Introduction

Type 2 diabetes is an increasingly common and costly chronic disease that predisposes affected individuals to substantial risk in terms of morbidity and mortality [1, 2]. A knowledge of the risk factors and protective factors associated with type 2 diabetes is essential for the development of prevention strategies. Diet is thought to have an important influence on the development of diabetes. The traditional approach used to investigate diet–disease associations focuses on single dietary components, such as single nutrients or foods. For example, it has been suggested that different types of fatty acids [3], cereal fibre [4], and magnesium [5] are associated with the risk of diabetes. However, diet is a complex entity, and therefore it is difficult to investigate the effects of individual dietary components separately. Thus, within the field of nutritional epidemiology, the use of dietary patterns has been proposed. A dietary pattern is a comprehensive variable of several foods or food groups that takes into consideration the interactions and cumulative effects of dietary components on disease risk. To date, patterns have been obtained using a hypothesis-oriented approach, such as diet indices, or an exploratory approach, such as cluster and factor analysis [6]. Factor analysis, the method most commonly used to identify patterns associated with diabetes [7–10], derives the typical dietary patterns of a specific study population, regardless of their relevance to health outcomes. Thus, the use of factor analysis for the identification of a pattern associated with a disease of interest is of limited value. Reduced rank regression (RRR) has recently been proposed as an approach that can be used to provide the link between overall diet and disease [11, 12]. This method combines the dietary information of the particular study population with prior scientific knowledge of the pathway from diet to disease.

The aim of this study was, first, to identify a dietary pattern cross-sectionally associated with plasma concentrations of diabetes-related biomarkers using RRR and, secondly, to investigate whether the pattern obtained is prospectively associated with type 2 diabetes.

Subjects and methods

Study population

A nested case–control study on the risk of developing type 2 diabetes was designed within the European Prospective Investigation into Cancer and Nutrition (EPIC)—Potsdam Study. The EPIC—Potsdam Study is a prospective cohort study of 27,548 individuals aged 35–65 years recruited from the general population. Baseline examinations, including self-administered questionnaires on dietary and lifestyle habits, a computer-guided interview on further lifestyle habits and medical history, anthropometric measurements, and blood sampling were carried out between August 1994 and September 1998. Follow-up questionnaires are mailed every 2–3 years to obtain updated measurements of baseline characteristics and to evaluate newly developed diseases, including type 2 diabetes and current medication. Case subjects were defined as participants without type 2 diabetes at baseline who developed type 2 diabetes during the first follow-up period. We identified 399 potential cases of incident type 2 diabetes from self-reports on incident diseases, current medication and current dietary treatment for diabetes. These cases were verified by the patients’ medical records. Up to November 2001, 201 cases of incident diabetes were medically verified. The subsequent biochemical analysis of the diabetes-associated antibodies GAD65 and IA-2 revealed that nine of these individuals had type 1 diabetes, while the remaining 192 individuals had type 2 diabetes. Each of the case subjects was matched with two control subjects in terms of sex and age (±1 year). Subjects with missing values for exposure variables or one of the confounding variables used in the subsequent statistical analyses were excluded. Consequently, the final study population consisted of 192 diabetic subjects and 382 control subjects.

The EPIC—Potsdam Study was approved by the Ethics Committee of the Landesärztekammer Brandenburg, Germany. Each participant gave written consent before enrolment into the study.

Dietary assessment

All participants were asked to complete a semi-quantitative food-frequency questionnaire. This questionnaire assessed the average frequency of intake and the portion size of 148 foods consumed during the 12 months prior to examination. Frequency of intake was measured using ten categories, ranging from ‘never’ to ‘five times per day or more’. Portion sizes were estimated using photographs of standard portion sizes. Information on frequency of intake and portion size was used to calculate the amount of each food item in grams consumed on average per day. The validity and reproducibility of the food-frequency questionnaire have been described previously [13, 14]. The 148 single food items of the questionnaire were aggregated into 48 food groups. The grouping criteria were based on the 44 food groups previously described in detail [15], and on additional categories for cooked vegetables (legumes and cooked vegetables), cheese (low- and high-fat), other dairy products (low- and high-fat) and soft drinks (low- and high-energy).

