Diet and gastric cancer risk: an umbrella review of systematic reviews and meta-analyses of prospective cohort studies

Literature search

The systematic literature search of PubMed, Embase, and the Cochrane Database of systematic reviews (CDSR) for meta-analyses of prospective cohort studies investigated the association between diet and incidence of gastric cancer (from database inception to April 30, 2021). Relevant studies in reference lists that met the inclusion criteria were also included. The search strategy of each database is presented in Supplementary Table S1.

Two independent authors (L.S.J. and X.J.M.) screened titles, abstracts, and the full text, and disagreements were resolved by the third author (Y.X.B.).

Selection criteria

Studies were included if they met criteria as follows: (1) participants: adult humans; (2) study design: systematic reviews or meta-analyses of prospective observational cohort studies; (3) exposure: any dietary factors; (4) outcome: incidence of gastric cancer; and (5) the multivariable-adjusted summary risk with 95% confidence intervals (CI) for dietary factors and outcome were reported. If the study performed subgroup analyses that were stratified by the study design, the results of cohort studies could be included.

We excluded studies if they met the following criteria: (1) non-English articles; (2) no summary estimate was reported; (3) insufficient data for quantitative assessment and synthesis. We also excluded studies that reported results for GC in combination with other cancers. When more than one systematic review or meta-analysis of the same research problem is performed, we included the article with the largest number of primary studies to avoid the inclusion of duplicate studies. If highest-to-lowest and dose–response analyses were used to assess the association between dietary factors and GC incidence, we included both.

Data extraction

For each included article, we recorded the first author, year of publication, exposure, number of primary studies, number of cases, number of participants, original article retrieval time, type of comparison, most adjusted risk estimates and corresponding 95% confidence intervals, and quality assessment tool. The data were extracted independently by the two authors (L.S.J. and X.J.M.) and validated by the third author (Y.X.B.).

Methodological quality assessment

Two independent investigators assessed the methodological quality of the SR/MA using AMSTAR-2 (A Measurement Tool to Assess Systematic Reviews 2) appraisal tool. A standardized checklist of 16 items was included in AMSTAR-2, which consists of the PICO of review questions, protocol registration, study design type, literature search, literature screening, exclusion of literature, data extraction, description of study characteristics, quality assessment, statistical methods, bias risk, heterogeneity, publication bias, and conflict of interest of the included studies. The items are classified as critical and non-critical domains. Each item could be answered with “yes”, “partial yes”, or “no”. The overall rating of the quality for the systematic reviews or meta-analyses can be rated as “high”, “moderate”, “low”, and “critically low”, if 0–1 non-critical weakness, > 1 non-critical weakness, 1 critical flaw with/without non-critical weaknesses, and > 1 critical flaw with/without non-critical weaknesses in the SR/MA, respectively (Shea et al. 2017).

Evaluation of quality of evidence

The GRADE grading system was used to evaluate the quality of evidence for included articles (Guyatt et al. 2011). Based on observational studies, five factors may lead to rating down the quality of evidence and three factors may lead to rating up. The downgraded factors were: (1) risk of bias, (2) inconsistency, (3) indirectness, (4) imprecision, and (5) publication bias. Three upgrade factors were (1) large effect, (2) plausible confounding would change the effect, and (3) dose–response (evidence of a gradient). Ultimately, the quality of evidence for outcomes was graded as “high”, “moderate”, “low”, and “very low” (Guyatt et al. 2011).

Statistical analysis/data synthesis and analysis

For each identified dietary factor of included studies, we recalculated the adjusted summary estimates and 95% CI with P values using the random-effects model, which considered both within and between-study variability (DerSimonian and Laird 2015). The 95% prediction interval (PI) of each SR/MA was calculated, accounting for the degree of heterogeneity between the studies and providing a 95% confident predicted range for the true effect in an individual study (Riley et al. 2011). Heterogeneity among the studies was assessed using the I² statistic. When the values of I² exceeding 50% or 75% are regarded as large or very large heterogeneity, respectively (Cumpston et al. 2019). We used Egger’s test for small-study effects to assess publication bias (Egger et al. 1997; Harbord et al. 2006). It indicates statistically significant publication bias in the smaller studies if the P value for Egger’s regression test is smaller than 0.10 (Sterne et al. 2011). All statistical analyses were conducted using Stata version 14.1.

