Diet Quality and Pancreatic Cancer Incidence in the Multiethnic Cohort

Pancreatic cancer is widely regarded as one of the most lethal cancers. The 5-year survival rate of 10% for this cancer is among the lowest of all cancers; so, even though incidence is relatively low, it is the third leading cause of cancer death in individuals between the age of 40 and 79 and the fourth leading cause of death across all age groups (1). The late onset of symptoms coupled with a lack of available screening tools have resulted in very poor outcomes (2). The American Cancer Society estimates that there will be 62,210 new cases of pancreatic cancer in the United States in 2022 with 49,830 deaths (3). The clinical, societal, and economic burden of pancreatic cancer is 3.9% of all disability-adjusted life years for cancers (4).

There is a clear need to identify reliable preventive measures in an attempt to reduce disease burden. Specifically, individuals in high-risk categories may benefit from modifiable measures that they can incorporate into their lifestyles to reduce risk. Well established modifiable risk factors for pancreatic cancer include smoking, diabetes, increased body mass index (BMI), and physical inactivity (5), but the association between overall diet quality and pancreatic cancer remains unclear, as results from cohort studies have been particularly mixed (6–10).

In this study, we aimed to address this research gap across diverse populations by examining the relationship of overall diet quality and the inflammatory potential of the diet with pancreatic cancer incidence in the Multiethnic Cohort Study (MEC). We also investigated if this relationship varied by sex, race and ethnicity, smoking status, or BMI because the impact of diet may be particularly important for specific high-risk subgroups.

The MEC is a population-based prospective study that was established to elucidate lifestyle factors and incidence of cancer and other chronic diseases across various races and ethnicities (11). The five races and ethnicities included in this study were African American, Japanese American, Latino, Native Hawaiian, and White. The cohort was assembled between 1993 and 1996 using a mailed questionnaire that was specifically designed to assess the participants’ lifestyle habits including a food frequency questionnaire (FFQ), consisting of over 180 food items that accounted for greater than 85% of the major nutrients of interest. At baseline, participants were between the ages of 45 and 75 years old and were recruited from the entire state of Hawaii as well as nine counties in California: Los Angeles, Riverside, Orange, San Bernardino, San Diego, San Francisco, San Mateo, Contra Costa, and Alameda (11). The FFQ showed satisfactory correlations for nutrients as energy densities (0.57–0.74) with three 24-hour recalls for all sex-ethnic groups in a calibration study (12). The baseline questionnaire also collected information on key demographic (i.e., age, sex, race and ethnicity, education, and occupation) and lifestyle history and habits including personal and family medical history, anthropometry (i.e., body weight and height), cigarette smoking, and physical activity. Questionaries were administered to all participants at baseline. Willingness to mail back the questionnaire was accepted as implied consent for participation in the study, which has been approved by the institutional review board at University of Southern California and University of Hawaii.

Of the more than 215,000 MEC participants, this analysis included 177,313 participants after the following exclusions: previous history of pancreatic cancer (n = 58), not in one of the five main racial and ethnic groups (n = 12,295), and missing information on covariates: smoking status or intensity (n = 7,835), BMI (n = 1,866), history of diabetes (n = 2), education (n = 861), physical activity (n = 8,833), and invalid diet based on total caloric intake or individual dietary components (n = 6,638). We determined invalid diet based on total energy intake or its components. Specifically, we computed a robust standard deviation (RSD) of total energy intake based on the middle 80% of normal distribution. Then, we excluded all individuals with energy values out of the ranges of the mean ± 3 RSD. A similar approach was applied to fat, protein, or carbohydrate intakes to exclude individuals outside the range of mean ± 3.5 RSD (13). Although the MEC specifically targeted five main race and ethnicity groups (African American, Japanese American, Latino, Native Hawaiian, and White), participants from other backgrounds were included in the study because actual ethnicity was not known at the time the questionnaire was mailed (11). Participants not in these five groups were not included in this analysis due to the small number of pancreatic cancer cases within each group as well as their distinct dietary patterns.

