Study designs and populations

The overall study design and participant flows are displayed in S1 Fig. Briefly, we derived a nutritional biomarker score in the MedLey trial as an objective measure of adherence to the Mediterranean diet, and we applied the score in EPIC-InterAct to test its association with incident T2D. We developed 2 protocols for earlier iterations of this research idea as parts of broader statistical analysis plans. Key analytical decisions pertinent to the current analysis were prespecified in these documents (S1 Protocol).

The MedLey trial is an RCT that tested effects of the Mediterranean diet on cardiovascular risk factors [13] and cognitive function [14] among healthy adults aged ≥65 years from metropolitan Adelaide, Australia. We randomised 166 participants to either the Mediterranean diet with maintenance of baseline body weight or continuation of habitual diet for 6 months in 2013 and 2014 and measured circulating carotenoids [15–18] and fatty acids [15,19–22] to assess compliance [23,24]. Circulating carotenoids are biomarkers of dietary carotenoids or intake of fruits and vegetable [25,26]. Fatty acid profiles reflect across intakes of individual fatty acids and their metabolism, and are affected by dietary fat composition, intake of fish and shellfish, dairy products, nuts, and added fats, including olive oil [27–29]. The Mediterranean diet intervention consisted of fortnightly sessions with the study dietitian and provision of key shelf-stable foods amounting to 30% to 35% of estimated energy requirements: virgin olive oil, low-fat Greek yoghurt, unsalted nuts, canned legumes, and canned tuna. Participants in the habitual diet control group received gift vouchers to local food stores [23,24].

We used the MedLey trial to derive a biomarker score of discrimination between the Mediterranean and habitual diet arms based on end-of-study nutritional biomarkers. Among 137 trial completers out of 166 randomised individuals, we excluded participants with missing nutritional biomarker data (n = 4) and those with at least 1 biomarker value outside of the 25th percentile minus 3 times the interquartile range (IQR) or the 75th percentile plus 3 times the IQR (n = 5). The analytical sample (n = 128) included 67 participants in the Mediterranean diet group and 61 in the habitual diet arm (S1 Fig).

EPIC-InterAct Study is a case-cohort study of T2D nested within the EPIC study in 8 European countries (Denmark, France, Germany, Italy, the Netherlands, Spain, Sweden, and the United Kingdom) [30]. Baseline data collection took place during 1991 to 1998. We ascertained and verified 12,403 individuals with incident T2D from a cohort of 340,234 participants with stored blood samples in the EPIC study. Incident T2D cases were ascertained from a combination of sources, including self-report, primary and secondary care registers, drug registers, hospital admissions, and mortality data, as described in detail previously [30]. From the cohort of 340,234 participants, a centre-stratified subcohort was assembled by randomly selecting 16,835 individuals. A total of 16,154 participants remained in the subcohort after exclusions (n = 548 with prevalent diabetes; n = 133 with uncertain diabetes status). As a result, a case cohort was established (n = 27,779, including 12,403 incident T2D cases; S1 Fig) [30].

For the current analysis, we excluded participants with missing biomarkers required to calculate the biomarker score (n = 5,577), leaving 22,202 participants, with 9,453 participants who developed incident T2D and 13,313 subcohort participants (S1 Fig) followed over an average of 9.7 years. Excluded participants were similar to those included in the analytical sample in terms of the mean baseline age (53.0 versus 51.6 years), BMI (26.3 versus 26.1 kg/m²), and the proportion of women (59.0% versus 62.8%) based on subcohort distributions.

All participants provided written informed consent and studies were approved by local ethics committees. The MedLey trial was prospectively registered in the Australian New Zealand Clinical Trials Registry (ACTRN12613000602729). This study is reported as per the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guideline (S1 Checklist) and Consolidated Standards of Reporting Trials (CONSORT) statement (S2 Checklist).

Statistical analysis

Stata 16.1 and R 4.0.2 were used for statistical analysis.

