Meta-analysis of Soy Consumption and Gastrointestinal Cancer Risk

Soy consumption has received considerable attention for its potential role in reducing cancer incidence and mortality. However, its effects on gastrointestinal (GI) cancer are controversial. Therefore, we performed a meta-analysis to evaluate the association between soy consumption and gastrointestinal cancer risk by searching for prospective studies in PubMed, Web of Science, EMBASE and the reference lists of the included articles. The study-specific odds ratio (OR), relative risk (RR) or hazard ratio (HR) estimates and 95% confidence intervals (CIs) were pooled using either a fixed-effect or random-effect model. Twenty-two independent prospective studies were eligible for our meta-analysis, including 21 cohort studies and one nested case-control study. Soy product consumption was inversely associated with the incidence of overall GI cancer (0.857; 95% CI: 0.766, 0.959) and the gastric cancer subgroup (0.847; 95% CI: 0.722, 0.994) but not the colorectal cancer subgroup. After stratifying the results according to gender, an inverse association was observed between soy product intake and the incidence of GI cancer for females (0.711; 95% CI: 0.506, 0.999) but not for males.

product group, except for those by Ward et al. 47 and Hedelin et al. 42 , which only reported the intake of isoflavones ( soy product among the included studies, and thirteen studies evaluated the intake of miso soup. In the subgroup study, we conducted a meta-analysis of miso soup intake and GI cancer risk. The data collection method that was used for the three studies was an in-person interview, while the remainder of the 19 studies used a self-administered FFQ.
Three studies adjusted for the confounding factors of age and sex, while the remaining 19 studies applied multiple adjustments. The exposure type and gender-specific risk estimates of GI cancer and the adjustments for confounding factors are shown in Table 2.
Quantitative synthesis. Soy consumption and GI cancer incidence. In our meta-analysis, the intake of mixed soy types had no cancer site-specific or gender-specific association with GI cancer incidence.
Eight studies reported the association between isoflavone intake and the incidence of GI cancer. The highest versus the lowest categories of isoflavone intake had no cancer site-specific or gender-specific associations with GI cancer.
Seven studies reported the association between miso soup intake and the incidence of GI cancer. No gender-specific or cancer site-specific associations were detected between miso soup intake and GI cancer incidence.
Soy consumption and GI cancer mortality. The estimated summary risk for the highest versus the lowest categories of soy consumption showed no association with the mortality of overall GI cancer, mortality of gastric cancer, or mortality of colorectal cancer. After stratifying according to gender, no association was observed for females or males.
In the subgroup analysis, we stratified by exposure, and no association was detected for soy product intake or miso soup intake.
Detailed results of the subgroup analysis are summarized in Table 3.  Table 3. Pooled risk estimates between lowest categories compared with highest categories of soy consumption and gastrointestinal cancer risk. * Statistically significant (P < 0.05).
Publication bias and sensitivity analysis. The results of the Begg-Mazumdar test and Egger's test indicated no evidence of a substantial publication bias for most of the analyses, except for the analysis of soy product consumption and the incidence of GI cancer for both genders. Although this analysis showed a publication bias under Egger's test, it did not show one under the Begg-Mazumdar or funnel test. We strictly followed our inclusion criteria, and therefore, we determined that the results did not suggest any publication bias. We applied a sensitivity analysis on our positive meta-analysis results. The overall pooled estimate did not substantially vary with the exclusion of any single study (Figs 4 and 5).

