A continued international decline in gastric cancer mortality has been observed for a last century,1 but it remains worldwide the 4th most common cancer and the 2nd leading cause of cancer-related deaths.2 This decline may be related with a decrease in the prevalence of Helicobacter pylori infection,2 and for a long time it was explained as the result of changes in dietary habits related to the general improvement in living conditions.3 Diet, as a risk factor for gastric cancer, however, remains controversial. The protective effect of fruits and vegetables consumption is less pronounced than initially accepted4 and in addition to an increased risk with salt exposure,5 no further consistent associations were disclosed. Some studies suggest a relatively greater impact of environmental factors on the development of the intestinal cancer,4, 6 as well as differences between cardia and non-cardia cancers.7
Study design drawbacks, such as inexistence of information on the anatomical location of the tumor or its histological type, overlooking H. pylori infection as a potential confounding or effect modifier of the association between lifestyles and gastric cancer, or unreliable dietary intake assessment have contributed to conflicting findings. Studies evaluating individual foods or nutrients are limited in their ability to consistently disclose meaningful associations with cancer. There are important correlations between the intakes of various foods and nutrients, largely due to individual behavioral patterns,8 and the analysis of dietary patterns allows the aggregation of individuals with similar diets, providing an additional tool to understand the role of diet as an etiological factor for cancer.
Among Southern European countries, Portugal has the 2nd highest death rate for gastric cancer in males, and the highest in females,9 reflecting the late inflection in the mortality trends, observed only in the 1970s,10 and a mediocre decline in the last decades.11 The high prevalence of H. pylori infection in the adult Portuguese population, over 80%,12 contributes to the high gastric cancer incidence9 and mortality rates.1 Such research setting is appropriate to assess the effect of other environmental factors among infected subjects, and to understand which factors modulate the progression to cancer when infection is present.
The aim of this study was to identify dietary patterns in a Portuguese population and to quantify their association with the occurrence of gastric cancer, by location and histological type, accounting for the effect of H. pylori infection status.
Material and Methods
This is a case-control study involving incident cases of gastric adenocarcinoma admitted to the Surgery wards of the 2 major public hospitals caring for cancer patients in the North of Portugal (Hospital de S. João and Instituto Português de Oncologia Francisco Gentil, both in Porto) and appropriate community controls selected among Porto dwellers.
From June 2001 to December 2006, we evaluated 709 cases of stomach adenocarcinoma following a previously described procedure.13 In brief, subjects were eligible if there was no previous cancer diagnosis, except for skin non-melanoma, and no subtotal gastrectomy for benign conditions. The interview took place during in-hospital stay, shortly after admission, mostly before surgical treatment. To assess cognitive function, all individuals above 64 years had a Mini Mental State Examination, which resulted in the exclusion of 42 cases scoring less than 18.14
Cancer was diagnosed according to the routine procedures of both institutions. To ensure a standard pathologic classification, a single experienced pathologist reviewed all pathology reports, and slides were reassessed by 3 pathologists whenever routine information was considered insufficient or inconsistent, allowing its reclassification according to the Laurén's criteria. The diagnosis was based on gastrectomy specimens, endoscopic biopsy material, or the evaluation of metastasis. Patients submitted to gastrectomy had the tumors classified as intestinal, diffuse, or unclassifiable according to Laurén. No significant differences were found between these patients and those from whom only endoscopic biopsies or metastatic tissue was available regarding age (median: 64 vs. 62 years, p = 0.965), gender (males: 63.1% vs. 56.6%, p = 0.203) and education (median number of school years: 4 vs. 4, p = 0.467). All cases had the anatomic site classified following image or pathology descriptions, and gastric adenocarcinoma considered of the cardia (defined as cardioesophageal junction, oesophagogastric junction and gastroesophageal junction), non-cardia (all other specified sites) or nonaccurately classifiable. A cardia location was more frequent in cases not submitted to gastrectomy (26.9% vs. 11.3%, p < 0.001).
