The UK Biobank is a large-scale population-based prospective cohort study of >500,000 participants aged between 40 and 69 years and recruited between 2006 and 2010 from the UK. Participants attended one of 22 assessment centres located across England, Scotland, and Wales, where they completed a baseline assessment comprising a touchscreen questionnaire and a series of physical assessments. Further information on the study protocol has been previously reported [14]. All UK Biobank participants provided written informed consent at recruitment and the study received ethical approval from the NHS North West Multicentre Research Ethics Committee (Ref. 11/NW/0382).

Participants who completed fewer than two 24-h dietary assessments (n = 372,173), had implausible energy intakes (>17,573 KJ or <3347 KJ for men and >14,644 KJ or < 2092 KJ for women [15]) (n = 3953), or withdrew their consent during the study follow-up were excluded from the analysis. Participants were also excluded if they had diabetes (n = 5009), CVD (n = 1868), cancer (n = 5920), were diagnosed with T2D before the last 24-h dietary assessment (n = 301), or if they were pregnant at baseline assessment (n = 46) (Supplemental Fig. S1).

Further details on the Oxford WebQ dietary assessment tool, including the method used to estimate flavonoid intakes are provided in the Supplementary Appendix. The top three foods that contributed the highest percent to total flavonoid and flavonoid subclass intakes were considered eligible for inclusion in the FDS (excluding fruit juices). In addition, dark chocolate was also considered eligible for inclusion within the score as it typically contains high concentrations of flavonoids. However, in this UK population, intakes of dark chocolate were low, with a median (range) of 0 (0–2.5) servings/day, and thus, did not contribute highly to flavonoid intakes. Based on mean intakes (servings/day) from a minimum of two 24-h dietary assessments, the score was made up of ten Oxford WebQ food items (tea (black and green), red wine, apples, berries, grapes, oranges (including satsumas), grapefruit, sweet peppers, onions, and dark chocolate). A final FDS was calculated by summing the total number of servings consumed across the ten selected food items. Tea (black and green) intakes were capped at a maximum of 4 servings/day due to the high intakes reported in this British population (median (range) 2.3 (0–11.5)). From this, the FDS was categorised into sex-specific quartiles. As we were interested in flavonoids rather than alcohol consumption, (a) we adjusted for total alcohol consumption in our statistical models on the FDS and T2D risk, and (b) we re-ran our statistical analyses on the FDS excluding red wine consumption.

As previously conducted for food groups in the UK Biobank [16], intraclass coefficients (ICCs) were calculated to test the reproducibility of the FDS over time and Spearman’s rank correlation coefficients were calculated to test the reliability of flavonoid intakes over time, using average flavonoid-rich food intakes from the Oxford WebQ dietary assessments at the second and third assessment (collected in February–April 2011 and June–August 2011) versus the fourth and fifth (October–December 2011 and April–June 2012). A subgroup of 21,543 participants who completed each of these dietary assessments were included in this analysis. Further, Spearman’s rank correlation coefficients were calculated to test the agreement between flavonoid intakes estimated from U.S Department of Agriculture (USDA) and Phenol-Explorer databases.

Sociodemographic, dietary, and lifestyle factors were self-reported at the baseline assessment between 2006 and 2010. Covariates considered in this study included: sex, age at recruitment, ethnicity, geographical region of recruitment, education, physical activity, alcohol intake, smoking status, energy intake, BMI, waist circumference, polypharmacy index, multimorbidity index, genetic risk of T2D (polygenic risk score (PRS) generated and issued by the UK Biobank for use upon request [17, 18]), Townsend deprivation index, family history of diabetes, hypercholesterolemia, hypertension, menopausal status, number of completed dietary assessments, wholegrain intake, sugar-sweetened beverage intake, red and processed meat intake, coffee intake and the healthful plant-based diet index [19]. Further details on how the covariates were classified for this study can be found in Table S1.

Incident T2D cases were defined as primary type 2 diabetes mellitus according to the International Classification of Diseases 10th edition (ICD-10) (E11), using UK Biobank linked hospital inpatient data on admissions and diagnoses available until the 30th of September 2021 from the Hospital Episode Statistics (HES) for England, 31st of July 2021 for Scottish Morbidity Records (SMR), and 31st of March 2016 for the Patient Episode Database for Wales (PEDW). Follow-up for incident T2D analyses was censored at date of hospitalisation, death, or end of follow-up, whichever occurred first.

