The UK Biobank is a large-scale biomedical database started from 2006. The aim of the program is to investigate the influence of genetic and environmental factors on disease development. It has recruited more than 500,000 volunteers aged 40 to 69 across the UK and will follow them over the next 30 years. Participants were invited to the assessment centers to complete a series of lifestyle, health and socioeconomic interviews, and physical measurements. In addition, biological samples were also collected. All disease conditions, prescription drug use, and deaths of them during the whole study period will be recorded through the centrally managed UK National Health Service system (22).

The initial 502,507 participants were invited to provide information on their food consumption in the past year through a touch questionnaire at the assessment centers. Subsequently, a sub-sample of about 20,000 population was also invited to repeat the questionnaire at assessment centers every 4 years, to examine the possible changes in participants' responses to the questionnaire and their dietary intake over time. Meanwhile, in order to gather more detailed information about the actual amount of food or food groups actually consumed by the participants, the UK Biobank also adopted an online 24-h dietary assessment tool named Oxford WebQ. It was developed for use in large population studies and had been validated in previous studies (23, 24). It collected additional detailed dietary intake information of 210,970 participants at least once through the 24-h recall. Participants were asked whether they have consumed the predefined 206 foods or 32 drinks in the past 24 h (24). After that, some of them were also invited to repeat the online 24-h questionnaire for a total of four times through emails between February 2011 and June 2012 every 3–4 months. At the same time, participants were also asked whether they consumed over the previous 24 h were fairly typical for their daily life (25).

Participants who reported consumption of at least 7 foods included in the EAT-LDP based on the 24-h dietary assessment tool were included in the analysis first (N = 69,686). Later, who reported history of any cancer (2,953), cardiovascular disease (1,923), and diabetes (1,269) at the baseline were also excluded. We also excluded participants with abnormal total energy intakes (<2,093 or >14,650 kJ/day in female and <3,349 or >16,743 kJ/day in male participants). Finally, we excluded participants who were followed for <1 year (N = 236) to minimize the potential for reverse causality bias. As a result, a total of 59,849 participants were considered for inclusion in the following main analysis (Figure 1). All participants had given written informed consent.

The EAT-LDP score used in this study was designed by Knuppel et al. (20). The distribution of the diet score can be seen in Supplementary Table 1. Definition of portion size and food items used in this study can be seen in Supplementary Table 2. The EAT-LDP score is consisted of 8 main dietary components, including whole grains, tubers and starchy vegetables, vegetables, fruits, dairy foods, protein sources, added fats, and added sugars. Participants were assigned with a point for meeting each of the recommendations. Each dietary component contributed 0 or 1 point resulting in a total score ranging from 0 to 14 points. The higher dietary scores indicated a greater adherence to the individual healthy eating patterns.

The online 24-h dietary assessment tool did not record the concrete weight of consumed food, but the number of predefined portion size was defined using the UK's standard food composition database (26, 27) (e.g., how many bowls of cereals, how many serving of Quorn they ate in the last 24 h, Supplementary Table 2). We then calculated the quantity by multiplying the portion size by the number of portions consumed for each food item.

The primary outcome of this study was the incident cases of T2D. The T2D diagnosis during follow-up was ascertained according to the International Classification of Diseases, 10th edition (ICD-10) (28) code from the hospital where the inpatient records containing data on admissions and diagnoses linked to the Hospital Episode Statistics for England, Scottish Morbidity Record data for Scotland, and the Patient Episode Database for Wales. Record linkage was available until March 2020.

Sociodemographic characteristics and lifestyle factors were recorded at baseline, including age, sex, ethnicity (White, Asian or Asian British, Black or Black background, Chinese, mixed background, etc.), Townsend deprivation index (an index constructed based on the material deprivation degree) (29), education attainment (college or university degree, professional qualifications, etc.), smoking status (previous, current, and never), drinking status (previous, current, and never), body mass index (BMI, kg/m²), waist circumference (cm), total energy intake (kJ/day), and metabolic equivalent (MET; min/week for all activity including walking, and moderate and vigorous activity). A separate category was introduced for those who had no available data on smoking (164 missing) or drinking (64 missing) status.

We used mean and SD to express normally distributed continuous variables, and number (percentages) to express categorical variables. Baseline characteristics are summarized and compared by using the χ² test for categorical, or ANOVA or Mann–Whitney U-test for continuous variables, as appropriate. If covariate information was missing, we imputed the mean values for continuous variables or used a missing indicator approach for categorical variables.

