Pancreatic cancer (PC) is a highly fatal disease with a 5-year overall survival of approximately 10 % in the USA(Reference Siegel, Miller and Jemal1). PC risk factors include family history, chronic pancreatitis, type 2 diabetes, obesity and heavy tobacco usage(Reference Mizrahi, Surana and Valle2–Reference Koyanagi, Matsuo and Ito4). Chronic inflammation is implicated in PC and supports cancer cells to evade immune elimination and accelerates malignant progression and metastasis to distant organs(Reference Aziz, Sideras and Aziz5,Reference Stone and Beatty6) . C-reactive protein, TNF-α, tumour growth factor-β, IL-1β, IL-6, IL-10 and IL-17 have been suggested as important roles in PC(Reference Feng, Morine and Ikemoto7–Reference Zhang, Zoltan and Riquelme11).
Scientific evidence shows the consumption of energetically rich diets evokes a state of chronic metabolic inflammation(Reference Christ, Lauterbach and Latz12). Exploring the association between higher inflammation in the diet and cancer risk is of great significance. A recent randomised cross-over trial reported whole grains diet reduced weight and systemic hypo-inflammation(Reference Munch Roager, Vogt and Kristensen13). Another randomised controlled trial found low-fat dietary intervention was associated with reduced PC incidence(Reference Jiao, Chen and White14).
The dietary inflammatory index (DII®) is based on the published review of the articles evaluating the effects of the specific foods or food components on six biomarkers of inflammation (C-reactive protein, TNF-α, IL-1β, IL-4, IL-6 and IL-10)(Reference Shivappa, Steck and Hurley15). Recent systematic reviews showed that the higher the DII score, the higher the risks of gynaecological cancer, urologic cancer, breast cancer and colorectal cancer(Reference Lu, Ren and Zhang16–Reference Shivappa, Godos and Hébert20). Some studies have described the link between DII score and PC risk, but the results are inconsistent. Up to now, we have not been able to identify any systematic review or meta-analysis assessing the relationship between DII score and PC risk. We conducted this meta-analysis to summarise the evidence on the association between DII score and PC risk.
Methods
Search strategy
Four electronic databases (PubMed, Embase, Web of Science and the Cochrane Library) were searched up to 22 November 2020, using the following keywords: (dietary inflammatory index OR inflammatory diet OR anti-inflammatory diet OR pro-inflammatory diet OR inflammatory potential of diet OR dietary score) AND (pancreatic cancer OR pancreatic carcinoma OR pancreatic neoplasm OR pancreatic adenoma OR pancreatic ductal adenocarcinoma). Search only for articles published in English. Search history is shown in online supplementary material, Supplemental Table 1. References of all relative articles were also manually checked to identify any potential additional studies. Two independent authors (Z.G., Y.H.) reviewed the titles, abstracts and full text of all the articles we identified. In case of any disagreements, a third investigator (Y.C.) was sought.
Eligibility
The inclusion criteria included: (1) studies conducted on human beings aged ≥ 18 years old who had completed all questionnaires; (2) studies that reported DII score category as exposure; (3) studies that reported the incidence of PC as the outcome measure; (4) studies that provided multi-covariate adjusted OR or hazard ratios and 95 % CI of PC risk associated with the DII score (the highest category v. the lowest category) and (5) randomised controlled trials, observational studies or case–control. The exclusion criteria included: (1) studies without reporting multi-covariate adjusted risk estimates of PC risk; (2) DII score measured as a continuous variable (see online supplementary material, Supplemental Table 2). The meta-analysis conformed to the Preferred Reporting Items for Systematic Reviews and Meta-Analysis checklist(Reference Moher, Liberati and Tetzlaff21). The protocol for this review was registered with PROSPERO (CRD42021237985).
