egger's regression test, respectively [37]. If publication bias was detected, the Duval and Tweedie trim and fill approach was employed [38].