Diet quality was assessed by previously developed and validated DQIs: the Healthy Eating Index-2015 (HEI-2015), the Alternative Healthy Eating Index-2010 (AHEI-2010), the alternate Mediterranean Diet (aMED) score, and the Dietary Approaches to Stop Hypertension (DASH) score. The inflammatory potential of the diet was assessed with the energy-adjusted Dietary Inflammatory Index (E-DII). These measurements of diet were selected and computed for the MEC as part of the Dietary Patterns Methods Project (DPMP) due to their particular relevance for dietary guidance commonly used in U.S. populations (14). The components of each measurement of diet are shown in Supplementary Table S1. The HEI-2015 is based on a 13-component system of food groups, food intake variability, and nutrients (15). It has a maximum score of 100 accumulated from a score of 0 to 5 on 6 of the components and a score of 0 to 10 on each of the seven components; a higher score represents adherence to a healthier diet. The AHEI-2010, a measure of dietary patterns consistently associated with lower risk of chronic disease, is based on 11 components each scored from 0 (nonadherence) to 10 with a maximum score of 110 (16). The aMED is an adaptation of the Mediterranean diet that was originally developed by Trichopoulou and colleagues (17). It is scored on the basis of nine components, each component was assigned 1 point for intakes above the median and 0 points for intakes below the median, resulting in an overall score between 0 and 9 (18). The DASH diet score is based on eight components. For components where higher intake was desired, such as fruits, vegetables, and whole grains, quintiles with the lowest adherence were given 1 point whereas the highest received 5 points; because red or processed meats, added sugars, and intake of sodium were meant to be minimized by the use of this diet, these components were reverse coded resulting in an overall score of 8 to 40 (19). The E-DII was developed on the basis of an extensive literature review carried out to assess the role of diet in influencing 6 inflammatory markers (20, 21). Both the DII and E-DII were designed to quantify the inflammatory potential of the entire diet on a scale from maximally anti-inflammatory to maximally proinflammatory based on individual nutrients and other dietary constituents. A global reference database was used to compute DII scores (20); for the E-DII, a global reference database of energy-adjusted nutrient scores was used (21). Only 28 of the 45 food components of the E-DII were included in the E-DII calculations for the MEC because data on the 17 components were not available from the FFQ (Supplementary Table S1; ref. 22). In contrast with the other DQIs, a lower E-DII score denotes a more anti-inflammatory diet, whereas a higher score represents a more proinflammatory diet and is indicative of poorer diet quality.

Incident cases of pancreatic cancer were identified by linkage of the MEC with statewide Surveillance, Epidemiology, and End Results (SEER) cancer registries in Hawaii and California. Cases were identified using the International Classification of Diseases for Oncology, 3rd Edition [ICD-O-3] codes C25.0-C25.9, excluding C25.4 (endocrine tumor). Deaths were identified on the basis of state death certificate files and the National Death Index. A total of 1,779 cases of pancreatic cancer were identified among 177,313 eligible participants after an average of 19.3 years of follow-up through December 31, 2017.

Hazard ratios (HRs) and 95% confidence intervals (CI) were calculated using Cox proportional hazards regression with age as the time metric to estimate an association between the dietary indices and pancreatic cancer risk. Each diet score was assessed using quintiles based on distributions across the entire cohort where the quintile representing the poorest diet was used as a reference for the other four quintiles of each diet separately. The results using sex-specific and race/ethnicity-specific quintiles remained similar, therefore we used the common cut-points in the current analysis to make the interpretation of the results more straightforward. For the HEI-2015, the AHEI-2010, the aMED, and the DASH, the poorest diet was represented by the lowest quintile whereas for the E-DII, the poorest diet was the highest quintile. Multivariate analyses were adjusted for the following covariates: age at cohort entry, history of diabetes (yes or no), family history of pancreatic cancer (yes or no), total energy intake (log-transformed kcal/day), BMI (<25, 25 to 29.9, and ≥30 kg/m²), and education (≤12 years of schooling, some college or vocational training, and college graduate and above). Sex (male or female), race and ethnicity (African American, Japanese American, Latino, Native Hawaiian, and White), vigorous physical activity (hours of vigorous activity per day categorized as 0, below the median number of hours, or above), and pack-years of smoking for former and current smokers (never, past with <20 pack-years, past with ≥20 pack-years, current with <20 pack-years, and current with ≥ 20 pack-years) were used as strata variables in evaluating baseline hazard functions. Covariates were selected for inclusion into the model based on a priori knowledge of risk factors associated with pancreatic cancer in the MEC (23). Models for the HEI-2015 and the DASH were further adjusted for alcohol use (<14, 14 to 28, and ≥28 g/day) as alcohol was not included in these two indices. Proportional hazard assumptions were tested using Schoenfeld residuals (24); covariates that violated the assumptions (smoking status with pack years and vigorous activity) were included as strata. Heterogeneity was tested using the Wald statistics for cross-product terms.