Data extraction and quality assessment
Two researchers (Z.G., Y.H.) independently collected the information from each eligible studies as follows: surname of first researcher, time of publication, countries and regions, study design, sample size, gender proportion, mean age or age range (years), method of dietary assessment, number of DII components, risk estimates and their 95 % CI for DII score category and PC risk, the DII stratification interval, DII score comparison, covariates adjusted in the multivariate model and follow-up time for cohort studies (years). The quality assessment of incorporated studies was conducted by the same two authors using the Newcastle–Ottawa Scale(Reference Wells, Shea and O’connell22) (studies ≤ 3 scores are classified as low quality, 4–6 scores as moderate quality and ≥ 7 scores as high quality). The Kappa statistic for the agreement of the two investigators for quality assessment as well as data collection was computed. The disagreement was resolved through discussion or, if required, consulting a third researcher (Y.C.) (see online supplementary material, Supplemental Table 3).
Statistical analysis
The pooled risk ratio (RR) and 95 % CI were calculated for the highest DII score category v. the lowest DII score category. The reported OR and hazard ratios were considered as equivalent to RR(Reference Greenland23). Cochrane’s Q and I 2 statistics were used to assess the heterogeneity across studies(Reference Higgins, Thompson and Deeks24). The Cochrane’s Q test P-value < 0·10 or I 2 > 50 % was considered to be significantly heterogeneous(Reference Higgins, Thompson and Deeks24). In the presence of significant heterogeneity, the random-effects model was chosen to combine the results; otherwise, the fixed-effects model was selected.
We used the generalised least-squares trend estimation, the methods developed by Orsini et al. (Reference Orsini, Bellocco and Greenland25), to measure the linear dose–response relation of the DII levels and PC risk. The results were combined using a random-effects model. The median in each category of the DII was assigned. If medians were not reported, the midpoint of lower and upper limits was designated as the assigned dose. If highest category or the lowest category was open-end, we used the reported maximum and minimum range of the DII, respectively. If the maximum and minimum range had not been reported, we estimated it from the reported mean and sd values (mean ± 3 sd) in the study. The potential nonlinear dose–response relationship was examined through restricted cubic splines with 4 knots at fixed percentiles (5, 35, 65 and 95 %) of the distribution(Reference Orsini, Li and Wolk26). Then the estimates were combined using the restricted maximum likelihood method in a multivariate random-effects meta-analysis(Reference Jackson, White and Thompson27). A P nonlinearity of the meta-analysis was calculated by testing the null hypothesis that the coefficient of the second spline was equal to zero(Reference Desquilbet and Mariotti28).
Stratification analyses were conducted to explore the potential sources of heterogeneity whenever possible, by stratifying study designs, regions, the number of the DII components and adjustment factors. Sensitivity analyses were conducted to test the potential effect of each study on pooled effect size, by excluding each single study in each turn. We did not test publication bias with the formal statistical tests, because they have limited power when there are < 10 studies. The quality of evidence for result was assessed using the Grading of Recommendations Assessment, Development, and Evaluation and ranked as high, moderate, low or very low. All calculations were implemented using Stata/se version 15.0 (Stata Corp).
Results
Study characteristics
Figure 1 shows the detailed retrieval and screening process of literature. A total of 404 studies were identified. Eleven full-text articles were evaluated for eligibility. Finally, six studies(Reference Accardi, Shivappa and Di Maso29–Reference Zheng, Merchant and Wirth34) met the inclusion criteria for meta-analysis, including two prospective cohort studies(Reference Zheng, Wirth and Merchant33,Reference Zheng, Merchant and Wirth34) with a total of 634 705 individuals (3152 incident cases) and four case–control design studies(Reference Accardi, Shivappa and Di Maso29–Reference Shivappa, Bosetti and Zucchetto32) with 2737 cases and 4861 controls. Three studies were performed in the USA(Reference Antwi, Oberg and Shivappa30,Reference Zheng, Wirth and Merchant33,Reference Zheng, Merchant and Wirth34) , two in Italy(Reference Accardi, Shivappa and Di Maso29,Reference Shivappa, Bosetti and Zucchetto32) and one in the USA, Italy and Asia(Reference Antwi, Bamlet and Pedersen31). The FFQ were used in all studies to estimate dietary intakes and the DII was calculated by the method developed by Shivappa et al. (Reference Shivappa, Steck and Hurley15). All studies reported the adjustment for smoking as the exposure. Four studies(Reference Accardi, Shivappa and Di Maso29,Reference Shivappa, Bosetti and Zucchetto32–Reference Zheng, Merchant and Wirth34) reported the adjustment for alcohol consumption and one study(Reference Antwi, Bamlet and Pedersen31) reported adjustment for history of PC as the exposure. The Kappa statistic for the agreement of the two investigators for data collection was 0·9511. The Newcastle–Ottawa Scale scores of all included studies reached with 7–8 stars and the Kappa statistic for the agreement of the two investigators for quality assessment was 0·6667. More characteristics of the included studies are shown in Table 1.