We calculated each participant's person-years from the date of the return of the baseline questionnaire (2006–2010) to the date of diabetes diagnosis, death, loss to follow-up, or the end of follow-up (March 31, 2020), whichever came first. Incidence rates of T2D events per 1,000 person-years were calculated by the EAT-LDP score. We used Cox proportional hazards models to examine the association between the diet score and the subsequent risk of T2D. The proportional hazards assumption was tested with Schoenfeld residuals against survival time.

We divided the study participants into 3 groups according to their summed EAT-LDP scores and compared the HRs with Tertile 1 as the reference group. The cut-offs for the groups were chosen so that the number of participants in each group is reasonable and similar within the allowable range of data distribution (Supplementary Figure 1). Meanwhile, the mean food intakes at baseline were calculated from the participants within each group who had reported consuming the specific food component.

Four models were estimated, and covariates were added in a stepwise manner. The Model 1 was the crude model. In Model 2, we further adjusted for sociodemographic covariates, including age sex, deprivation, qualifications, and ethnicity. In Model 3, we additionally adjusted for lifestyle factors included smoking, drinking, energy intake, and MET. And Model 4 additionally included baseline BMI and waist circumference. In addition, we explored the potential non-linear associations using 3-knotted restricted cubic spline regression. We also viewed the diet score as a linear variable to examine the risk reduction associated with a 1-point increment.

A mediation analysis was considered to explicate the association of diet score with T2D; indirect associations acting through BMI as a mediating variable and direct associations not mediated by BMI were quantified. At the same time, a mediation analysis was also conducted when waist circumference was included in the Cox models. We used paramed package in Stata version 16 (StataCorp., College Station, TX, USA) to conduct the mediation analysis. This is a parametric regression approach that estimates two models. For example, one model for the mediator (BMI) conditional on exposure (diet score) and covariates, and another model for the T2D conditional on diet score, BMI and covariates. It extends statistical mediation analysis to allow for the presence of exposure-mediator (diet score and BMI) interactions in the outcome regression model using counterfactual definitions of direct and indirect effects. The proportion of the association mediated by BMI or waist circumference was calculated by the formula: NIE/(NIE + NDE) (32). NIE represents the natural indirect effects, and NDE represents the natural direct effects of EAT-LDP score on T2D. We also used testparm package to test the statistical significance of potential interaction between the exposure variable and covariates (30). The p-values for interaction were evaluated using interaction terms and likelihood ratio tests.

Several sensitivity analyses were performed to test the robustness of our findings as follows. First, we conducted stratified analyses that were defined a priori by ethnicity and other potential risk modifiers, including age, sex, deprivation, education attainment, smoking, and drinking status. Second, we repeated all analyses after excluding all participants who developed T2D during the first 3 years of follow up to reduce the possibility of spurious association due to reverse causation. Third, we conducted the multiple imputation (mi package) to impute the missing covariates, and 5 imputed datasets were pooled using Rubin's rules. The statistical analysis was conducted using Stata version 16 (StataCorp., College Station, TX, USA). The statistical significance was set as p < 0.05 (two-sided test).

We described the baseline characteristics of the included participants according to the tertiles of EAT-LDP score in Table 1. A total of 59,849 participants were included in the analyses, with an average age of 55.9 years. Among them, there are 34,512 (57.67%) female and 25,337 (42.33%) male participants. The mean dietary score among these targeted participants was 5 (range from 1 to 10). Participants with higher EAT-LDP scores tended to be older [mean (SD) age, 56.96 (7.89), p < 0.001], female (63.92%, p < 0.001), white (94.94%, p < 0.001), be least deprived (21.79%, p < 0.001), have other education attainment (51.22%, p < 0.001), be never smokers (60.84%, p < 0.001), with lower BMI [mean (SD), 26.36 (4.44), p < 0.001], smaller waist circumference [mean (SD), 86.66 (12.65), p < 0.001], and have higher total physical activity [PA, mean (SD), 2,756.06 (2547.09), p < 0.001]. Overall, the group with the lowest diet score had the highest incidence rate of T2D (Table 2).

The proportion of target population adhering to each diet component is shown in Supplementary Table 3. The mean intake for each food component (for whole population and in each tertile) at baseline is also shown in Supplementary Table 4. In this study, 82% of the 24-h diet recalls were considered to reflect typical eating habits. Of the 59,849 selected participants finally included in the analysis, more than 50% of them consumed the recommended weights of whole grains and refined grains, all vegetables, all fruits, dairy foods, eggs, and soy foods. Participants with the highest EAT-LDP scores (3rd tertile) were more likely to consume the recommended weights of potatoes, all vegetables, dairy foods, protein foods, legumes, added fats, and added sugars.