Stratified analysis was performed for sex, race and ethnicity, smoking status, and BMI category. For the analysis by smoking status, current and past smokers were grouped together as ever smokers with adjustment for pack-years. For analysis involving BMI, participants were divided into two groups based on their BMI (>25 or ≤25 kg/m²), and adjusted for continuous BMI within each group.

The eight components of the DASH diet score were analyzed individually to test for independent associations to see if any one particular component of the DASH score was driving the association; separate Cox models were run for each component adjusted for the specified covariates as well as the other seven components of the DASH by removing each component individually from the total DASH score. All analyses were performed with SAS 9.4 (SAS Institute, Inc.). All P values were two-sided.

The de-identified data underlying this article can be shared on reasonable request to the corresponding author upon approval by the MEC Research Committee.

At baseline, the entire MEC cohort was 45% men and 55% women; participants were 33% age 45 to 54, 32% age 55 to 64, and 35% age 65 to 75 (11). Of the 177,313 participants, 81,408 (46%) were men and 95,905 (54%) were women. The largest racial/ethnic group was Japanese Americans (n = 51,682), followed by Whites (n = 44,791), Latinos (n = 39,208), African Americans (n = 28,670), and the smallest group was Native Hawaiians (n = 12,962). Compared with men and women in the lowest quintile of the DASH, both men and women in the highest quintile tended to be Whites followed by Japanese Americans and African Americans. Participants with the highest DASH scores also tended to be older, more educated, and less obese; men in this quintile were more likely to be former smokers with less than 20 pack years followed by never smokers, whereas women were primarily never smokers. All key baseline characteristics are shown in Table 1.

There was no association between healthier diets, as defined by the DQIs and E-DII, and pancreatic cancer risk after adjusting for covariates (Table 2). Results were similar in men and women; tests for heterogeneity showed no sex differences in association (all P_{heterogeneity} ≥ 0.48). No significant associations were found among the diet scores and pancreatic cancer incidence in racial and ethnic-specific analysis (Supplementary Table S2). Tests for heterogeneity were not statistically significant across the five racial and ethnic groups (all P_{heterogeneity} ≥ 0.44).

In the subgroup analysis by smoking status, risk of pancreatic cancer tended to be lower among ever smokers who adhered to the DASH diet (P_{heterogeneity} = 0.004). Ever smokers in the highest quintile of the DASH diet compared with those in the lowest quintile had a 25% decrease in pancreatic cancer risk (95% CI, 0.61–0.93; P_trend = 0.01), whereas no risk reduction was found among never smokers (HR, 1.05; 95% CI, 0.83–1.32; P_trend = 0.54). For the other four dietary indices we investigated, reduction of risk was not observed among either never or ever smokers, and testing for heterogeneity did not show differences in associations (Fig. 1; Supplementary Table S3).

Participants were divided by BMI as underweight or normal weight (<25 kg/m²) and as overweight or obese (≥25 kg/m²) for subgroup analysis (Supplementary Table S4; ref. 25). Although no significant heterogeneity (P = 0.15) or P_trend (P = 0.08) was observed, among overweight or obese individuals, the highest quintile of the DASH diet was associated with a 21% reduction in pancreatic cancer risk (95% CI, 0.65–0.97; Figure 2) compared with those in the lowest quintile. For under/normal weight, the HR was 1.23 (95% CI, 0.95–1.58; P_trend = 0.22; Supplementary Table S4).