A representative sample of the adult population of Porto was assembled as part of a health and nutrition survey and used as source of controls. A detailed description of the general selection procedures and participants characteristics was published.15, 16 In brief, participants were recruited by random digit dialing using households as the sampling frame, followed by simple random sampling to select one eligible person among permanent residents in each household. The selected participants were invited to visit our department for interview, without allowing replacements of refusals. The overall sample comprised 2,485 community controls, aged 18 to 93, corresponding to a participation proportion of 70%.16 Twenty-one subjects aged above 64 years scoring less than 18 in the Mini Mental State Examination14 were not considered eligible.
Evaluation of participants
Trained interviewers inquired both cases and controls using a structured questionnaire providing information on demographic, social, behavioral and medical characteristics, as previously described.13 A venous blood sample was collected and serum kept frozen at −20°C until analysis.
Dietary habits were recorded using a semiquantitative food frequency questionnaire (FFQ) comprising 82 food items or beverage categories. It was designed according to Willett17 and adapted by inclusion of a variety of typical Portuguese food items. The choice of relevant food items to be included in the FFQ was based on their contribution to the between-person variance of intake of total energy, proteins, fat, carbohydrates, cholesterol, dietary fiber, vitamin A, provitamin A carotenoids, vitamins C and E, calcium, alcohol and caffeine. Certain foods with similar nutrient composition were grouped together as a single item. The FFQ was validated by 2 different methods, with a 7-day food record in 75 female and 71 male community participants and, regarding the fatty acid composition, with the composition of subcutaneous adipose tissue in 64 female and 52 male subjects.18
For each FFQ item subjects were asked the average frequency of consumption (9 possible responses ranging from never to 6 or more times per day), the portion size usually consumed (based on a photograph manual with small, medium and large portion sizes). This information was used to estimate, for each FFQ item, the consumption in grams (or milliliters for beverages), corrected for seasonality, by multiplying the reported consumption by the ratio between the number of months during which the food item was reported to be consumed and 12 (months).
The FFQ evaluated dietary intake for the previous year period, or the year before onset of symptoms, when appropriate, for cases. Complete information on dietary habits was obtained for 628 out of 667 cases, from which, 37 declared to have changed their food intake 12 or more months before interview, due to gastrointestinal symptoms, and were excluded from this analysis. Subjects that modified their habits during the previous year were not excluded, but were asked to recall dietary intake in the year before the change. For each of the remaining 591 cases we selected up to 4 age and gender frequency-matched controls, resulting in 1,585 controls. As more than 1 control per case was usually available, we opted for those more recently evaluated. Cases and controls differed in age (median: 63 vs. 59 years, p < 0.001) and gender (males: 62.1% vs. 48.9%, p < 0.001) distribution because available controls did not always allowed a 1 to 4 ratio, especially for older males.
Among participants eligible for analysis of dietary habits, 449 out of 591 cases, and 1,306 out of 1,585 matched controls had a blood sample available. Assessment of serum anti-H. pylori IgG titers was performed by ELISA [Anti-H. pylori ELISA (IgG), Eurolmmun®, Lubeck, Germany]. Participants were classified as negative if they had less than 16 RU/ml and as positive if their antibody concentration was equal or greater than 16 RU/ml, according to the manufacturer's instructions. Cases for whom a blood sample was not available were not significantly different regarding age (median age: 65 vs. 62 years, p = 0.146), gender (males: 68.3% vs. 60.1%, p = 0.080) or education (median number of school years: 4 vs. 4, p = 0.275). Controls for whom a blood sample was not available were significantly older (median age: 62 vs. 58 years, p < 0.001) and less educated (median number of school years: 4 vs. 6, p < 0.001), but there were no significant gender differences (males: 47.7% vs. 49.2%, p = 0.652).