Result

The 4028 records were identified after the initial search; after removing 420 duplicates, 3515 articles were excluded through browsing titles and abstracts, and 77 articles were excluded through full-text screening. Finally, we selected 16 remaining articles (Ren et al. 2012; Tian et al. 2014; Fang et al. 2015; Bae and Kim 2016; Weng and Yuan 2017; Kim et al. 2019; Kong et al. 2014; Song et al. 2015; Zhou et al. 2016; Ye et al. 2017; You et al. 2018; Yang et al. 2019; Kang et al. 2010; Deng et al. 2016; Wang et al. 2017a; Du et al. 2020), including 57 diet associations in this umbrella review (Fig. 1).

Description of included studies

The characteristics of 16 eligible studies are given in Table 1. The included 16 studies were published between 2010 and 2020. The number of prospective cohort studies included in the SRs ranged from 2 to 30, with a median number of 6. The number of participants ranged from 14,133 to 3,562,414, and the number of cases ranged from 306 to 8482. Fifty-seven exposures of four categories (food and food groups, micronutrient, beverage, and dietary pattern) assess the relative risks and 95% CIs as the effect size. Among 16 studies, the methodological quality of 13 was assessed by the Ottawa Quality Assessment Scale (NOS) (Tian et al. 2014; Fang et al. 2015; Weng and Yuan 2017; Kim et al. 2019; Kong et al. 2014; Song et al. 2015; Zhou et al. 2016; Ye et al. 2017; You et al. 2018; Yang et al. 2019; Deng et al. 2016; Wang et al. 2017a; Du et al. 2020). This scale assigns a maximum of nine points to each study and score of ≥ 7 was considered as high quality. The average range of NOS scores was 7.4–9.0, indicating that most of the original studies in the 13 MAs were of high quality. Three studies did not conduct a methodological quality assessment.

Summary effect size

The meta-analyses of the 57 associations were re-performed using a random-effects model shown in Table 2.

Heterogeneity and small-study effects

Most associations (n = 44, 77%) showed low heterogeneity (I² < 50%). Ten (18%) associations had substantial heterogeneity estimates (I² > 50%) and three (5%) associations (vitamins, vitamin A/retinol and heme iron) had a high heterogeneity with I² > 75% (Table 2).

We found the presence of small-study effects (potential publication bias) using Egger’s test (P < 0.10) for eggs and bread in meta-analyses comparing highest versus lowest intake. In addition, the presence of small-study effects of alcohol was noted in dose–response meta-analyses. However, only an increase of 12.5 g per day alcohol of these associations contained a sufficient number of studies (n ≥ 10) for Egger’s test to be statistically powerful enough to identify small-study effects (Table 2).

Methodological quality and quality of evidence assessment

The results of AMSTAR-2 for each eligible MA are shown in Fig. 2. According to the AMSTAR-2 standard, two (13%) meta-analyses were classified as critically low confidence, 13 (81%) meta-analyses were classified as low confidence, and only one (6%) meta-analyze was rated as high confidence. All the review authors did not report funding sources for the included studies. All of the studies used appropriated meta-analysis methods and considered publication bias. All the published meta-analyses search relevant studies using at least two databases and provide the search strategies, but only one study searched the gray literature. The meta-analyses assessed with the low methodological quality did not provide the list of excluded studies. Eight (50%) of the articles had no explanation of the selection of study design for inclusion. The reasons for the heterogeneity were not stated or discussed in the two included studies (13%). Most (81%) of the studies assessed the risk of bias in included studies using the Newcastle–Ottawa Scale, which is considered a reliable quality assessment tool.

Of the 57 associations, seven associations (12%) were graded as “moderate,” about 16% were graded as “low,” and 70% were graded as “very low.” Only one piece of evidence was stratified as “high.” These meta-analyses were derived from articles from observational studies that were initially rated as “low.” The reasons for downgrading the quality of evidence were mainly low methodological quality, limited sample size, risk of bias, and the 95% confidence interval containing invalid values. The grading of the evidence is shown in Table 2 and more details in Supplementary Table S3.