Patient and public involvement

The MedLey trial intervention was piloted in a group of volunteers that led to minor adjustments in its design and delivery [56]. Patients or public were otherwise not involved in study design.

Derivation of the biomarker score

The biomarker score consisted of a linear combination of 23 biomarkers in total (S2 Table). The score distinguished 2 arms of the Mediterranean and habitual diet groups with C-statistic = 0.88 (95% CI: 0.82 to 0.94) and moderate calibration (S3 Fig). We did not find any evidence of interaction between medication use and the biomarker score (S3 Table). After standardisation of the score in the MedLey trial, mean end-of-trial score values were higher by 1.81 (95% CI: 1.45 to 2.18) points in the Mediterranean than the habitual diet group. In EPIC-InterAct, baseline medians of the biomarker score standardised using the overall subcohort distribution ranged between countries from −0.35 in Spain to 0.71 SD in Sweden (Fig 1 and S4 Table). The C-statistic values for discrimination between extreme categories of self-reported Mediterranean diet by the biomarker score were 0.63 (95% CI: 0.38 to 0.87; n = 42) in the MedLey trial baseline sample and 0.59 (95% CI: 0.58 to 0.61; n = 4,298) in the EPIC-InterAct subcohort, with country-specific values ranging from 0.54 in Germany to 0.68 in Spain.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Nutritional biomarker score of the Mediterranean diet: study-specific distribution by the MedLey trial arms post-intervention and the EPIC-InterAct subcohort by country at baseline.

The biomarker score was derived as a discriminatory model between the Mediterranean and habitual diet in the MedLey randomised partial-feeding controlled trial. Circulating carotenoids and fatty acids were used to calculate the score as linear predictions from the discriminatory model. The score was standardised separately within the MedLey trial and the EPIC-InterAct subcohort. Boxes denote the IQR and medians inside; and whiskers, values up to 1.5 IQR outside of these percentiles. EPIC, European Prospective Investigation into Cancer and Nutrition; IQR, interquartile range; UK, United Kingdom.

https://doi.org/10.1371/journal.pmed.1004221.g001

The association of the biomarker score with incident T2D

The biomarker score of the Mediterranean diet was inversely associated with incident T2D (Table 2). In the adiposity-adjusted multivariable model, the HR (95% CI) for the top fifth of the biomarker score compared to the bottom fifth was 0.38 (0.30 to 0.50) (p trend = 0.012). The HR (95% CI) per 1 SD was 0.71 (0.65 to 0.77) with inverse associations in all countries, moderate heterogeneity between country-specific estimates (I² = 67%) and a 95% prediction interval of 0.55 to 0.91 (Fig 2). We found weak evidence of departure from linearity (p nonlinearity = 0.044, Fig 2) where the inverse association levelled off in approximately the upper fifth of the distributions of the score. The association remained statistically significant after additional adjustments for individual and all biomarkers or interaction terms included in the score (S5 Table). The estimated PAF (95% CI) was 11% (7% to 14%), indicating that the incidence of T2D could be reduced by 11% if the biomarker score were increased by 10 percentiles, assuming a causal relationship. For comparison, the estimated PAF (95% CI) for a 10 percentile lower BMI was 28% (20% to 35%).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Association between the nutritional biomarker score of the Mediterranean diet and incidence of T2D in the EPIC-InterAct case-cohort study (n = 22,202).

