Modifying effect of metabotype on diet–diabetes associations

Analyses were performed on data from the population-based KORA FF4 (2013/2014) study, the second follow-up of the KORA S4 health survey conducted in the region of Augsburg in Southern Germany between 1999 and 2001 [26]. In brief, of the 4261 participants included in S4, 2279 individuals also participated in the 14-year follow-up FF4 study. Detailed information on the participation response has been given elsewhere [27]. All individuals answered self-administered questionnaires, and participated in a standardized physical examination as well as in a computer-assisted face-to-face interview conducted by trained investigators at the study center. A detailed description has been provided previously [5].

All participants gave their written informed consent, and the study was approved by the Ethics Committee of the Bavarian Chamber of Physicians and conducted in accordance with the Declaration of Helsinki.

Prevalent diabetes was defined by either current intake of antidiabetic medication or a self-reported diagnosis, both validated with the respective treating physician. All participants without previously known diabetes took part in a standard oral glucose tolerance test (OGTT) and their glucose tolerance status was classified according to the 2003 American Diabetes Association (ADA) diagnostic criteria [28]. Further details have been outlined elsewhere [29]. An OGTT value of ≥ 7.0 mmol/L fasting or ≥ 11.1 mmol/L 2-h glucose was defined as undetected diabetes mellitus (UDM), also called screen-detected diabetes. Participants with isolated impaired fasting glucose (IFG 5.6–6.9 mmol/L fasting glucose), isolated impaired glucose tolerance (IGT 7.8–11.0 mmol/L 2-h glucose) or combined IFG/IGT were classified as prediabetic. Individuals with fasting glucose levels < 5.6 mmol/L and 2-h glucose levels < 7.8 mmol/L were classified as normal glucose tolerant (NGT).

Dietary intake was assessed in 1602 KORA FF4 participants with up to three 24-h food lists [30] and a food frequency questionnaire. Combining this information, the usual dietary intake was estimated in an advanced blended two-step approach, which follows the idea of the National Cancer Institute (NCI) method and the Multiple Source Method (MSM) [31, 32] to separate the calculation of consumption amount and consumption probability. The consumption probability and the consumption amount on consumption days were estimated separately with models both including the same covariates to link the two parts. Then, the usual dietary intake of all food items was calculated for each participant by multiplying the consumption probability of a certain food item by the usual consumption amount on a consumption day. The food groups were categorized according to the European Prospective Investigation into Cancer and Nutrition (EPIC)-Soft classification system [33] and nutrients were derived using the National Nutrient Database (Bundeslebensmittelschlüssel BLS 3.02). For the analysis, we selected the 17 following food groups and nutrients in g/d associated with T2DM in the literature [9, 10, 34, 35]: fruits, vegetables, potatoes, total meat, red meat (beef and pork), poultry, processed meat, eggs, total dairy, milk, yogurt, cheese, coffee, fruit and vegetable juice, sugar-sweetened beverages (SSB), alcohol and fiber.

The selection of covariates was based on theoretical considerations and the existing literature on diet and diabetes [9]. These included age (years), sex (reference = male), energy intake (kcal/day), waist circumference (cm), family history of diabetes [yes, no (= reference), do not know], physical activity [active in summer and in winter and active for ≥ 1 h per week in at least one season, inactive (= reference)], smoking status [never (= reference), former, current], hypertension [≥ 140/90 mmHg or antihypertensive medication given that the participants were aware of having hypertension; yes, no (= reference)] and education [< 10 years (= reference), 10 to < 13 years, ≥ 13 years, in accordance with the German education system]. Waist circumference and blood pressure were measured at the study center under standardized conditions by trained staff. All other covariates were assessed during a standardized personalized computer-assisted interview or via a self-administered questionnaire.

We performed all statistical analyses using the statistical software package RStudio version 1.0.136 that uses R version 3.2.2 (R Development Core Team, 2010, http://​www.​r-project.​org). p values of < 0.05 were considered statistically significant.

In the KORA FF4 study population, a low intake of fruits and a high intake of total meat, processed meat and SSB were significantly associated with UDM/prevalent T2DM. Taking into account metabolic differences between individuals, each of these associations remained significant only in one of both metabotype subgroups. In the combined clusters 1 and 2, the intake of total meat and processed meat showed a positive association with UDM/prevalent T2DM. In cluster 3, the intake of fruits was negative and the intake of SSB was positively associated with UDM/prevalent T2DM. Despite these differences in significant associations between metabotype subgroups, only the interaction effect between intake of SSB and metabotype was significant. In addition, the association between the intake of eggs and UDM/prevalent T2DM was also significantly different between metabotype subgroups, however, egg intake was not significantly associated with UDM/prevalent T2DM in the individual metabotype subgroups potentially caused by loss of statistical power due to stratification. Thus, further significant associations or significant differences in diet–diabetes associations between metabotypes as for example for intake of coffee (p value for interaction = 0.09) and processed meat (p value for interaction = 0.11) might be detectable in larger cohorts.