Principal findings

The impact of dietary factors included foods, beverages, and dietary patterns on the incidence of GC has been analyzed in many published meta-analyses. For a more comprehensive summary of existing evidence, we conducted an overview of these meta-analyses. We also provide an appraisal of the methodological quality and evidence quality of these associations. Overall, we included 16 published meta-analyses and evaluated 57 associations between dietary factors and GC risk. The methodological quality of most of the published meta-analyses was low. Of the statistically significant associations, one association (total alcohol) was supported by high evidence, two (processed meat and processed meat per 50 g/day) were supported by moderate evidence, three (total fruit, vitamin E, and carotenoids) were graded as low quality of evidence, and two (citrus fruit and pickled vegetables/foods) were graded as very low. The existing evidence found a possible inverse association between citrus fruit, carotenoids, total fruit, and vitamin E intake and GC incidence. A higher intake of processed meat, pickled vegetables/foods, total alcohol, and taking 50 g processed meat per day may increase the risk of GC.

Fruits and citrus fruits may be protective factors against the development of GC. A prospective study (Wang et al. 2017b) in Asia showed that increased fruit intake was associated with a reduced risk of non-cardia gastric cancer (OR = 0.71, 95% CI 0.52–0.95), suggesting that high fruit intake may inhibit gastric cancer, especially non-cardia gastric cancer. Citrus fruits include oranges, tangerines, lemons, grapefruits, etc. Epidemiological surveys on GC have found a low incidence of GC in the southern United States, where citrus fruits are abundant (Song 1978). Fruits are rich in bioactive compounds like vitamin C and flavonoids. Vitamin C is a cofactor of enzymes and a scavenger of active oxygen, which can inhibit the formation of carcinogen nitrosamines in the stomach and reduce the oxidative damage of the gastric mucosa (Jain et al. 2017). Flavonoids are secondary metabolites of plants, widely found in vegetables, fruits, pasture, and medicinal plants, with antioxidant, anti-inflammatory, free radical scavenging, immunomodulatory functions, and also promote the absorption of vitamin C (Hazafa et al. 2020), thus reducing the occurrence of GC.

The umbrella review suggests that a high intake of carotenoids may be inversely associated with GC. A nested case–control study with 36,745 participants followed by 9 years found that people with low plasma levels of α-carotenoids and β-carotenoids have a higher risk of GC (Persson et al. 2008). Experimental showed that the potential mechanism of carotenoids act as antioxidants affecting gastric cancer may be the neutralization of reactive oxygen species, thereby protecting DNA from oxidative damage (Velmurugan and Nagini 2005), reducing cell proliferation and inducing apoptosis, and regulating immune function (Hua et al. 2009; Cui et al. 2007). Furthermore, carotenoids can reduce Helicobacter pylori infection and gastric inflammation by transforming the T-lymphocyte response from Th1-response dominated by g-interferon to Th1/Th2-response dominated by g-interferon and interleukin-4 (Liu and Lee 2003).

Although the role of vitamin E in GC is still controversial, previous studies (Zheng et al. 1995; Carman et al. 2009) have suggested that vitamin E may have a potential preventive effect on GC. This effect may be related to vitamin E succinate (VES), an esterified derivative of vitamin E. VES has strong antitumor activity, which can inhibit tumor development by inhibiting cell proliferation, regulating cell gene expression, and inducing tumor cell apoptosis (Li et al. 2012). Experimental studies have shown that VES can induce autophagy and inhibit DNA synthesis in human gastric cancer SGC-7901 cells (Yu et al. 2017; Cao et al. 2021).

Nitrite and nitrate are common food additives. We observed that the association between Nitrates, Nitrites, and NDMA (Nitrosodimethylamine) and GC is inconclusive, while pickled vegetables/foods and preserved meat are associated with an increased incidence of GC. In marinating vegetables, foods, or meat, some components like heme iron, amines, and amides of the food combine with nitrate or nitrite to form nitroso compounds. Nitroso compounds can cause damage to the cells that form the gastrointestinal barrier (Mirvish 1995). After the injury of the gastric biscuit, the gastric parietal cells start to regenerate and even over proliferate. During this process, the probability of errors in DNA replication will increase, and excessive proliferation may also lead to hyperplasia, which induces GC (Wang et al. 2016). Pickled food can also reduce gastric acid secretion, thus inhibiting the synthesis of prostaglandin E (Wallace 2001). Prostaglandin E is a protective substance that can improve the resistance of the gastric mucosa. The reduction of its synthesis will make the gastric mucosa vulnerable to various attacks. In addition, pickled food can increase the risk of Helicobacter pylori infection, and the synergistic effect of the two pathogenic factors can further increase the risk of GC (Raei et al. 2016). A case–control study (Lin et al. 2014) conducted in China shows that eating pickled meat and preserved vegetables are positively associated with GC. Reducing salt-processed food may be an effective measure to prevent GC.