Covariate assessment

Body height and weight, and waist and hip circumferences were measured by trained and quality-monitored personnel, with participants wearing light underwear [16]. BMI was calculated as body weight in kilograms divided by body height in metres squared, and WHR as waist circumference divided by hip circumference.

Information on leisure-time activity, smoking status, educational level, and intake of lipid-lowering and anti-inflammatory drugs was obtained from a self-administered questionnaire and a personal computer-guided interview. Leisure-time activity was calculated as the average of hours of physical activity per week during the summer and winter. Smoking status was defined as current smoker, ex-smoker for <5 years, and ex-smoker for ≥5 years or non-smoker. Educational level was defined as less than high school education, high school education, or university degree. Users of lipid-lowering or anti-inflammatory drugs were defined as subjects reporting a regular intake of lipid-lowering drugs, or use of antiphlogistics, analgesics, antirheumatics or corticoids, respectively.

Selection of response variables

Type 2 diabetes is a complex condition with various antecedents. For our analysis we chose the four plasma parameters HbA1c, HDL cholesterol, C-reactive protein (CRP), and adiponectin as response variables as they are affected by diet and are related to the pathophysiology of diabetes.

HbA1c reflects elevated blood glucose concentrations [17] and is commonly used as an indicator of long-term glycaemic control [18]. Since impairment of glucose homeostasis increases the risk of type 2 diabetes development [19–21], we used HbA1c as one of the four response variables in the analysis. HDL cholesterol was selected because studies report that insulin resistance and type 2 diabetes are commonly accompanied by decreased HDL cholesterol concentrations, with an increased flux of NEFA to the liver considered as an underlying metabolic effect [23]. Studies indicate that diabetes is associated with a chronic subclinical inflammatory process [24] that is characterised by changes in concentrations of acute-phase proteins [25]. Of these, an elevated CRP concentration is currently the best marker of the future risk of type 2 diabetes [26–28]. The adipocyte-derived peptide adiponectin has been reported to have insulin-sensitising properties and anti-inflammatory effects [29–31]. Although the mechanisms linking adiponectin to diabetes risk remain speculative, the plasma adiponectin level seems to be an independent predictor of type 2 diabetes [32].

Evidence suggests that the four diabetes-related biomarkers we chose are affected by dietary components, especially by alcohol consumption [33–36], fat composition [37, 38], and certain types of carbohydrates [39, 40].

Laboratory procedures

At the baseline examination of the EPIC—Potsdam Study, peripheral venous citrate blood samples were taken and centrifuged at 1,000×g for 10 min at 4°C. Plasma was removed and stored in aliquots at −80°C until assays of biochemical markers were performed. The diabetes-associated antibodies GAD65 and IA-2 were measured by RIA (Medipan Diagnostics, Selchow, Germany). HbA1c was analysed by enzyme immunoassay (DAKO Diagnostics, Hamburg, Germany; intra-assay CV 2.0–3.0%, interassay CV 1.9%). HDL cholesterol was measured by enzyme calorimetric assay (ABX Diagnostics, Göppingen, Germany; interassay CV 2.1%). CRP was determined by high-sensitivity ELISA (Immun Diagnostik, Bensheim, Germany; intra-assay CV 4.9–6.2%, interassay CV 13.2%). The limit of detection for CRP was 0.124 μg/ml. For the subsequent statistical analysis, CRP values below the limit of quantification were set at 0.7 times the respective limit of quantification [41]. Adiponectin was measured by RIA (Linco Research, St Charles, MO, USA; intra-assay CV 0.1–6.2%, interassay CV 5.0%). All assay procedures were performed as described by the manufacturer.

Statistical analysis

Analyses were performed by using SAS software, Version 8.02 (SAS Institute, Cary, NC, USA). To test for significant differences in baseline characteristics between type 2 subjects and control subjects the non-parametric Kruskal–Wallis test was used for continuous variables and the Mantel–Haenszel test was used for categorical variables.

