Two observers (HYQ and TCY) independently and systematically searched PubMed, Web of Science, Cochrane Library from inception to November 10, 2021, to identify observational studies of systematic reviews and meta-analyses assessing the association between non-genetic risk factors and GC using the following search algorithm: (“Stomach Neoplasms” OR “Gastric Cancer” OR “Cancer, Gastric” OR “Cancers, Gastric” OR “Gastric Cancers” OR “Gastric Neoplasm” OR “Gastric Neoplasms” OR “Stomach Neoplasm” OR “Neoplasm, Stomach” OR “Neoplasms, Stomach” OR “Neoplasms, Gastric” OR “Neoplasm, Gastric” OR “Stomach Cancers” OR “Cancers, Stomach” OR “Cancer, Stomach” OR “Stomach Cancer” OR “Cancer of the Stomach”) AND (“systematic review” OR “meta-analysis”). The reference lists of retrieved eligible papers were further hand-searched to avoid missing other potentially related studies. Only articles published in English were included.

Articles were deemed qualified if they satisfied all of the following inclusion criteria: (1) the articles were systematic reviews and meta-analyses of observational studies (i.e., cohort and case-control and cross-sectional studies); (2) the study evaluated the association of environmental (non-genetic) factors and GC risk, but not for screening, diagnostic, prognostic purposes; and (3) the study provided enough data to perform the analyses. The title and abstract of all eligible papers were initially screened, and then the full text of possible qualified articles was retrieved for further perusal based on the pre-established inclusion and exclusion criteria. Disagreements between two investigators were settled through a discussion. If multiple meta-analysis examined the identical scientific issue, we chose the largest number of studies to avoid repeated evaluation of the same primary studies (10, 11).

From each included meta-analysis, two researchers (HYQ and TCY) independently extracted the following data: the examined risk factors, the first author's name, year of publication, the epidemiological design and number of included studies, and the number of participants and cases. The study-specific relative risk estimates [i.e., relative risks (RRs), odds ratios (ORs), and hazard ratios (HRs) together with their corresponding 95% confidence intervals (CI)], heterogeneity, and publication bias for every risk factor were also collected in each study. Divergence during data extraction was clarified by discussion until a consensus was reached.

Two researchers (HYQ and TCY) independently appraised the methodological quality of all included systematic reviews using the updated 16-item AMSTAR 2 instrument (a measurement tool for assessing the methodological quality of systematic reviews) (12). This tool is used to classify the methodological quality into four grades: high, moderate, low, and critically low. No or only one non-critical flaw is considered high quality, more than one non-critical defect is considered moderate quality, only one critical defect with or without non-critical flaws is low quality, and more than one critical defect with or without non-critical flaws is considered critically low quality. Any differences between the AMSTARS 2 scores were resolved through a discussion or arbitration by the third investigator (LDL).

Two authors (HYQ and TCY) independently assessed the strength of the epidemiologic evidence using the following criteria (11, 13, 14):

(1) precision of the estimate (p-value <0.001 (15, 16), a threshold related to significantly fewer false-positive results, and >1,000 cases), (2) the heterogeneity between studies was not relatively large (I² <50% and p-value of Cochran Q test>0.10), and (3) no evidence of small-study effects (p-value of Egger's test>0.10). The strength of the epidemiologic evidence was classified as high (when all of these criteria were met), moderate (when a maximum of 1 criterion was not met and p-value < 0. 001 was satisfied), or weak (in other cases, p-value < 0.05). When the p-value was not directly reported, it was recalculated from the 95% confidence interval of the pooled effect estimate using an established method (17). In case of doubt during the evaluation of evidence quality, discrepancies were settled through arbitration with a third investigator (LDL).

Based on the extracted raw data from every included study, we reanalyzed and presented the random-effects estimate whenever the fixed-effects model was initially used (14). And in the case of missing measures, we calculated them when enough data were available. A p-value of the pooled estimate of effect size of < 0.05 was deemed significant. I² and Q test was used to determine the heterogeneity (18) among studies, while Egger's test was utilized to evaluate the small-study effects (19); a p-value of < 0.10 (Q test) indicated a significant heterogeneity and a publication bias. Values (I² test) exceeding 50% were generally considered to indicate high heterogeneity. All p-values were two tailed, and all statistical analyses were performed using Stata version 16.0.

Overall, the initial search identified a total of 8,347 articles (2,414 from PubMed, 5,736 from Web of Science, and 197 from the Cochrane Library), of which 76 were finally deemed eligible (20–95). The process of selecting contained meta-analyses is displayed in Figure 1, while the general characteristics of the included eligible articles are summarized in Supplementary Table S1. The publication dates of contained studies ranged from 2008 to 2021. Among the meta-analyses reported in our study, the median number of original articles included in each meta-analysis was 12 (range: 2–73), the median number of cases was 4,745 (range: 51–137,451), and the number of cases was >1,000 in 113 (82.48%) meta-analyses. All 76 articles examined 112 unique risk factors and 137 associations, among these associations, 93 (67.88%) reported significant summary effects with a p-value of <0.05, while 47 (34.3%) reported a p-value of <0.001 (Supplementary Table S1).

Obesity is a well-recognized risk factor for multiple adverse health outcomes. Waist circumference and waist–to–hip ratio were associated with an increased risk of GC in the highest vs. lowest comparisons (RR 1.48; 1.24–1.78 and RR 1.33; 1.04–1.70, respectively) (20). Similarly, a higher body mass index (≥30 vs. 18.5-25) was also associated with GC (OR 1.13; 1.03–1.24) (22) (Figure 2).