The 82 FFQ items were initially combined into 23 different food and beverages groups, as follows: alcoholic beverages (including wine, beer, spirits, Porto wine and liqueurs); bread (including bread, whole grain bread and cornbread); cereals and potatoes (including rice, pasta, potatoes and cereals); codfish; coffee; canned fruits; dairy products (including semi-skim milk, skim milk, whole milk, cheese, yogurt and ice-cream); eggs; fast-foods/fried snacks (including fried snacks, pizzas, ham, bacon and sausage); fish and seafood (including fish, except codfish, and seafood); fruits (including fruits and natural fruit juices); soft-drinks (including nonalcoholic beverages, except tea, coffee, water and natural fruit juices); dried fruits, hereafter referred as nuts; oils and fats (including olive oil, vegetable oil, butter, margarine and mayonnaise); olives; red meat (including beef, pork and hamburgers); white meat (including chicken, turkey and rabbit); salads (including onions, lettuce, carrots, tomatoes, sweet peppers and cucumber); ketchup hereafter referred as sauces; vegetable soup; sugar and sweets (including biscuits, croissants, chocolate, marmalade and sugar); black tea, hereafter referred as tea and vegetables (including spring greens, turnip greens, broccoli, green beans, turnips, beans, peas). The overall intake in each food group was established by adding up the amounts of single items or groups consumed per day.
Principal components analysis was conducted to derive noncorrelated variables based on the 23 food and beverages groups. Because principal components analysis is sensitive to outliers, 122 controls for whom the standardized variables were ≥5 times or ≤5 times the standard deviation of the mean were excluded. Principal components analysis identifies foods that are frequently consumed together, aggregating food/beverage items or groups on the basis of the degree to which they are correlated with each other. To identify the number of principal components to be retained, we used the criterion of eigenvalues greater than 1 and the scree plot. The factors were rotated by an orthogonal transformation to achieve a simpler structure with greater interpretability. After Varimax rotation, factor scores were saved from the principal components analysis for each individual.
The factors defined by principal components analysis were used in cluster analysis to identify groups with similar characteristics. The number of clusters was defined with hierarchical cluster analysis, performed with squared Euclidean distances used in the proximities matrix, and Ward's method used as clustering method. The Calinski and Harabasz pseudo-F stopping rule index and the Duda and Hart Je(2)/Je(1) index were used jointly to select the number of clusters for the analysis.19
Proportions were compared using the χ2 test. Comparisons of quantitative variables between dietary patterns were conducted using the Kruskal–Wallis test for multiple comparisons, considering a level of significance adjusted for multiple comparisons [significance level divided by k (k − 1), where k represents the number of groups], of 0.0083. The association between dietary patterns and gastric cancer risk was quantified with odds ratio (OR) and the corresponding 95% confidence intervals (95% CI) computed by unconditional logistic regression. The statistical analysis was conducted using STATA®, version 9.2.
Eight uncorrelated variables were defined by principal components analysis, explaining 49.7% of the total variance (Table 1), and 3 dietary patterns were identified, by cluster analysis, which may be described as follows: pattern (I) high consumption of dairy products, fruits, salads and vegetables, and low consumption of meat and alcoholic beverages; (II) low consumption of most food groups, specifically dairy products, fish and seafood, fruits, salads, vegetables and meat; (III) high consumptions of most food groups and lowest vegetable soup intake (Table 2).
Table 1. Eight major groups identified by principal components analysis
Table 2. Characteristics of consumption (grams per day) across the 3 dietary patterns
Controls were included predominantly in pattern I (57.6%), and less often in pattern III (17.3%). Compared to controls in patterns I and II, those in pattern III were more likely to be male (74.7% vs. 38.7% and 49.7%), younger (median age: 50 vs. 60 and 64 years), more educated (median number of education years: 8 vs. 6 and 4 years) and to have a higher total energy intake (2,651 vs. 2,084 and 1,885 kcal/day). (Table 3). Among cases, a similar proportion of subjects was included in patterns I and II, and only 11.8% were in pattern III. There was a male predominance in all patterns, but pattern III had a higher proportion of males (77.1%). The age and total energy intake distribution across patterns was similar to the observed in the controls and there were no differences regarding median education.