Fig. 1 Literature search and study selection process for inclusion in meta-analysis of dietary inflammatory index and pancreatic cancer risk
Table 1 General characteristics of included studies in the meta-analysis of DII score and pancreatic cancer risk
DII, dietary inflammatory index; NOS, Newcastle–Ottawa Scale; NR, not reported.
* a and b represent different studies with the same author name.
Meta-analysis of dietary inflammatory index score and pancreatic cancer risk
Overall, the pooled RR indicated that individuals in the highest category compared with the lowest category had a 45 % increased PC risk (RR = 1·45; 95 % CI 1·11, 1·90; P = 0·006) when using a random-effects model (see Fig. 2). However, significant heterogeneity among the incorporated studies was also found (I 2 = 88·8 %, P < 0·001).
Fig. 2 Forest plots for risk ratios (RR) of the highest compared with the lowest category of DII score and pancreatic cancer. *a and b represent different studies with the same author name
Two studies(Reference Accardi, Shivappa and Di Maso29,Reference Antwi, Bamlet and Pedersen31) did not present cut-off points for DII. The dose–response meta-analysis of four studies(Reference Antwi, Oberg and Shivappa30,Reference Shivappa, Bosetti and Zucchetto32–Reference Zheng, Merchant and Wirth34) indicated that a 1-unit increase in the DII score was associated with the PC risk (RR = 1·08; 95 % CI 1·002, 1·166; P = 0·045; I 2 = 94·1 %, P < 0·001). Nonlinear dose–response meta-analysis showed an increased risk of moving from fewer to more inflammatory borders with increasing DII (P nonlinearity = 0·003; I 2 = 76·5 %, P < 0·001) (Fig. 3). The Wald test indicated dose–response relation was consistent with a nonlinear model (χ 2 = 9·8, P = 0·0074).
Fig. 3 Dose–response associations between the dietary inflammatory index and pancreatic cancer risk. 
, spline model; 
, linear model; 
, 95 % CI
Subgroup analyses
Due to the high heterogeneity among the incorporated studies, subgroup analyses were performed by stratifying study designs, regions, the number of the DII components and adjustment factors. The results are shown in Table 2. The summary RR indicated a significant positive correlation between high DII score and PC risk was found in the case–control subgroup (RR = 1·70; 95 % CI 1·16, 2·50; P = 0·007), with high heterogeneity (I 2 = 87·8 %, P < 0·001). On the contrary, the cohort subgroup had lower heterogeneity (I 2 = 39·3 %, P = 0·20) but no significant increase in the risk of PC (RR = 1·03; 95 % CI 0·80, 1·34; P = 0·81) (see Fig. 2). In subgroup analysis on countries and regions, a nonsignificant association was found. A significant positive association appeared to be in studies with ≤ 31 DII components (RR = 1·76; 95 % CI 1·14, 2·72; P = 0·011) but not in studies with >31 components (RR = 1·05; 95 % CI 0·83, 1·32; P = 0·70), and both groups had medium–high heterogeneity (≤31 DII components, I 2 = 87·5 %, P < 0·001; > 31 DII components, I 2 = 57·0 %, P = 0·127). Positive correlations were not observed in studies stratified to adjust for alcohol consumption (RR = 1·28; 95 % CI 0·98, 1·66; P = 0·07; I 2 = 84·8 %, P < 0·001), PC family history (RR = 1·23, 95 % CI 0·92, 1·65; P = 0·17), race (RR = 1·77; 95 % CI 0·87, 3·59; P = 0·12; I 2 = 91·0 %, P = 0·001) and total energy intake (RR = 1·03; 95 % CI 0·80, 1·34; P = 0·81; I 2 = 39·3 %, P = 0·20). Subgroup analyses indicated study design and adjustment for total energy intake appeared to be the potential sources of the heterogeneity.