The statistical method RRR, also known as the maximum redundancy analysis, was employed using the PLS procedure in SAS to derive dietary pattern scores. The use of this method in dietary pattern analyses has recently been described in detail and has been compared with other methods elsewhere [11]. Briefly, the method allows the calculation of scores similar to those extracted by factor analysis. Whereas factor analysis determines scores, i.e. linear combinations of predictor variables, by maximising the explained variation of all predictor variables (e.g. food groups), RRR determines linear combinations of predictor variables by accounting for as much of the variation in response variables (e.g. disease-related biomarkers) as possible. In the special case of only one response variable, RRR is identical to multiple linear regression.

We used intake data on 48 food groups as predictors, and logarithmically transformed data for HbA1c, HDL cholesterol, CRP, and adiponectin levels as responses to derive a dietary pattern score. To reduce confounding by medication, HDL cholesterol values were adjusted for intake of lipid-lowering drugs, whereas values of CRP and adiponectin were adjusted for intake of anti-inflammatory drugs. As the number of extracted scores cannot be higher than the number of selected responses, four scores were obtained. Because the first dietary pattern score explained much more variation than the subsequent ones, this was the only one considered in the subsequent analysis. Factor loadings represent the correlations of each food group with the dietary pattern score. To focus on food groups that significantly contributed to the dietary pattern score we only considered food groups with absolute factor loadings >0.2.

Crude and adjusted (for age, sex, leisure-time activity, smoking status, educational level and total energy intake, with or without adjustment for BMI and WHR) odds ratios (ORs) and 95% CIs for the risk of developing type 2 diabetes according to quintiles of the dietary pattern score were determined by conditional logistic regression analysis.

To examine the robustness of results, several sensitivity analyses were performed using different sets of response variables. First, we selected the four possible combinations of three response variables. Second, we added IL-6 and TNF-α separately to the four originally chosen responses. Finally, we chose all six responses together. The first dietary pattern score was derived in each case. Multivariate-adjusted ORs and 95% CIs for the incidence of diabetes across quintiles of the respective scores were determined again by logistic regression analysis. A p value of less than 0.05 was considered statistically significant for all analyses.

Results

The baseline characteristics of the study population are summarised in Table 1. Compared with control subjects, the diabetic subjects had higher mean values for BMI and WHR and were less likely to have a university degree. Leisure-time activity, total energy intake and percentage of current smokers did not differ significantly between the two groups. Intake of lipid-lowering drugs was more frequent among diabetic study participants than control subjects, whereas use of anti-inflammatory drugs was similar. Plasma concentrations of biomarkers differed significantly between diabetic subjects and control subjects, with the former having higher values for HbA1c and CRP, and lower values for HDL cholesterol and adiponectin.

Table 1 Baseline characteristics of the study population
Full size table

Dietary pattern scores were derived using intake data on 48 food groups as predictors and logarithmically transformed medication-adjusted concentrations of the four biomarkers as responses in RRR. Four scores were obtained. The first score explained 10.2% of the variation in adiponectin (compared with 15.5% maximally explainable by all 48 food groups), 9.4% of HDL cholesterol variation (14.2%), 9.1% of HbA1c variation (14.8%), 1.1% of CRP variation (13.4%), and 7.4% (14.5%) of the total variation in the four selected biomarkers. The second, third and fourth scores accounted for only 3.6%, 1.9% and 1.5% of the total variation in biomarker levels, respectively. We therefore only included the first score in further analyses.

The association between each food group and the dietary pattern score is indicated in Table 2. Food groups were sorted according to decreasing absolute values of factor loadings. The factor loadings represent the correlations of the single food groups with the score. The mean intakes of the 48 food groups across quintiles of the score are also provided. Red meat, beer, legumes, poultry, high-caloric soft drinks, processed meat, and bread (excluding wholegrain bread) were inversely correlated with the score (factor loadings <−0.2) and were characterised by a steady decrease in intake across quintiles of the dietary pattern score (p<0.0001 for trend). In contrast, fresh fruit was directly associated with the score (factor loading >0.2) and showed an increased intake according to quintiles of the score (p<0.0001 for trend). All other food groups made a smaller contribution to the dietary pattern score (absolute factor loadings <0.2).

