Intake of legumes and cardiovascular disease: A systematic review and dose–response meta-analysis

We have conducted a systematic literature search in PubMed, Scopus and Web of Science for original studies published up to March 22, 2022 (Supplemental Table 1), that reported the association between the intake of legumes and CVD (comprising any cardiovascular disease), as well as CHD and stroke in particular. The exposure of interest was defined as the intake of beans, peas, lentils or other species belonging to the leguminosae family [], in their dried or fresh form, reported individually or in combination. Since the association between the intake of soy and CVD has already been extensively reviewed [], soy or soy products were not considered under the exposure definition, unless combined with other legume subtypes.

A study was considered eligible if 1) it was an intervention or a prospective or retrospective observational study that 2) evaluated the association between the intake of legumes and CVD, CHD or stroke in 3) the general adult population. Additionally, studies were suitable for inclusion if 4) they reported risk estimates as rate ratios (RRs), hazard ratios (HRs), or odds ratios (ORs), with the corresponding 95% confidence intervals (CIs), or data to calculate the respective variance, for all categories of intake. Studies were excluded if 1) not involving human subjects, 2) not presenting original data (e.g., review articles, editorials, reports, comments or guidelines), 3) performed among children (<18 y) and/or pregnant women, 4) having a cross-sectional design, ecological studies, case series, case reports and qualitative studies, 5) addressing exposures other than the intake of legumes as defined above, 6) addressing outcomes other than CVD, CHD or stroke. In addition, studies in which legume intake was expressed as a continuous variable could not be used in the dose–response meta-analysis. Lastly, no restrictions were imposed regarding the study's language, geographical location or publication date.

The titles/abstracts and full-texts of studies identified in the original search were reviewed independently by two authors (VM and ANi) in accordance with the exclusion criteria. Any disagreements were resolved by consensus or after discussion involving a third author (AN). The literature search was complemented by back and forward citation tracking of eligible papers and other relevant systematic reviews and meta-analyses.

Two authors (VM and ANi) have independently extracted the information using a standardized data collection form. Since no intervention studies were identified at the study selection stage, the following data were extracted: 1) study characteristics (study design, publication year, duration of follow-up when relevant), 2) sample characteristics (country, study cohort, sex, age, sample size), 3) exposure characteristics (definition, data collection methods and categories of intake), 4) outcome characteristics (definition, outcome assessment methods, number of cases) and 5) association measures with CIs for fully-adjusted models and covariates. Estimates were extracted for the whole sample or for different strata, as available. Authors of eligible studies with missing information that was relevant for the calculation of the pooled estimates were contacted and reply was provided by investigators of the Finnish Mobile Clinic Health Examination Survey on legume intake for males and females combined [].

For each intake category we extracted the mean or median intake value, depending on the available data. If these were missing, the midpoint of each intake category was computed. When the highest and lowest exposure categories were “open”, a value that was 20% higher or lower than the closest cut-off point was calculated. For studies where the legume intake was only reported as number of servings/portions, intake was converted into grams using a predefined weight of 100 g per portion, in order to be consistent with the serving size used in previous meta-analyses [, , , , ,].

The internal validity of eligible studies was assessed independently by two authors (VM and ANi) using the ROBINS-I (Risk Of Bias In Non-randomised Studies of Interventions) tool []. In case of disagreement, consensus was achieved involving a third author (AN). The following risk of bias domains were considered: (1) bias due to confounding, (2) bias in selection of participants into the study, (3) bias in classification of interventions, (4) bias due to deviations from intended interventions, (5) bias due to missing data, (6) bias in measurement of outcomes, (7) bias in selection of the reported results. For each domain, studies were judged to be at low, moderate, serious or critical risk of bias. The tool was tailored to address our specific research question and to accommodate the nature and methodology of studies selected for the analysis, in accordance with the ROBINS-I guidance document [].