The biomarker score was derived as a discriminatory model between the Mediterranean and habitual diet in the MedLey randomised partial-feeding controlled trial. Circulating carotenoids and fatty acids were used to calculate the score as linear predictions from the discriminatory model. Associations were assessed with the Prentice-weighted Cox regression and pooled by random-effects meta-analysis. Top: The diamond and error bars of the pooled estimate represent the 95% confidence and prediction intervals. Bottom: Restricted cubic splines with 5 knots were used to model the nonlinear association. The p-value for the test of nonlinearity was 0.044. Black solid line represents point estimates of HRs and purple area denotes the 95% CI. The 10th percentile of the subcohort distribution was used as reference. Associations were adjusted for: age (as timescale), sex, recruitment centre, prevalent cancer, cardiovascular disease, hypertension and hyperlipidaemia, familial history of type 2 diabetes, smoking status (never, former, current smoker), physical activity index (inactive, moderately inactive, moderately active, active), seasonality (sine and cosine function of the day of the year), fasting status (<3, 3–6, >6 h), current use of vitamin or mineral supplements, marital status (single, married or cohabiting, divorced or separated, widowed), educational attainment (none, primary school, technical or professional school, secondary school, post-secondary school education), current employment, body mass index and waist circumference, and in women, menopausal status (pre-, peri-, postmenopausal, bilateral oophorectomy) and current hormone replacement therapy use. CI, confidence interval; HR, hazard ratio; SD, standard deviation; T2D, type 2 diabetes.

https://doi.org/10.1371/journal.pmed.1004221.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Associations between the nutritional biomarker score of the Mediterranean diet* with incidence of T2D in the European Prospective Investigation into Cancer and Nutrition-InterAct (n = 22,202).

https://doi.org/10.1371/journal.pmed.1004221.t002

The main result was robust to multiple sensitivity analyses which explored the effects of time of follow-up, reverse causation, alternative analytical decisions at the stage of derivation of the biomarker score, and additional adjustment for use of medications (S5 Table). Among the potential effect modifiers, we found evidence of interaction of the biomarker score with age and the use of dietary supplements (S6 Table). The HR (95% CI) per 1 SD of the biomarker score was 0.76 (0.69 to 0.84) in supplement users and 0.69 (0.62 to 0.75) in non-users. The stratum-specific estimates by age at baseline <45, 45–60, and >60 years were 0.54 (0.42 to 0.69), 0.74 (0.69 to 0.80), and 0.74 (0.67 to 0.82), respectively. The results from complete-case analysis were similar to the multiple imputation estimates (S6 Table). The HR (95% CI) per 1 SD of the score of self-reported Mediterranean diet was 0.90 (0.86 to 0.95).

Comparison with previous studies

Attempts to derive biomarkers of the Mediterranean diet and other dietary patterns have previously been largely confined to metabolomic profiling using cross-sectional designs [7,58]. In a subgroup analysis in one of the centres of the PREDIMED trial, urinary metabolomic profiles at 1 or 3 years post-randomisation were able to correctly classify 93%, 85%, and 68% of participants to their respective intervention arms of the Mediterranean diet with either olive oil or nuts and the control lower fat diet [8]. Only 2 previous studies considered scores comprised of nutritional biomarkers [5,59], one of which used a hypothesis-free approach that was likely prone to confounding by non-dietary regulation of nutritional biomarkers [59]. The second study used a novel feeding design of habitual diets and performed data-driven variable selection from a wider range of nutritional biomarkers than available in our investigation: serum phospholipid fatty acids, carotenoids, tocopherols, retinol, B vitamins, and 24-h urinary nitrogen, sodium, potassium, and energy expenditure estimated from the doubly labelled water technique [5]. The study derived a biomarker score that was well correlated with quantified habitual adherence to the Mediterranean diet (r_{cross-validated} = 0.60). It included 2 circulating carotenoids and 7 fatty acids that constituted 81% of the prediction performance, as well as urinary potassium and energy expenditure [5]. This finding corroborates our results and the hypothesis that combinations of nutritional biomarkers, and in particular circulating carotenoids and fatty acids, can be used to objectively assess adherence to the Mediterranean diet [15–22]. Importantly, it also highlights the remaining biomarkers not selected into the score as potentially having limited value for assessment of the Mediterranean dietary pattern, which is supported by interventional evidence for some of these analytes, e.g., for tocopherols [16]. Studies on other dietary patterns likewise suggest utility of combinations of carotenoids and fatty acids for their objective assessment [5,60–62]. It is prudent for future research on biomarker scores of the Mediterranean diet to evaluate further groups of nutritional biomarkers in a data-driven manner. For example, biomarkers of polyphenols are promising [21,63], however, the evidence is limited and their utility has so far not been assessed in the context of multivariate modelling.