Total alcohol intake may increase the risk of stomach cancer. Alcohol is not a carcinogen, but its metabolites acetaldehyde and ROS have genetic toxicity and carcinogenic effect (Orywal and Szmitkowski 2017). Aldehyde binds to the DNA of digestive gland cells in vivo, interferes with DNA synthesis, and destroys folic acid (Salaspuro 2011). Alcohol releases myeloperoxidase, oxygen-free radicals, and ROS metabolites such as superoxide anion and protease, which promote the occlusion of large blood vessels, and leads to inflammation and injury of gastric mucosa, promoting the occurrence of GC (Zhang 2018). In addition, alcohol can accelerate the absorption of nitrosamines, mycotoxins, polycyclic aromatic hydrocarbons, and other carcinogenic substances in the human body.

Comparison with other studies

The Continuous Update Project (CUP) of the WCRF/AICR issued third expert recommendations on diet, nutrition, and physical activity for cancer (Clinton et al. 2020). They assessed the association between diet and GC incidence. However, the principles of meta-analyses they included and their criteria for judging and grading evidence differ from the existing umbrella review. Our umbrella review included a wider range of dietary exposure factors derived from prospective cohort studies.

Our finding supports existing guidance, which recommends limiting the intake of alcohol and foods preserved by salting (pickled vegetables/foods, processed meat). In addition, our results suggest that an increase of 50 g of processed meat per day increases the risk of GC by 21%. In contrast, a greater intake of total fruit, citrus fruit, carotenoids, and vitamin E may reduce the risk of GC by 7%, 13%, 22%, and 18%, respectively. However, we found only low-quality evidence supporting a high intake of fruit, carotenoids, and vitamin E to prevent gastric cancer, as well as very low-quality evidence supporting increased consumption of citrus fruit reduce GC incidence (the association between fruit, citrus fruit intake and gastric cancer was also reported to be limited-suggestive in the WCRF/AICR report (Clinton et al. 2019)). Although the quality of evidence supporting these associations was not strong, these findings still have important public health implications. The underlying biological mechanism is plausible and needs to be validated by high-quality investigations are required.

Strengths and limitations

Our study provides the first umbrella review of a systematic review and meta-analysis of dietary factors and GC incidence. Our overview has several strengths. Firstly, systematic reviews and meta-analyses have been considered as high-quality evidence-based. However, they are often affected by methods, publication bias, and other factors during the evaluation process, leading to a decrease in quality. Compared with individual systematic reviews or meta-analyses, our umbrella review provides a more comprehensive critical appraisal of published associations, and thus provide more credible evidence for guidelines and clinical practice. Second, exposure data in prospective cohort studies are collected before the disease occurred, there was little recall bias or selection bias. The results are more reliable and robust than in case–control studies. To reduce heterogeneity, we restricted the outcome of interest to gastric cancer incidence only. Third, our study includes dose-relationship analyses that reveal how the risk of gastric cancer varies with the level of exposure factors, showing the direction of the effect. Linear dose–response analysis is considered important, since demonstrating a biological gradient increases the weight of evidence that the relationship may be causal.

This study has some limitations. Although only 5% of the associations indicated publication bias based on Egger’s test, only 18% of the dietary factors included more than ten studies, and 31% of the associations included even fewer than five studies, which may affect the reliability of the results. More work is required to investigate these exposures, which are based on the few studies included. We included the meta-analysis that contained the largest number of original studies for each dietary factor. The meta-analyses included may not be the highest quality evidence. However, one of the main reasons for the low quality of the evidence in our quality assessment was the small number of primary studies included in the meta-analysis. Therefore, meta-analyses that included fewer primary studies were less likely to yield higher quality evidence than meta-analyses that included our umbrella review. In addition, it was impossible to control the bias of primary studies included in the meta-analyses or systematic reviews, which may influence the results.

Conflict of interest

The authors declare that they have no conflict of interest.