Table 3. Sample characteristics across the 3 dietary patterns in controls and in cases1
Compared to dietary pattern I, pattern II was associated with a significant 1.7-fold increase in the risk of gastric cancer, but no significant association was found for pattern III (OR = 0.80, 95% CI: 0.57–1.14). H. pylori infection was not an effect modifier (P for interaction = 0.166) of the association between dietary patterns and cancer or responsible for an important confounding effect (Table 4).
Table 4. Association between dietary patterns and gastric cancer, according to Helicobacter pylori infection status
The ORs for the comparison of patterns II or III with pattern I were similar in cardia (II vs. I: OR = 1.71; III vs. I: 0.91) and non-cardia cancers (II vs. I: OR = 1.64; III vs. I: 0.85) (Table 5).
Table 5. Association between dietary patterns and gastric cancer, by cancer location and histological type
Considering the Laurén's histological type, when comparing pattern II with pattern I the association was stronger for the intestinal type (OR = 1.87, 95%CI: 1.30–2.67) than for the diffuse type (OR = 1.32, 95% CI: 0.83–2.08). When comparing pattern III with pattern I there was no association for the intestinal type tumors (OR = 0.99, 95%CI: 0.59–1.66), but more than a 50% decrease in risk in those of the diffuse type (OR = 0.43, 95% CI: 0.22–0.87) (Table 5).
A dietary pattern characterized not only by low intake of fruit, salads and vegetables, but also low total energy intake and low intakes of meat, fish and dairy products was associated with an increased gastric cancer risk when compared to the pattern mainly characterized by high consumption fruit, salads and vegetables. The pattern with the highest energy intake and high consumptions of not only most food groups, namely red meat or cereals and potatoes, but also fruit and vegetables, despite having the lowest intake of vegetable soup was not associated with an overall increased risk. Findings were similar for cardia and non-cardia cancers, but the latter pattern was associated with a decreased risk of diffuse type gastric cancer.
Our study is one of the few investigations20–26 addressing the association between dietary patterns and gastric cancer, and the first conducting a stratified analysis according to the histological type and taking into account the potential effect of H. pylori infection. However, some limitations have also to be discussed, namely those related to the methodological shortcomings of case-control studies and specifically the design and data analysis options of the present investigation. Cancer registry data could not be used to identify cases because of the local delay in cancer notification. Therefore we opted for approaching every case admitted to the Surgery wards of the 2 major public hospitals caring for cancer patients in the North of Portugal, both in Porto. The sex- and age-distribution of gastric cancer cases identified in this study is similar to the observed in cases from Porto.27 However, the patient's social class may be different in the hospitals involved in our study and in institutions that were not included, especially private clinics, which may be responsible for an overestimation of the association between education and gastric cancer in this study, but the adjusted estimates for the association between diet and cancer are not expected to be biased.
An accurate classification of gastric cancer subtypes is essential to ensure the validity of the findings in a study assessing the risk of gastric cancer according to cancer location and histological type. Only the cases undergoing gastrectomy, therefore allowing an adequate sampling of the tumor, were considered for histological type analysis, and experienced pathologists were involved in the assessment of histological type (including the standardized review of pathology reports and reassessment of slides whenever necessary), contributing for an accurate classification of the histological type, and overcoming the most frequent limitations of previous research on this topic.13 We used the best available information to classify tumors' topography, but those described as “cardia cancers” may include a mixture of neoplasms arising from the lower oesophagus as well as the gastric cardia, to a higher or lesser extent depending on the frequency of oesophageal and gastric adenocarcinomas on the populations under study. This potential for misclassification brings important challenges to the study of cardia cancer aetiology, as the origin of these cancers can hardly be determined by examining them grossly or microscopically.28 However, recent studies28, 29 suggest the coexistence of 2 aetiologically distinct types of cardia cancer, one associated with H. pylori-induced atrophic gastritis, similar to non-cardia cancer, and the other associated with nonatrophic gastric mucosa and resembling oesophageal adenocarcinoma. A practical implication of these findings is that the state of the non-neoplastic gastric mucosa may provide a key to determining the 2 subtypes of gastric cardia cancer.28 In our study, most cardia cancer cases occurred in patients with atrophic gastritis in the non-neoplastic mucosa (data not shown) that is consistent with them being of aetiology similar to non-cardia cancer arising from H. pylori-induced atrophic gastritis, and different from lower esophageal adenocarcinomas associated with gastroesophageal reflux.29
In this study, the interviewers were carefully trained to approach cases and controls in the same way, the questionnaires were precoded, and interviewers evaluated both cases and controls, which contributes to overcome potential information bias resulting from the impossibility of blinding interviewers for the case or control status of participants.