Cox proportional hazards regression models were used to assess the relationship between the FDS, flavonoid subclass intakes, and intakes of individual flavonoid-rich foods, and incident T2D, producing hazard ratios (HR) and 95% confidence intervals (CI). The FDS was divided into sex-specific quartiles according to summed scores. Flavonoid intakes and major flavonoid-food contributors were also grouped into sex-specific quartiles, with the lowest quartile serving as the referent group across all analyses. The FDS, flavonoid subclass intakes, and flavonoid-rich foods were also modelled as continuous exposure variables when carrying out linear trend tests (P-trend).

Two models were used for adjustment in all analyses. Model 1 was adjusted for potential sociodemographic confounders; sex (female, male) and education (Low: CSEs or equivalent, O levels/GCSEs or equivalent; Medium: A levels/AS levels or equivalent, NVQ or HND or HNC or equivalent; High: College or University degree, other professional qualifications eg: nursing, teaching; unknown/missing/prefer not to say (6.6%)), stratified by age (<45 years, 45–, 50–, 55-, 60–, ≥65 years) at recruitment and geographical region of recruitment (ten UK regions). Model 2 was further adjusted for body mass index (BMI) (≤18·5 kg/m2, 18·5–24·9 kg/m2, 25·0–29·9 kg/m2, ≥30 kg/m2, or unknown/missing (0.1%)), waist circumference (continuous scale, cm), ethnicity (Asian, Black, Multiple, White, Other, or unknown/missing (0.3%)), physical activity (METs hr/week in quintiles, or unknown/missing (1.8%)), smoking status (never, previous, current, or unknown/missing (0.2%)), alcohol intake (<1 g/day, 1–7 g/day, 8–15 g/day, 16+ g/day, or unknown/missing (16.3%)), energy intake (continuous scale, kJ/day), polypharmacy index (total number of self-reported medications taken at baseline; 0, 1–3, 4–6, 7–9, >10, or unknown/missing (0.0%)), multimorbidity index (number of pre-existing long-term conditions; 0, 1, 2, or >3), hypercholesterolemia (no, yes), hypertension (no, yes), Townsend deprivation index (quintiles from low to high deprivation index, or unknown/missing (0.1%)), family history of diabetes (no, yes), menopausal status (no, yes, not sure (among women), or unknown/missing (0.1%)), PRS (tertiles from low to high PRS for T2D, or unknown/missing (2.1%)), number of completed dietary assessments (continuous scale, ranging between 2 and 5), wholegrain intake (continuous scale, servings/day), sugar-sweetened beverage intake (continuous scale, servings/day), red and processed meat intake (continuous scale, servings/day), and coffee intake (continuous scale, servings/day). Sensitivity analyses were also conducted, providing an alternative statistical model which involved replacing wholegrain, sugar-sweetened beverages, red and processed meat, and coffee intake with the healthful plant-based diet index (with a score ranging from 31 to 84 points) as a potential confounder. Further, a sensitivity analysis excluding study participants with less than two years of follow-up time to examine whether observed associations were due to reverse causality was also carried out. On completion of the Schoenfeld residuals test, there was no evidence to suggest the violation of the proportional hazards assumption.

To investigate effect modification, stratified analyses were carried out across potential risk modifiers smoking status (never, ever), sex (male, female), BMI (<25, ≥25 kg/m2), education (low: GSEs/O-Levels/GCSEs or equivalent, NVQ/HND/HNC/A-Levels/AS-Levels or equivalent; high: Other professional qualifications, College/university degree), ethnicity (white, non-white), and alcohol intake (<1 g/day, >1 g/day). Further, heterogeneity was assessed across strata of polygenic risk (T2D) (PRS tertiles; low, intermediate, high). Likelihood ratio tests (LRT) were used to test for potential effect modification of associations between the FDS and T2D risk by covariates. Multiplicative terms between the FDS (continuous) and covariates along with the main effect term were added to fully adjusted Cox regression models. LRTs were used to compare models with and without multiplicative terms.

A mediation analysis was also carried out to assess the association between potential mediators (biomarkers of obesity/sugar metabolism, inflammation, kidney and liver function, and lipid metabolism) on the FDS-T2D pathway. Details on the rationale for the choice of potential mediators and on the statistical methods are provided in the Supplementary Appendix.

Stata version 17.0 (Stata Corp LP, College Station, TX) was used to conduct all analyses. All reported p values were two-sided, with statistical significance set at p < 0.05 for main analyses. To account for multiple testing in flavonoid-rich food and flavonoid subclass analyses, Bonferroni correction was used [20].