We performed traditional meta-analyses to investigate the association between the intake of legumes and CVD, CHD and stroke by comparing the study-specific highest versus the lowest intake category. Summary RRs and respective 95% CIs were calculated by applying random-effects models [], as they are generally considered preferable to meta-analyse data presented in peer-reviewed publications [] and to allow the generalization of results beyond the studies considered []. Moreover, in the overall analysis ORs were assumed to approximate RRs in accordance with the “rare disease assumption”, a practice commonly followed in this field [,]. Heterogeneity was assessed using the I² statistic []. In order to investigate possible sources of heterogeneity, we conducted subgroup analyses for each outcome of interest according to the a) study design (case–control vs. prospective cohort studies), b) study location (Asian vs non-Asian studies), c) follow-up time (<10 years vs. ≥10 years) d) validity of the dietary assessment method (validated vs. not validated), e) number of cases (<500 vs. ≥500 cases) and f) sex (females vs. males vs. both). Sensitivity analyses were performed for each outcome by excluding studies that specifically include soy in their exposure definition and studies that were classified as having a serious/critical risk of bias. Additional sensitivity analyses included the application of alternative standard values (i.e., ±15% and ±25% instead of ±20%) to estimate the highest and lowest exposure categories when these were open []. Furthermore, all meta-analyses were repeated by removing each single study (leave-one-out method) in order to evaluate its influence on the pooled effect estimates. We assessed the possibility of publication bias through visual inspection of funnel plots and the Egger's regression test.

Using the methodology established by Greenland and Longnecker (1992) [], and advanced by Orsini et al. (2012) [] and Crippa et al. (2019) [], we have performed a dose–response meta-analysis to identify the association between the intake of legumes and the risk of CVD, CHD and stroke. We applied the ‘one-stage’ approach [,], that allows to include studies assessing only two levels of exposure, using a restricted cubic spline model with 3 knots at fixed percentiles (10, 50, and 90%) of the intake distribution. The restricted cubic spline model was fitted with a generalized least-squares regression taking into account the correlation within each set of the effect estimates and combining the effect estimates using the restricted maximum likelihood method in a random-effects meta-analysis [,,]. All analyses were performed with the Stata statistical software, version 14 (StatCorp, College Station, TX, USA) and RStudio (Version 1.1.456).

Out of the total of 26 studies identified in the systematic review, Martínez-González MA et al., 2011 [] was excluded from the analyses addressing the association between the intake of legumes and CVD overall for being performed using the same sample as Mata-Fernández A et al., 2021 []. Twenty-five studies were thus included, comprising as outcomes CVD as a composite measure, CHD or stroke (Supplemental Table 2). Of these, twenty were prospective cohorts [,, , , , ,, , , ,,, , , , , , , ,] (mean follow-up: 14.6 y) and five were case–control studies [,,,,]. One study was performed in Africa [], six in Asia [,,,,,], nine in Europe [,,,,,, , , ] and nine in North America [, , ,, , , , ,]. The pooled sample included a total of 1 703 121 participants and 44 181 cases and age ranged between 19 and 83 years. All studies used Food Frequency Questionnaires (FFQs) for the dietary assessment, with the exception of Mizrahi, A et al. (2009) [] who relied on interviewer-administered diet history. Three studies have explicitly included soy-legumes in their definition along with non-soy ones [,,], while eleven have explicitly excluded them [,,, , ,,,, , ,]. For the remaining eleven studies, no details on the exposure definition were provided. CVD cases were ascertained using medical records and vital statistics, with the exception of: Baik, I et al. (2013) [] (based on self-reporting and vital statistics); Scarmeas N et al. (2011) [] (using magnetic resonance imaging (MRI)); and Durga AV & Manorenj S (2019) [] (using MRI or computed tomography scans). Maher MA & Gutbi SS (2017) [] did not report on the method used to identify new cases.

A summary of the risk of bias assessment is presented in Supplemental Table 5. Four studies were judged to be at serious risk of bias due to either limited control of confounding (i.e. failure to adjust for important confounders, including age, sex, body mass index (BMI) and energy intake; in the absence of BMI, controlling for energy intake and physical activity levels was sufficient; in the absence of energy intake, adjusting for overall food intake was acceptable) [,] or concerns regarding exposure classification [,,], while one study was judged to be at critical risk of bias [] due to possible bias in exposure classification (not using a validated dietary assessment method and not quantifying intake levels).