One of the major concerns in case-control studies is differential recall bias due to the knowledge of the disease status. Furthermore, cases might have changed diet due to their disease, and the recall of past dietary intake can be influenced by current habits.30 Because illness duration may modify the accuracy of dietary reporting, our study was limited to incident cases, dietary data referred to the year preceding the diagnosis or the change in dietary habits, when applicable, and cases declaring to have changed dietary habits more than 1 year before the interview were excluded from analysis to improve the recall of dietary intake. Although these options contribute to increase the validity of our results, dietary changes may have occurred without being perceived by the patients30 and information bias may persist to some extent. However, data were available regarding weight variation in cancer patients (weight at the time of interview minus usual weight before diagnosis), which may reflect the potential for dietary changes not acknowledged by the patients. We conducted a stratified analysis according to weight variation (median variation was used as cut-point), and the adjusted OR estimates were not significantly or meaningfully different when considering the cases with more pronounced weight losses (pattern II vs. pattern I: OR = 1.91, 95% CI: 1.39–2.63; pattern III vs. pattern I: OR = 0.94, 95% CI: 0.58–1.50) or those with smaller weight changes (pattern II vs. pattern I: OR = 1.48, 95% CI: 1.06–2.05; pattern III vs. pattern I: OR = 0.71, 95% CI: 0.43–1.15). These results support the validity of our conclusions regarding the association between pattern II and an increased risk of gastric cancer.
The limitations of the case-control design to assess the relation between H. pylori infection and gastric cancer, especially in high risk settings, are also likely to account for the similar prevalence of infection in cases and in controls observed in our study. When H. pylori infection is evaluated retrospectively, false-negative results in serological assessment of infection status (classified as ever in life) are more frequent among cases, especially those with advanced gastric cancer, and in settings with a high proportion of subjects with chronic atrophic gastritis31 and intestinal metaplasia,32 due to spontaneous disappearance of the infection and consequent underestimation of its life-prevalence.31 Moreover, the prevalence of infection gradually increases with age in settings where the overall prevalence is low, but in high-risk populations, namely the Portuguese, it approaches 80–90% at young adulthood,12 which contributes to lower OR estimates than the observed in lower risk settings. This phenomenon finds a parallel in the weaker association between infection and gastric cancer in the age-groups, due to the increasing prevalence of infection with age.7, 33
Most previous studies assessing the association between dietary patterns and the risk of gastric cancer relied either on factor analysis (or principal components analysis) or cluster analysis. We opted for principal components analysis to derive new noncorrelated variables, aiming to reduce the dimensionality of the data by obtaining fewer new components able to explain a large proportion of the variation in the dietary habits across participants. Although several studies have concluded their analysis with the constructions of the factors,23, 25 we considered that the variance explained by our components was lower than desirable (“rule of thumb”: lower than 90%34) and further conducted a cluster analysis. Cluster analysis is a useful tool for summary and descriptive purposes, but is data dependent. As the selection of the number of clusters is largely subjective, we used the Calinski and Harabasz pseudo-F stopping rule index and the Duda and Hart Je(2)/Je(1) index jointly to select the number of clusters aiming to reduce the subjectivity of this approach. Cluster analysis also depends on the choice of clustering method; consequently, different techniques could yield different solutions with the same data. We used the Ward's hierarchical clustering method, as it is known to perform well when the data may aggregate in natural groups with the same size,19 as expected in the present analysis. To examine the reproducibility of our findings we randomly divided the sample into 2 groups, applied the same method to both groups and compared the results between them and with our final model. The results showed very similar dietary patterns (data not shown).