Table 2 Subgroups analyses of studies reporting the pancreatic cancer risk for the highest v. the lowest category of DII score
DII, dietary inflammatory index; RR, risk ratio.
Sensitivity analyses
Sensitivity analysis (Fig. 4), after excluding the study of Antwi (a) et al. (Reference Antwi, Oberg and Shivappa30), also obtained positive affected results (RR = 1·26; 95 % CI 1·01, 1·58; P = 0·038), and heterogeneity slightly decreased (I 2 = 80·0 %, P = 0·001). Meanwhile, the exclusion of another individual study also did not have a significant impact on the results of the meta-analysis, indicating the stability of the results.
Fig. 4 Sensitivity analysis was conducted by removing each study in turn and recalculating the pooled risk ratio (RR) estimates. *a and b represent different studies with the same author name
Grading of Recommendations Assessment, Development, and Evaluation evidence
The Grading of Recommendations Assessment, Development, and Evaluation evidence for meta-analysis of DII score and PC risk is shown in online supplementary material, Supplemental Table 4. The quality of evidence for the result of the meta-analysis of DII score and PC risk was moderate.
Discussion
This meta-analysis summarised the evidence on the relationship between DII score and the incidence of PC. The current study included two prospective cohort studies and four case–control studies, with a total sample size of 642 303 participants. Meta-analysis findings indicated that the highest DII score might increase PC risk up to 45 % compared with the lowest DII score. The dose–response meta-analysis indicated that a 1-unit increase in the DII score was associated with an 8 % increase in PC risk. Nonlinear dose–response meta-analysis showed an increased risk of moving from fewer to more inflammatory borders with increasing DII score. Previous studies suggested better diet quality was associated with a reduced PC risk. Meanwhile, a recent systematic review including eight cohort studies and eight case–control studies indicated that PC risk and the animal products, starch-rich and western dietary patterns had significant positive associations. Significant inverse correlations were observed between the risk of PC and the specified dietary pattern of vegetables and fruits, vitamins and fibre(Reference Zheng, Guinter and Merchant35). A review from the International Agency for Research on Cancer and another 2017 review both reported being obese or overweight increases the risk of developing PC, probably because obesity produces an inflammatory state, increasing IL-6 in pancreatic tissue(Reference Lauby-Secretan, Scoccianti and Loomis36,Reference Kerr, Anderson and Lippman37) . This review further indicated that a higher DII score (means a more pro-inflammatory and less anti-inflammatory diet) may increase the risk of PC(Reference Lauby-Secretan, Scoccianti and Loomis36).
Subgroup analyses showed case–control subgroup had a more significant positive correlation between the DII score and PC risk compared with the cohort subgroup. However, case–control studies are more prone to bias in selection, recall and reverse causality. This difference may be due to the fact that there are only two articles included in the cohort subgroup, and both articles are from the same author(Reference Zheng, Wirth and Merchant33,Reference Zheng, Merchant and Wirth34) . In the subgroup analysis, the number of components used to calculate DII scores decreased and the associations increased instead, suggesting that the reduction in the number of components strengthened the link between DII scores and PC risk. However, owing to the small number of studies, we were unable to explore how this relationship changed with changes in DII components. Therefore, this result had some limitations and needed to be interpreted carefully. The eating habits of Western and Asia are different. However, in subgroup analysis based on countries and regions, three subgroups have similar results on the link between DII score and PC risk. The reason may be that only one study involved the Asian race(Reference Antwi, Bamlet and Pedersen31). Meanwhile, the current study did not perform subgroup analysis based on region, only based on the whites and other races. Moreover, the results were adjusted for race. We further explored the risk factors adjustment, including a family history of PC, race, alcohol consumption and total energy intake, using subgroup analyses, but the positive association was not found.