Discussion
We systematically reviewed the existing literature from three main databases and identified 22 prospective epidemiological studies that assessed the association between soy consumption and GI cancer risk. The findings showed that there was no association between soy consumption and GI cancer risk. Cancer site-specific and soy subtype-specific subgroup analyses revealed that the highest versus the lowest categories of soy product consumption were inversely associated with the incidence of overall GI cancer and the gastric cancer subgroup, but not the colorectal cancer subgroup. A gender-specific analysis showed that this protective effect that the soy product has on the incidences of GI cancer and gastric cancer was only observed in females.
Our results did not find any association between soy consumption and colorectal cancer risk, which was consistent with some previous meta-analyses, including Yan et al. 22 and Jin et al. 48 . However, Tse et al. 15 , Yu et al. 26 and Zhu et al. 49 reported that soy consumption had an inverse association with CRC. Although the previous studies were inconsistent, our study included the newly reported articles by Umesawa et al. 12 and Hedelin et al. 42 , both of which reported no association between GI cancer risk and soy consumption. Woo et al. (2013) performed a meta-analysis of the risks of gastric and colorectal cancer with flavonoids intake 14 . The inclusion of this study showed no association between colorectal cancer risk and flavonoids intake when case-control designed studies were excluded, while a significant inverse association was detected when case-control designed studies were included. Our meta-analysis included only prospective studies, which minimized the recall bias and selection bias from case-control studies, while most retrospective studies reported a significant inverse association. Thus, our most updated and prospective studies included only a meta-analysis, which was more reliable.
Several mechanisms may account for the inverse association between soy product consumption and the incidence of gastric cancer. Two of the major soy isoflavones are genistein and daidzein, which have anti-inflammatory and antioxidative effects 50 . Genistein is known to inhibit the growth of H. pylori 51 and the activation of the nuclear factor-kappaB (NF-κB) signaling pathway. The classical activation pathway of NF-κB signaling has been identified in regulating inflammation-associated gastrointestinal tract malignancies [52][53][54] . Genistein also reduced the growth and proliferation of gastric cancer cells by cell cycle arrest and the Akt signaling pathway, which increased apoptosis and inhibited angiogenesis 55-57 . Interestingly, this protective effect was only found for soy product consumption but not for the mixed exposure. Of all of the included studies, seven studies reported the association between soy product consumption  Scientific RepoRts | 7: 4048 | DOI:10.1038/s41598-017-03692-y and the incidence of gastric cancer 6,40,41,45,46 . Wada et al. 6 and Hara et al. 40 reported this association in females and males, respectively. Thus, we considered them to be two independent studies. Those seven studies that had a clear statement on the measurement of the intake of the mixed types of soybean products are shown in Table 1. However, there were three studies 12,28,44 that reported the relationship between miso soup consumption and the incidence of gastric cancer. When we combined those three studies with the previous seven studies that included a mixed exposure, the above-mentioned protective effect was not observed. Miso soup is a traditional Japanese food with high salt that is made from fermented soybeans 58 . The fermented soy foods contain N-nitroso compounds. High concentrations of sodium in the diet were reported to enhance the carcinogenicity of N-nitroso compounds and H. pylori infection, as well as weaken the protective effect of the mucous barrier 12,59,60 .
In our study, the beneficial effect of soy consumption was found among the female population but not among the male population. Chandanos et al. reported that women with a longer fertility life and those who are on hormone replacement therapy seem to have a decreased risk of gastric cancer, and men who have been treated with estrogen for prostate cancer also have a decreased risk 61 . The mechanism for this decrease in risk remains unknown. Isoflavones have a similar structure to 17β-estradiol and act as estrogen agonists or antagonists in environments of different estrogen levels, which may contribute to the different beneficial effects of soy consumption in females and males 62 .
Moderate heterogeneity was found from some of our results. First, while every study adjusted for age and gender in the calculation of risk estimates, not every included study has been adjusted for total energy intake and body mass index, which are confounding factors 63 . Second, the effects that soy intake has on GI cancer risk might differ among different preparations or fermentations of soy foods. Three included studies adjusted and analyzed fermented and non-fermented soy food 6,40,46 . The high intake of non-fermented soy food was more likely to be inversely associated with gastric cancer risk 6 . A higher salt intake increased the risk of GI cancer, and miso soup, one of the soy subtypes, was considered a high salt food 12,64 . Third, the data gathering methods that were used might also contribute to the heterogeneity. Four studies relied on a personal interview, while the remaining studies came from the self-reported Food Frequency Questionnaires (FFQ) 28,31,39,46 . The participants may have different understandings of the questionnaire by different methods. Fourth, thirteen studies used a validated FFQ mixed with nine non-validated FFQs. The validated FFQ listed various types of soy foods, leading to precise estimates of soy or isoflavone intake. Fifth, we have pooled cohort studies and a nested case-control study with different estimates of OR, RR and HR. HR and OR were considered to be approximations of RR because CRC is a rare outcome in humans. We used a random effects method to determine when the heterogeneity (I 2 ) was larger than 40% to enhance the credibility of the results.
Our meta-analysis has several strengths. First, our study was based on only prospective studies, which enabled us to minimize the food exposure recall bias and selection bias. To our knowledge, this is the first time that the association between both GI cancer incidence and mortality with soy intake from prospective studies has been summarized. Most previous meta-analyses collected both retrospective and prospective studies. Woo et al. (2013) reported that a case-control design created a significant association between the flavonoid subclasses and cancer risk, while cohort studies did not observe this association 14 . Second, all included studies strictly followed our inclusion criteria, which made our results more stable. Third, our sample size is an important strength, as we included a total of 12,901 cancer cases from a total of 965,466 participants. Combining a large number of participants renders us sufficient power to detect potential, modest associations. Fourth, according to our sensitivity analysis, the inverse association did not vary with the exclusion of any single study. Similar to all other meta-analyses, our study has some limitations. First, moderate heterogeneity was observed from some of our results. We have discussed the reasons above; however, the sensitivity analysis showed that our inverse association was stable and reliable. Second, the included studies were reported from different countries and populations and the measurement of soy intake and soy type varied among them.
In summary, no association was found between soy consumption and GI cancer incidence or mortality. A higher intake of soy product is associated with the decreased risk of overall GI cancer and gastric cancer, but not colorectal cancer. This protective effect was observed in females but not in males.