The meta-analysis comparing the highest versus the lowest category of intake (31 comparisons) showed a protective association between the consumption of legumes and CVD risk (RR = 0.94; 95%CI:0.89, 0.99), with a moderate level of heterogeneity across included estimates (I² = 43.1%) (Fig. 2). The Egger's regression asymmetry test showed no publication bias (Supplemental Figs. 1–B), supported by the substantially symmetric distribution of the corresponding funnel plot (Supplemental Figs. 1–A). The pooled estimate was largely determined by prospective cohort studies (20 out of the 25 studies considered) and as expected, the RR remained materially unchanged when the analysis relied only on this study design (n = 20, RR = 0.96, 95%CI: 0.91; 1.00; I² = 39.5) (Fig. 2 and Supplemental Table 8). Conversely, the association was stronger when only case–control studies were considered (n = 5, RR = 0.72, 95% CI: 0.60, 0.85; I² = 0.0%). An inverse association was also observed among studies conducted in Asia (n = 6, RR = 0.82, 95%CI: 0.74, 0.90; I² = 9.9%) and the pooled estimate for non-Asian studies was similar to that observed for the main analysis (n = 19, RR = 0.97, 95%CI: 0.92, 1.02; I² = 34.0%) (Supplemental Table 8). The association was stronger considering studies that did not rely on a validated dietary assessment method [,,,,,,] (n = 6, RR = 0.85, 95%CI: 0.74; 0.96; I² = 30.0%), when compared with those using validated methods [, , , , ,, , , , ,,, , , , , , , ,] (n = 19, RR = 0.96, 95%CI: 0.91, 1.01; I² = 42.2%) (Supplemental Table 8). Subgroup analyses based on follow-up duration, number of included cases and participants' sex did not reveal significant changes (Supplemental Table 8). The inverse association between the intake of legumes and CVD risk also remained after the exclusion of studies specifically assessing soy intake [,,] (RR = 0.93; 95%CI: 0.88; 0.98; I² = 47.7%) and judged to be at serious/critical risk of bias [,,,,] (RR = 0.94; 95%CI: 0.90; 0.99; I² = 37.5%). Furthermore, after selectively removing each individual study (Supplemental Fig. 2), the pooled estimates have consistently remained within the 95% CI of the main analysis. In addition, pooled estimates remained the same after the application of ±15% or ±25% (instead of ±20%) to the closest available boundary (lower or upper) in open categories of intake, indicating that results are robust to the choice of a standard value for these estimations.

Of the twenty-five studies included in the analysis comparing extreme categories of intake, seventeen provided data that were considered in the dose–response analysis [,, , , ,, , ,,,,,, , , , ], comprising a total of 1 489 243 participants and 37 166 CVD cases. Considering an intake level of 0 g/week as reference, one portion of cooked legumes (100 g) per week was associated with a reduced CVD risk (RR = 0.97; 95%CI: 0.90, 1.04), which was further attenuated per each increasing portion (Fig. 3). The subgroup analysis based on study design is presented in supplemental figures 3-A and 3-B. Based on evidence from 14 prospective cohort studies, the inverse association between legume intake and CVD risk remains but is attenuated (Supplemental Figs. 3–A). Conversely, the dose–response analysis performed using data from case–control studies only has indicated a U-shaped association between the intake of legumes and CVD (Supplemental Figs. 3–B). However, the number of included studies was small (n = 3) and the interpretation of this result should be made with caution. Dose–response associations between legume intake and CVD risk by study location are presented in Supplemental Fig. 4. The gradually increasing benefit for every additional weekly portion was observed in non-Asian populations (14 studies, Supplemental Figs. 4–B) and it was more pronounced in Asians (3 studies, Supplemental Figs. 4–A), who have also reported higher intake levels.