The Mediterranean diet (MedDiet) and dietary total antioxidant capacity (D-TAC) were associated with significant reductions in GC risk for the highest vs. lowest comparisons (RR 0.7; 0.61–0.8 and RR 0.63; 0.53–0.73; respectively) (23, 26). Fiber intake was inversely associated with GC (28), whereas, refined grain consumption of the highest dosage was related to a significantly increased risk of GC (OR 1.36; 1.21–1.54) (30). Intake of meat, particularly red and processed meat was associated with GC (34, 94). High salt consumption, especially salt-preserved foods, could increase the GC risk (36, 37). Fruit and vegetable intake were widely reported as protective factors for GC (40, 41, 43); of note, vegetable fat also had beneficial effects on GC (RR 0.55; 0.41–0.74) (32). Heavy alcohol drinking and chili intake might increase the risk of GC (48, 49), but no association between tea and GC was found (50) (Figure 3).

Higher vitamin consumption was associated with reduced GC risk (RR 0.73; 0.68–0.78) (55), especially antioxidant vitamins (vitamin A, vitamin C, vitamin E, β-carotene, and α-carotene). An inverse association between total polyphenol intake and GC was also found (OR 0.67; 0.54–0.81, for the highest vs. lowest intake comparisons) (62) (Figure 4).

Regular proton pump inhibitors (PPI) use was associated with GC (RR 1.78; 1.38–2.31) (66). Conversely, regular aspirin use could reduce the risk of GC (RR 0.6; 0.51–0.82) (67). In addition, statin use and menopausal hormone therapy might also be related to a decreased risk of GC (OR 0.65; 0.45–0.93 and RR 0.77; 0.64–0.92, respectively) (71, 72) (Figure 5).

Smoking can be linked to the development of many cancers, including GC (73, 74). Physically active people were protected from subsequent GC (RR 0.81; 0.73–0.89) (76). Higher toothbrushing frequency and refrigerator use also reduced the risk of GC (OR 0.84; 0.77–0.92 and OR 0.7; 0.56–0.88, respectively) (77, 78) (Figure 6).

Depression was associated with an increased GC risk (OR 1.84; 1.61–2.09) (79). Non-alcoholic fatty liver disease (NAFLD) could significantly increase the development risk of various extrahepatic cancers, including GC (OR 1.74; 1.03–2.95) (80). Based on the contribution of autoimmunity to gastric carcinogenesis, autoimmune diseases (e.g., systemic lupus erythematosus, pernicious anemia and diabetes mellitus, type 1) were closely associated with an increased risk of GC (RR 1.34; 1.05–1.72; RR 2.84; 2.30–3.50; RR 1.41; 1.2–1.67, respectively) (81, 83) (Figure 7).

In addition to well-recognized HP and Epstein-Barr virus, other viruses (e.g., hepatitis B virus, hepatitis C virus, human cytomegalovirus, human papillomavirus and John Cunningham virus) were also associated with a higher incidence of GC (86–89) (Figure 8).

A strong inverse association was found between socioeconomic position indicators (educational level and household income) and GC risk (OR 0.60; 0.44–0.84 and OR 0.65; 0.48–0.89, respectively) (90). Of note, we found that blood group O and longer duration of fertility were protective factors of GC, conversely, blood group A was associated with a higher GC risk (72, 91) (Figure 9).

With regard to heterogeneity, in the 137 unique meta-analyses, only 43 (31.39%) associations showed absence of heterogeneity (I² < 50% and p-value of Cochran Q test ≥ 0.10), while the remaining 94 (68.61%) associations indicated significant heterogeneity (I² ≥ 50% or p-value of Cochran Q test < 0.10). In terms of publication bias, 110 (80.29%) meta-analyses showed no evidence of significant small-study effects (p-value ≥ 0.10 of Egger's test), whereas 25 (18.25%) meta-analyses demonstrated publication bias (p-value < 0.10 of Egger's test). With regard to the associations of GC with fiber intake (10 g/d increment) (28) and HTLV-1 infection (89), small-study effects were not applicable as only 2 observational studies were included in each meta-analysis.

The methodological qualities of the 76 included articles were assessed and graded using the 16-item AMSTAR2 tool (Supplementary Table S2); three (3.95%) studies were determined to have low methodological quality, while the other 73 (96.05%) studies were determined to have critically low methodological quality (Figure 10). Based on the AMSTAR 2 scores, none of the eligible articles had high or moderate methodological quality. The most common critical flaws were as follows: lack of registered protocol (n = 64, 84.21%), incomplete literature search (n = 75, 98.68%), and the absence of list of excluded studies (n = 64, 84.21%).

The outcomes of the epidemiologic evidence measurement are presented in Supplementary Table S3. Among the 93 statistically significant associations, only 10 (10.75%) showed high epidemiologic evidence for an association with GC according to the abovementioned prespecified credibility criteria (with >1,000 cases, p-value of < 0.001, and absence of large heterogeneity and small-study effects), including two risk factors (waist circumference and bacon) and eight protective factors (D-TAC, vegetable fat, cruciferous vegetable, cabbage, total vitamin, vitamin A [HvL], vitamin C [100 mg/d increment], and years of fertility) (20, 26, 32, 43, 55, 56, 72, 94). A total of 26 (27.96%) associations demonstrated moderate epidemiologic evidence, the remaining 57 (61.29%) associations presented weak epidemiological evidence (Figure 11).