Results of the associations between the EAT-LDP adherence and risk for T2D are shown in Table 2. During a mean follow-up time of 10 years, a total of 2,461 adults experienced the incident T2D (4.11 per 1,000 person-years, 95% CI: 3.95–4.27). Across tertiles of the EAT-LDP score (from lowest to highest), the incidence of T2D decreased from 4.74 to 3.44 per 1,000 person-years. All models, except the Model 4 adjusted for BMI and waist circumference, consistently showed a significant dose–response relationship between the increased diet score and risk of T2D. In multivariate analysis (Model 3), compared with the group with the lowest diet score (1st tertile), the HR for T2D was 0.81 (0.72–0.90).

After adjusting for potential confounders, each additional point of the EAT-Lancet diet score was associated with a 6% decreased risk of T2D (HR: 0.94, 95% CI: 0.91–0.97). A linear association between the diet score and incident T2D using restricted cubic spline regression can be found in Figure 2.

Supplementary Figure 2 shows the association between the adherence of specific food recommendations and incident T2D. We found that adhering to the recommended intake of potatoes, all vegetables, all fruits, dairy foods, beef, lamb and pork, and eggs was associated with lower incidence of T2D in this analysis. In our research, 914 participants met the recommendation ( ≤ 31 g/day) of all sugars. The result shows that the consumption of all sugars was significantly associated with the T2D risk (HR = 1.85, 95% CI: 1.45–2.37). Considering that the purpose of this research is to explore the relationship between the specific diet pattern and T2D, a single-diet component is not the focus of this study.

Table 3 presents the total, direct, and indirect associations between the EAT-LDP score and T2D, and 44% of this association was mediated by BMI as mediator variable. The total effect was significant (β = 0.56, 0.54–0.57, p < 0.0001), with the natural direct effects (β = 0.85, 0.82–0.88, p < 0.0001) and the natural indirect effect (β = 0.66, 0.65–0.67, p < 0.0001). Similar results were found within sex-stratified analyses. The relationship between the EAT-LDP score and risk of T2D was 36% mediated by BMI among male and 53% among female participants.

Table 3 also shows that the mediating role of the waist circumference in the association between the EAT-LDP score and incident of T2D (40%). Similarly, the relationship between the EAT-LDP score and risk of T2D was 35% mediated by waist circumference among male and 50% among female participants.

We observed similar results in the analysis that included participants who were followed for <3 years, which reinforced the robustness of our findings (Supplementary Table 5). The results did not significantly change when the multiple imputation was conducted (Supplementary Table 6). We also explored the associations between the EAT-LDP score and incident T2D stratified by age, sex, ethnicity, Townsend deprivation index, education attainment, smoking status, and drinking status. In the stratified analysis, compared tertile 3 of the EAT-LDP score with tertile 1, the hazard ratio (HR) for T2D in female participants was 0.72 (95% CI: 0.61–0.86). The protective effects of adherence to higher EAT-LDP score can be seen in the white, never and previous smoking, and current drinking participants (Supplementary Table 7). Meanwhile, the effect modification by these covariates is not significant.

This study has the advantages of prospective design and large sample size of diet habits to explore the association between the healthy and sustainable EAT-LDP and incident of T2D. Our results of the main analysis are shown to be consistent with the sub-analyses. There are some potential limitations warrant consideration. First, each component of EAT-LDP score was constructed as a binary variable (adherence to the target intake levels vs. non-adherence). This may lead to the loss of some dietary information. A more refined scoring method should be developed to investigate its association with the possible health status. Second, participants in our study can only represent middle-aged and elderly people. And 93% of people are white. Target participants included in our main analysis were those who reported consumption of at least 7 foods according to the EAT-LDP. Therefore, the obtained results could not be generalized to other population with different characteristics. At the same time, the dietary information used in this analysis mainly comes from the baseline assessment, which may not reflect the potential changes in participants' eating habits. Third, patients diagnosed early were followed for a longer period of time than patients diagnosed in recent years. Longer follow-up time would allow the increase of the duration between nutritional assessment and assessment of the T2D. Fourth, T2D in this research was diagnosed through inpatient medical records. Although doctors' diagnosis is a more common and precise way, the actual incidence of T2D could be underestimated. Last but not least, although we have adjusted for different confounding factors, there may be residual of unmeasured confounding factors that cannot be excluded in the observational studies. More validation is needed for reliable estimation of the associations between EAT-LDP and the possible adverse health outcomes.

In light of the increasing global burden of diabetes, our results seem to be clinically relevant for diabetes prevention, and the EAT-LDP is an achievable and sustainable objective that should be promoted.