Table 2 Factor loadings and intake of food groups in the study population (n=574) according to quintiles of the first identified dietary pattern score
Full size table

The mean plasma biomarker concentrations according to quintiles of dietary pattern score are presented in Table 3. HbA1c values consistently decreased across quintiles of the score (p<0.0001 for trend), whereas concentrations of HDL cholesterol and adiponectin increased (p<0.0001 for trend). No significant trend was evident for CRP concentration across quintiles of the score (p=0.18 for trend).

Table 3 Plasma concentration of response variables in the study population (n=574) according to quintiles of first dietary pattern score
Full size table

Table 4 shows the risk of developing type 2 diabetes for each quintile of the dietary pattern score. In the crude model (model 1), the risk was consistently and significantly reduced across quintiles of the score. Subjects in the highest quintile had an 80% lower diabetes risk than those in the lowest quintile (OR=0.20, 95% CI 0.10–0.40, p<0.0001 for trend). After adjusting for known risk factors (except for BMI and WHR) a further slight decrease in diabetes risk was observed (OR=0.18, 95% CI 0.09–0.37, p<0.0001 for trend). Additional adjustment for BMI and WHR, which are strong predictors of type 2 diabetes, reduced the magnitude of the decrease in risk estimates across quintiles of the score. However, subjects in the highest quintile still had a >70% lower risk of developing diabetes than subjects in the lowest quintile (OR=0.27, 95% CI 0.13–0.64, p=0.0006 for trend).

Table 4 Risk of type 2 diabetes in the study population (n=574) according to quintiles of first dietary pattern score
Full size table

To examine the dependence of the results on the choice of biomarkers we derived the first dietary pattern scores for varying sets of biomarkers (Table 5). The risk of type 2 diabetes significantly decreased with increasing quintiles for each of the dietary pattern scores.

Table 5 Risk of type 2 diabetes in the study population (n=574) according to quintiles of first dietary pattern scores derived from different combinations of response variables
Full size table

Discussion

We have successfully identified a dietary pattern that is associated with plasma concentrations of four diabetes-related biomarkers and with the incidence of type 2 diabetes. A high pattern score reflected a diet that included a high intake of fresh fruit and low intakes of high-caloric soft drinks, beer, red meat, processed meat, poultry, legumes and bread (excluding wholegrain bread). Together, these eight food groups explained about 70% of the variation in dietary pattern score (data not shown). Furthermore, subjects with a higher dietary pattern score were characterised by higher plasma concentrations of HDL cholesterol and adiponectin, and by lower plasma concentrations of HbA1c and CRP.

The results of our study show parallels with previous studies. In a population-based cross-sectional study, a so-called “healthy balanced” dietary pattern was identified. This pattern was mainly characterised by a frequent fruit intake, and was positively associated with HDL cholesterol concentration and inversely correlated with type 2 diabetes [7]. In a large prospective study of men, a ‘prudent pattern’ with a high intake of fruit was identified that was associated with a moderate reduction in the risk of subsequent diabetes [9]. Within the same study population the authors found that red meat and processed meat were the two main components of a diet named “western pattern”. This pattern contributed to an increased incidence of type 2 diabetes [9]. Alone, a high intake of processed meat, adjusted for known risk factors or other dietary components, was associated with an increased risk of diabetes in men and women, respectively [42, 43]. Other studies have reported that a dietary pattern including a high intake of refined bread is correlated with a detrimental biomarker profile [44] and an increased risk of diabetes [8, 9].

Studies investigating the relationship between alcohol consumption or type of alcoholic beverage and diabetes risk have shown inconsistent results. In our study, subjects with a lower dietary pattern score consumed larger amounts of beer and showed an increased risk for type 2 diabetes. Furthermore, these subjects were characterised by a higher consumption of high-caloric soft drinks. Both types of beverage may have several diabetogenic effects, including contributing to an excess energy intake, to overweight and disturbances in carbohydrate metabolism [45–48].