Methods
Search strategy. We systematically searched three databases, PubMed, ISI web of science and EMBASE, for studies that were published in any language (up until December 7, 2016). We combined the key words of the three following items: terms for outcome (colorectal cancer, gastric cancer, or gastrointestinal cancer), terms for exposure (soy product or isoflavone), and terms for epidemiology (cohort, prospective, or observational study).
In addition, we reviewed the reference lists of all of the eligible studies to identify more potential studies.

Study selection.
The following inclusion criteria were applied in the screening of articles: (1) original reported data that evaluated the association between soy consumption and GI cancer incidence or mortality, (2) studies with a prospective study design, (3) studies that used risk point estimates, e.g., odds ratio (OR), relative risk (RR) or hazard ratio (HR) estimates with 95% confidence intervals (CIs), and (4) studies with population-based control samples. We did not include the studies that reported the associations between the serum concentrations of isoflavones and GI risk. When there were multiple published reports from the same study population, the most recent or the most informative report was selected for analysis.
Data extraction. The extracted data that were used included the first author's name, year of publication, participants' ages, study name, location, sample size, cancer type, study period, method used for the food intake measurements, validity of FFQ, method used in the cancer and/or death ascertainment, exposure items, soy consumption type, the risk estimates or data used to calculate the risk estimates, 95% CIs and adjustments for potential confounding effects. When more than one adjusted ratio was reported, the ratio with the most adjustment variables was chosen.

Credibility of meta-analysis results. We performed this meta-analysis under the guidance of Preferred
Reporting Items for Systematic reviews and Meta-Analyses (PRISMA) 65 and Meta-analysis of Observational Studies in Epidemiology (MOOSE) 66 . All enrolled studies were in strict compliance with well-designed inclusion criteria and exclusion criteria. To protect from bias, there was no change of results when any of the studies were excluded by the sensitivity analysis. Two observers independently evaluated the quality and eligibility of the included studies.
Statistical analysis. We extracted the association between soy consumption and GI cancer incidence or mortality by the ORs, RRs or HRs that were reported in the included studies. Soy type was defined as being one of three subgroups: soy product, isoflavone or miso soup. When more than one adjusted ratio was reported, the ratio with the most adjustment variables was chosen. ORs, RRs or HRs and 95% CIs were estimated based on the most adjusted variables for the highest versus the lowest soy consumption. In situations where the incidence was low, the odds ratio approximates the relative risk and hazard ratio. Therefore, for studies of GI cancer (a rare event), it is acceptable to compare the OR, RR and HR estimates [67][68][69] . The outcomes are presented as a forest plot with the 95% CIs. We used I 2 and Cochrane Q statistics, which are quantitative measures of inconsistency among studies, to test for possible heterogeneity across the studies 70 . When I 2 was from 0% to 40% and had a P > 0.10, the heterogeneity might not be important. If the meta-analysis has no heterogeneity, a fixed-effects model with the Mantel-Haeszel method 71 would be used to combine the individual studies. Otherwise, the random-effects method 72 was used for pooling.
To estimate multiple modification effects, cancer site-specific, gender-specific and soy type-specific analyses were performed. Additionally, we did a single study sensitivity analysis for each of the statistically significant results. Sensitivity analyses were conducted by excluding each study, in turn, to evaluate the stability of the results.
The Egger's regression test 73 and Begg-Mazumdar test 74 were used to assess for publication bias. P < 0.05 was considered to be a statistically significant publication bias.
All reported P-values were two-sided. All statistical analyses were performed using STATA (version 11.0; Stata-Corp, College Station, TX).