In component analysis, the association between the DASH diet score and pancreatic cancer risk was not changed by removing any one of the eight dietary components (Supplementary Table S5).

To address possible reverse causality in the short term, we excluded cases diagnosed within the first 2 years of cohort entry (n = 68), and our results remained consistent. We also performed analyses with the 10-year follow-up questionnaire data, and results were similar to the original analyses. When the analyses were restricted to participants with no history of diabetes, similar results were also observed. The association between the highest quintile of the DASH diet and pancreatic cancer risk for individuals with BMI ≥25 kg/m² was 0.80 (95% CI, 0.64–0.99). For ever smokers, the association between adherence to the DASH diet and pancreatic cancer risk in participants with no history of diabetes was 0.76 (95% CI, 0.61–0.96).

In this large prospective study, diet quality was not associated with pancreatic cancer risk overall. This null association was similar in men and women and across racial and ethnic groups. However, we found that the DASH diet was associated with reduced pancreatic cancer risk in participants who were current or former smokers at cohort entry. Our results also suggest an inverse association between the DASH diet and pancreatic cancer risk among overweight or obese individuals.

Past studies have shown mixed results when examining pancreatic cancer risk with various dietary patterns, with some cohort studies showing no association as in this study. Our findings on E-DII are consistent with null results from two U.S. cohort studies, which examined E-DII as the dietary index of interest (6, 7). The Prostate, Lung, Colorectal, and Ovarian Cancer Screening Trial (PLCO) cohort study (6) with a small number of cases (n = 328) as well as the much larger NIH and American Association of Retired Persons (NIH-AARP) prospective cohort study (n = 2,824; ref. 7) found no association between the E-DII and risk of pancreatic cancer. When comparing the highest versus lowest quintile of HEI-2005 in the NIH-AARP study (n = 2,383; ref. 9), there was an inverse association with pancreatic cancer risk (HR, 0.85; 95% CI, 0.74–0.97). A slightly stronger association was seen among obese individuals (BMIs ≥ 25 kg/m²; HR, 0.81; 95% CI, 0.69–0.96) than among those with a normal weight. Results from the Singapore Chinese Health Study (8) based on 311 pancreatic cancers found that the highest quartiles of the AHEI-2010, aMED, and DASH diets were associated with a decreased pancreatic cancer risk overall (all P_trend < 0.05). In subgroup analysis, adherence to the DASH diet showed an inverse association among ever smokers but not in never smokers (P_{heterogeneity} = 0.02), but this was based on only three cases of pancreatic cancer among ever smokers in the highest quartile of DASH. The authors suggested that pancreatic cancer risk maybe mitigated by the protective effects of a high-quality diet more robustly in individuals with a higher risk for pancreatic cancer, such as smokers. A recent cohort study also based on the NIH-AARP Diet and Health Study (n = 3,137 cases) found a statistically significantly lower PDAC risk in the entire cohort for the HEI-2015, aMED, and DASH diet scores, but no association was shown for the AHEI-2010 (10). In the Women's Health Initiative (WHI) Study (26), adherence to a low-fat diet, similar to the DASH diet, was not associated with a decreased risk for all women in the study (HR, 0.86; 95% CI, 0.67–1.11). However, risk reduction was statistically significant among postmenopausal women with BMI >25 kg/m² (HR, 0.71; 95% CI, 0.53–0.96).

A 2017 systematic review of a priori dietary patterns as measured by the Mediterranean diet, the HEI-2005, or E-DII showed that eight case-control and cohort studies, including Arem and colleagues (9), consistently suggested an inverse association between diet quality and pancreatic cancer risk (27). Adjusted risk ratios ranged from 0.27 to 0.98. However, a stronger association was observed in case–control studies versus in cohort studies. Biases inherent to case–control studies could help explain the discrepancy. Recall bias can lead to differential misclassification biasing the results away from the null.