In general, the associations of intakes of fruits, total meat, processed meat and SSB with UDM/prevalent T2DM that we found in the total KORA FF4 study population are in line with previous meta-analyses and review articles on the respective food groups [39‐48]. However, results of a number of other studies were inconsistent [9, 10], and there are studies that have shown either no or only weak associations between these food groups and diabetes [49‐57]. All other selected dietary intake variables were not associated with UDM/prevalent T2DM in our study, in contrast to some of the previous studies [9, 10, 34, 35]. Our study was the first to investigate associations between diet and diabetes stratifying by metabotype, i.e., considering metabolic differences between individuals. As we identified intakes of fruits, total meat, processed meat and SSB to be significantly associated with diabetes in only one of both metabotypes, metabolic differences may partially explain conflicting results in diet–diabetes associations observed in previous studies. This holds mainly for the association between intake of SSB and diabetes, which was shown to be significantly different between both metabotype subgroups. Despite previously described age-related metabolic differences [58], the results of our sensitivity analysis restricted to older participants ≥ 60 years remained relatively stable. Consequently, these metabotypes may be relevant for adult populations with a large age range and not only for adult populations with specific age categories.

In the previous literature, there are numerous studies on metabotyping [16, 22, 23]. Some of these studies defined metabotypes including fasting plasma values, while a few studies classified metabotypes based on plasma parameter responsiveness to dietary interventions [17, 22, 59‐62]. Fewer studies tested metabotype subgroups for a differential responsiveness to dietary intervention in a disease-specific manner [63‐65]. For example, O’Sullivan et al. [63] identified a vitamin D-responsive metabotype subgroup concerning markers of the metabolic syndrome, Moazzami et al. [64] found subgroups of individuals with different insulin response after an intervention with breads, and Vázquez-Fresno [65] detected a responsive metabotype subgroup of cardiovascular risk patients to red wine polyphenols. In addition, O’Donovan et al. were the only ones developing dietary recommendations based on metabotype subgroups which was done using a decision tree approach [66, 67]. Thus to date, metabotypes have been rarely used for the development and establishment of targeted dietary recommendations for disease prevention. Further research should identify discrete differences between metabotype subgroups in the context of diet–disease relationships. By assigning individuals to metabotype subgroups, targeted dietary recommendations in disease prevention may be implemented in whole populations.

Concerning the metabotypes identified in KORA FF4, diabetes prevention may be especially relevant for cluster 3, defining an unfavorable metabotype concerning metabolic characteristics. In detail, this cluster showed the highest median concentrations of glucose and glycated hemoglobin, which are used in the diagnosis of prediabetes and prevalent T2DM [68]. Consequently, there were high numbers of individuals with prediabetes, UDM and prevalent T2DM in this cluster. However, clusters 1 and 2 defining a rather beneficial metabotype could also benefit from targeted prevention due to the high prevalence of prediabetes, which is a strong risk factor for the development of T2DM [69]. Other known risk factors for diabetes such as age, obesity, physical inactivity, family history of diabetes and hypertension were most frequent in cluster 3 [1, 68, 70]. In addition, low education as seen in cluster 3 was linked to poor health [71]. The lowest percentage of current smokers and simultaneous highest percentage of ex-smokers in cluster 3 indicated smoking cessation due to the high diabetes prevalence in this cluster, as smoking is as well a strong risk factor for diabetes [1, 70]. When assessing dietary intake in our metabotype subgroups, individuals in cluster 3 showed a higher intake of total energy, total meat, red meat, processed meat and SSB than individuals in clusters 1 and 2. Simultaneously, individuals in cluster 3 consumed lower amounts of vegetables, total dairy, milk, yogurt and fruit and vegetable juice compared to individuals in clusters 1 and 2. This dietary pattern was shown to be associated with increased risk for diabetes [9, 10]. As targeted dietary advice may be more effective than general recommendations [11, 14, 16, 24, 25], the development of strategies for change in dietary behavior on the metabotype subgroup level, especially for the ‘high-risk’ cluster 3, could improve the prevention of diabetes.

One of the strengths of the study is the fact that associations between diet and diabetes were investigated in a large population-based study, allowing us to perform stratified analyses by metabotype with sufficient sample sizes. Due to the lack of a uniform definition of the term ‘metabotype’, metabotyping was performed in KORA FF4 analogous to Riedl et al. [23] in KORA F4 to get comprehensive metabotypes based on a broad range of parameters [22]. However, only 16 of the 34 biochemical and anthropometric parameters originally used in F4, were available in FF4. Repeating the identification of metabotypes in F4 with the reduced set of 16 parameters also available in FF4 and comparing to the originally identified metabotypes in F4 based on 34 parameters, revealed a similar allocation of individuals to the clusters (1513 of 1729 individuals or 87.5% of participants). In addition, the newly defined metabotypes in FF4 based on the 16 parameters showed a good distinction of demographic and metabolic characteristics. Consequently, we assume to have identified metabotypes in KORA FF4 that are still comprehensive despite the reduced set of biochemical markers and we assume that these metabotypes were appropriate to consider metabolic differences in diet–diabetes associations. Another strength is the availability of extensive dietary data assessed by food frequency questionnaire and up to three 24-h food lists, which enabled the investigation of a large number of food items and their association with diabetes. However, as with all dietary assessment methods, misreporting cannot be ruled out. Further strengths include the assessment of diabetes by either a physician-validated diagnosis or an OGTT, and the availability of a large number of confounders for adjustment. Limitations of the study include the fact that a large proportion of the original S4 participants (1982 of 4261 individuals) did not participate again in the second follow-up KORA FF4 study due to death, refusal and loss to follow-up, what could have biased our results. Furthermore, due to the cross-sectional study design, no causal relationships between dietary factors and diabetes could be established and longitudinal or intervention studies considering metabolic differences are needed. In addition, samples sizes, dietary intake amounts and the diabetes prevalence varied considerably between metabotype subgroups, which could have influenced our results.