Several potential mechanisms have been proposed to explain the possible association between pro-inflammatory diet (means higher DII score) and PC risk. In PC, down-regulated tumour suppressor miRNA and up-regulated oncogenic miRNA are considered to be associated with tumour growth and metastases(Reference Hawa, Haque and Ghosh38). Growth factors and cytokines might enhance or inhibit the expression levels of miRNA. Also, the expression of growth factors, cytokines and their receptors might be regulated by miRNA(Reference Padoan, Plebani and Basso39). Meanwhile, anti-TNF-α can up-regulate some miRNA (i.e., hsa-miR-23a, hsa-miR-197 and hsa-miR-221, etc.) which are considered to be associated with PC(Reference Padoan, Plebani and Basso39), which means cancer-associated inflammatory responses might be influenced by the balance between cancer-derived and inflammatory cell-derived cytokines and chemokines. In this case, TNF-α, produced by tumour cells themselves and tumour-infiltrating inflammatory cells, is involved in tumourigenesis, growth, metastasis and anti-cancer immune regulation(Reference Padoan, Plebani and Basso39). In a systematic review, the authors summarised that levels of IL-2, IL-6, IL-8, IL-10, macrophage colony-stimulating factor, macrophage inhibitory cytokine-1 and vascular endothelial growth factor were higher in patients with PC than in those without PC(Reference Yako, Kruger and Smith40). And the DII scoring system exactly focuses on the dietary inflammatory potential.
Another potential mechanism between inflammatory and PC was the activation of transcription factors NF-κB. A recent study reported that the activation of NF-κB promotes the production of growth and differentiation factor 15, a member of the tumour growth factor-β superfamily, in pancreatic cells. Growth and differentiation factor-15 acts on tumour-related macrophages by inhibiting NF-κB signalling and reduces the synthesis of TNF-α, thereby reducing TNF-α-dependent tumour cell apoptosis and enhancing tumour growth(Reference Ratnam, Peterson and Talbert41).
Some limitations should be considered in the present meta-analysis. First, owing to the small number of studies and no randomised controlled trials included, meta-analyses results may be affected, but the included studies all have good quality. Second, high heterogeneity was observed in the meta-analysis results. Further subgroup analyses and sensitivity analyses were conducted, and it was found that the heterogeneity might be due to the different designs of the included studies. Third, we did not assess the publication bias using the formal statistical tests as they have insufficient power when there are limited studies (n < 10). Besides, the research population is mainly limited to Europe and the USA, which may lead to the results cannot be promoted globally.
The present meta-analysis also has some strengths. First, the DII score used in each study is calculated based on the same dietary assessment tool (FFQ), which increases the comparability of the study. Second, the exclusion of another individual study also did not have a significant impact on the results of the meta-analysis, indicating the stability of the results.
In conclusion, the analysis of the evidence from included studies suggested dietary habits with high inflammatory features (high DII score) might increase PC risk. However, these findings should be interpreted with caution due to the limited number of studies and potential bias, and the need for further validation. Future studies would benefit from improved designs, larger sample sizes and better confounding controls and should emphasise the potential dose–response effect on DII score and PC risk.
Acknowledgements
Acknowledgments: We would like to thank Ph.D. Min Zhao for providing language help and writing assistance. Financial support: The current study was supported by the National Natural Science Foundation of China (grant no. 81701950), Medical Research Projects of Chongqing (grant no. 2018MSXM132) and the Kuanren Talents Program of the Second Affiliated Hospital of Chongqing Medical University (grant no. KY2019Y002). Conflict of interest: There are no conflicts of interest. Authorship: Z.G. and Y.C. designed the research; Z.G., Y.H. and Y.C. conducted the research; Z.G. analysed the data; Z.G. wrote the paper. Y.C. had primary responsibility for final content. All authors read and approved the final manuscript. Ethics of human subject participation: not applicable. Because analysis is only based on published article data.
Supplementary material
For supplementary material accompanying this paper visit https://doi.org/10.1017/S1368980021001579