These previously described studies differ from our current study in a variety of ways, but most notably in the racial and ethnic diversity of their study populations. The NIH-AARP cohort is predominately White (93.7% non-Hispanic White for males and 90.9% non-Hispanic White for females, 28), which could potentially explain why there are differences between some of our results and those of studies based solely on this cohort (9, 10). Luu and colleagues (8) also examined a much different study population than ours and recruited Chinese participants in Singapore and analyzed 311 pancreatic cancer cases. The systematic review performed by Zheng and colleagues (27) included three cohort studies on a priori diet; of these three cohort studies, two were based on the NIH-AARP study (9, 10), and the other recruited participants from Northern Sweden (29).

Some research has supported lowering the cut off for obesity among Asians compared with Whites (30, 31), whereas others have shown that Whites and Asians have similar BMI-mortality associations (32, 33, 34). For the statistical analysis of this study, a BMI of 25 kg/m² was used as a cutoff for subgroup analysis to compare individuals who were overweight to those who were not, which is slightly below the average, age adjusted mean BMI in the United States between 2015 and 2016 (29.1 kg/m² for males and 29.6 kg/m² for females; ref. 35). The CDC has maintained a cutoff of a BMI of 25.0 kg/m² as overweight or pre-obesity among all races and ethnicities (36), which supports the use of this cutoff in this study.

It also is important to note that to calculate the E-DII scores for the MEC, only 28 of the 45 food components originally included for the DII development were used. The 17 missing parameters are all anti-inflammatory; so, their absence will tend to result in skewing toward the right for E-DII values (i.e., toward higher, more proinflammatory values). This, of course, would be an issue in studies where these foods are actually eaten; and this would be more likely in the MEC than in a standard American population (21). Indeed, E-DII scores in the MEC are generally low (more anti-inflammatory). So, even though the E-DII scores were computed from fewer components, they have been demonstrated to perform well in predicting inflammatory markers and health outcomes in other studies (37, 38, 39) as well as in the MEC (22, 40, 41).

There are inherent limitations to using self-reported dietary questionnaires such as reporting bias which can lead to measurement error (42). Other limitations include the limited sample size in subgroup analyses, residual confounding, and dietary patterns could be associated with other lifestyle factors and risks that were not adjusted for in the models. Due to the rarity of pancreatic cancer, most previous studies had a modest number of cases. This study benefited from a large number of cases (n = 1,779), a long follow up time, its prospective design, and the availability of potentially confounding variables in analyses. To our knowledge, this study is the first study to include a racially and ethnically diverse population and provided risk association analyses with five different dietary indices.

In summary, this study provides limited evidence for an association between overall diet quality and a pro-inflammatory diet and pancreatic cancer risk in both men and women in the MEC. There is some indication, however, that the DASH diet may potentially be protective among current and former smokers. Care needs to be taken when interpreting subgroup analyses as these results could be attributed to chance. Additional research is needed to draw more definitive conclusions as previous literature is ambiguous when examining the association between diet quality and pancreatic cancer risk.

D.O. Stram reports grants from NIH during the conduct of the study. J.R. Hébert reports grants and other support from Connecting Health Innovations LLC (CHI) outside the submitted work; also has a patent for Federally registered trademark for the DII issued, licensed, and with royalties paid from Connecting Health Innovations LLC (CHI). No disclosures were reported by the other authors.

H. Steel: Formal analysis, writing-original draft, writing–review and editing. S.-Y. Park: Formal analysis, investigation, methodology, writing–review and editing. T. Lim: Formal analysis, writing–review and editing. D.O. Stram: Investigation, writing–review and editing. C.J. Boushey: Methodology, writing–review and editing. J.R. Hébert: Methodology, writing–review and editing. L. Le Marchand: Resources, funding acquisition, writing–review and editing. A.H. Wu: Investigation, writing–review and editing. V.W. Setiawan: Conceptualization, resources, supervision, funding acquisition, project administration, writing–review and editing.

V.W. Setiawan received R01CA209798, and L. Le Marchand received U01CA164973. This work was supported by the NCI at the NIH, grant nos. R01CA209798 and U01CA164973.

The publication costs of this article were defrayed in part by the payment of publication fees. Therefore, and solely to indicate this fact, this article is hereby marked “advertisement” in accordance with 18 USC section 1734.