We computed OR estimates adjusted for the potential confounding effect of age, gender, education and total energy intake. Tobacco smoking is also a potential confounder as it is an important behavioral risk factor for gastric cancer6 and is known to be associated with dietary habits,35 but the models additionally including smoking status (categorical variable: never, former and current smokers) yielded approximately the same results (pattern II vs. pattern I: OR = 1.67, 95% CI: 1.31–2.14; pattern III vs. pattern I: OR = 0.81, 95% CI: 0.57–1.14).
The dietary patterns approach is population-dependent; therefore, the external validity of these findings, the comparison across studies, especially between populations with different dietary habits is difficult. Moreover, the “labeling” of the identified patterns and the interpretation of their association with the occurrence of gastric cancer is also somewhat arbitrary. For example, Campbell et al.21 defined 2 patterns labeled as “prudent” (positively loaded on several vegetables, fruits and fish) and “western” (strongly correlated with soft drinks, French fries, white bread, hamburger, eggs, bacon, doughnuts, and hot dogs, and modestly negatively correlated with broccoli, spinach and tofu). Bahmanyar and Ye25 identified 3 major patterns labeled as “healthy diet” (high in vegetables, tomato, fruits, fish and poultry), “Western diet” (high in processed meat, red meat, sweets, high-fat dairy and high-fat gravy), and “alcohol drinker” (high in intakes of beer, liquor and French fries). Masaki et al.20 identified 4 major patterns: “vegetable and fruit” (loaded greatly on cabbage, lettuce, green leafy vegetables, carrots, oranges and other fruits), “Western breakfast” (bread, butter, cheese, ham sausage and coffee were consumed more often in contrast to rice, seaweeds, bean curd and pickled vegetables), “meat” (positively loaded on pork, beef, and chicken and negatively on tomatoes and other fruits), and “rice/snack” (rice, miso soup, cookies and orange juice loaded heavily in contrast to negative levels of vegetables).
Taking into account only the few dietary factors for which the association with gastric cancer is established (essentially an increased risk for low intakes of fruit and vegetables or high intake of salt or salted/salty foods)5 patterns I and III identified in our study may be labeled as “fruit and vegetables” or “healthy,” as both include high intakes of food items pertaining to the groups of fruits, vegetables and salads.
A Portuguese report of food consumption in Porto shows that the main supplier of sodium, intrinsic to foods, is bread. However, after a joint estimate of the intrinsic salt and salt added to home-made cooking, vegetable soup is the major supplier of the total ingested sodium, followed by pasta, rice and potatoes and bread,36 which may be explained by the fact that Portuguese vegetable soup is a salty beverage consumed in larges quantities. This could explain a somewhat lower risk of gastric cancer in subjects included in pattern III, as it is characterized by a lower consumption of vegetable soup, even if pattern I has the lowest consumption of cereals and potatoes. Another possible explanation for the differences observed between patterns I and III is that dietary habits are associated with healthy behaviors and better socio-economic status.37, 38 A recent Portuguese study addressed the relation between educational attainment and food choices, showing that higher education is associated with higher intake of vegetable soup.39 Therefore, we cannot exclude the possibility that not only dietary factors defined as dietary patterns, but also their related demographic and lifestyle factors, may affect the gastric cancer risk. In our study the adjusted models included age, gender, education and total energy intake, but residual confounding cannot be excluded, even taking into account that H. pylori infection status was considered in the analysis.
Given the complex nature of dietary patterns, it is difficult to translate the evidence generated in the present investigation into direct dietary advice to the public, which is out of the scope of this study. Our results confirm previous findings in the protective effect of high fruit and vegetables intake, and allowed the disclosure of a differential association between dietary pattern rich in fruit and vegetables according to histological type.