In one prospective study of men, poultry and legumes were included in the “prudent pattern” that was inversely associated with diabetes risk [9]. Furthermore, it was suggested that legumes are good sources of dietary fibre, which has a positive effect on concentrations of cholesterol, glucose and insulin in the blood [49], and contributes to improved glycaemic control [50]. Nevertheless, we found no direct evidence in the literature to suggest that either poultry or legumes are associated with the risk of type 2 diabetes. In our study we observed that the intake of poultry was directly correlated with a dietary pattern associated with an increased risk of diabetes, as was the intake of processed meat and red meat. Moreover, we found that a high consumption of legumes was linked to a dietary pattern score associated with an increased risk of type 2 diabetes. This finding may be explained by the fact that most of the intake of legumes was attributable to a stew of lentils, peas or beans, which is preferably accompanied by bacon, sausages, beef or pork in this region.

Only a few epidemiological studies have identified dietary patterns associated with the risk of type 2 diabetes. Compared with the single dietary component approach, the patterning approach takes account of the complexity of the dietary intake and the intercorrelations and cumulative effects of various nutrients or foods. Thus, patterns are appropriate measures of overall diet. Furthermore, we are not aware of any other studies that have derived patterns associated with diabetes risk using biomarker information. Analysis by RRR allowed us to use data on the diabetes-related biomarkers CRP, adiponectin, HDL cholesterol and HbA1c to derive a pattern score associated with diabetes risk. Compared with other data reduction techniques (e.g. principal component or factor analysis), RRR has the advantage that it may extract dietary pattern scores by maximising the explained variation in the concentration of biomarkers or other intermediate variables in the pathway that links diet to the disease of interest. Thus, RRR may be a useful tool for the identification of diet–disease associations. However, the method is limited by the fact that it requires prior knowledge of the disease-related variables chosen as responses.

A further advantage associated with our study is the prospective study design. The identification of the study population and assessment of diet before diabetes was diagnosed reduces the possibility of selection and recall bias. Our results should not be severely biased by misclassification of diabetes, since all potential cases were verified by medical records. The remaining misclassification based on non-identified incident or prevalent cases would serve to attenuate the observed risk estimates rather than overestimate the risks. Moreover, the high follow-up rate of 96% minimised the bias due to loss to follow-up.

Nevertheless, our study has several limitations. First, we assessed diet using a food-frequency questionnaire. Measurement errors may be introduced by the under- or over-reporting of the amounts of food groups usually eaten per day. Second, a crucial issue in selecting biomarkers as response variables for RRR is the time of blood sampling. Ideally, all biomarkers should be biomarkers of exposure or disease, or they should refer to the same stage of the pathway from exposure to disease [51]. In our study, the four responses CRP, adiponectin, HDL cholesterol and HbA1c probably reflect different stages in the pathogenesis of type 2 diabetes. HbA1c in particular may refer to a later stage of the pathogenesis of diabetes than the other three biomarkers. However, elevated HbA1c values that reflect higher average glucose concentrations and that are below the threshold for diabetes may refer to a preceding stage of type 2 diabetes. Additionally, the sensitivity analysis showed that a dietary pattern score derived using a biomarker set without HbA1c is still significantly associated with a reduced risk of diabetes (Table 5). Third, although the variations in the selected biomarkers elucidated by the dietary pattern score were small, these translated into relatively large variations when comparing the maximally explainable biomarker variation by all food groups. Fourth, a cohort approach could be more attractive than the nested case–control approach used in this study, because this design produces more realistic results. A cohort study would probably produce smaller risk estimates across quintiles than those observed in this study. However, a study based on a nested case–control design with available biomarker values for all cases and matched controls seemed the most appealing approach for our aims. Fifth, patterns may be difficult to reproduce in other studies, especially in study populations with different dietary habits.

In summary, we have identified a dietary pattern score that is directly associated with the intake of fresh fruit and inversely associated with the intake of high-caloric soft drinks, beer, red meat, poultry, processed meat, legumes and bread (excluding wholegrain bread). A higher pattern score was cross-sectionally related to higher concentrations of HDL cholesterol and adiponectin, and lower concentrations of HbA1c and CRP. It was also prospectively associated with a substantially lower risk of diabetes. The observed association between the pattern and diabetes risk was independent of BMI, WHR, smoking status, leisure-time activity, energy intake, educational level, age and sex. The results suggest that the identified dietary pattern, which was linked to a specific biomarker profile, may be a predictor of the subsequent risk of type 2 diabetes.