Our work on derivation of the nutritional biomarker score of the Mediterranean diet adds to the previous literature by incorporating the strength of an RCT. Owing to the RCT design of the MedLey trial, we were able to minimise confounding by non-nutritional factors, and thus derive a biomarker score with| potentially improved generalisability to external populations. A further novelty of our work is the application of the experimentally derived score to prospective associations between dietary patterns at large and disease outcomes. To our knowledge, the only other example of derivation and application of a putative biomarker of the Mediterranean diet to associations with disease endpoints was a metabolomics-based analysis in the Spanish PREDIMED study and prospective cohorts in the USA. Similar to our analysis, it observed inverse associations for cardiovascular disease using the metabolite score observationally derived in PREDIMED across all cohorts (HR range per 1 SD: 0.71 to 0.86) and weaker or null associations when using the self-reported Mediterranean diet [64]. Of note, these results also provide evidence in support of transferability of objective exposure assessment measures across populations for epidemiological investigations. This is particularly pertinent to the current manuscript in which we were unable to test the association of the biomarker score with incidence of T2D in the target population of the MedLey trial.

For incident T2D, the PREDIMED trial reported a 30% risk reduction in the Mediterranean diet intervention arms relative to the control lower fat diet group (273 incident cases in 3,541 participants) [65]. The CORDIOPREV trial reported a statistically nonsignificantly higher incidence of T2D in the Mediterranean diet arm compared to the lower fat control group (HR 1.35; 95% CI: 0.91 to 2.01; 107 cases in 462 participants) [66]. This finding from a secondary CVD prevention trial among participants predominantly with pre-diabetes (85% prevalence) is of lower relevance to prevention of T2D in the general population than that of the PREDIMED trial. Beyond moderating effects of population characteristics, the result of the CORDIOPREV trial may have been driven towards favouring the lower fat diet by a small mean weight loss over 5 years (−1.14 kg) compared to a small average weight gain in the Mediterranean diet arm (+0.78 kg) [66]. Inverse associations between the self-reported Mediterranean diet and incident T2D in middle-aged adults have previously been reported in EPIC-InterAct [12] and other prospective cohorts [2]. The largest reduction in the incidence between extreme categories of the self-reported adherence to the Mediterranean diet was 25% (top versus bottom fifth) [67], which is a substantially smaller effect size than the 62% observed in the current study using the biomarker score.

Strengths and limitations

The major strength of the current research was the use of a novel analytical approach that combined the derivation of an objective measure of the Mediterranean diet in a partial-feeding study (the MedLey RCT) and its application in a large observational study (the EPIC-InterAct Study). The RCT compared the effects of this dietary pattern (without weight loss) with continuation of habitual diet on nutritional biomarkers. Such an experimental design allowed us to derive a biomarker score of the Mediterranean diet in a manner that was free from influences of other dietary and non-dietary factors, while using a control group suitable for application of the score to study participants in an observational setting in Western countries. We used a set of biomarkers that reflect dietary exposures over the past weeks or months [68] that is a desirable timeframe for assessment of habitual diet for epidemiological applications. Our analysis was based on the largest study to date of nutritional biomarkers and T2D, including over 9,000 incident cases. Among other strengths, our observational analyses adjusted for a comprehensive range of potential confounding factors and included several sensitivity analyses, compared the main results against the Mediterranean diet assessed by self-report, and modelled the population impact of greater adherence to the Mediterranean diet on future T2D risk.

Our research had several limitations. The intervention in the MedLey trial was administered at one level of intensity. This allowed for modelling of the Mediterranean diet only as a binary variable and precluded objective evaluation of the dose-response relationship between the adherence to this dietary pattern and the biomarker score [69]. Second, the MedLey trial was a partial-feeding RCT that may have resulted in lower adherence to the dietary components not provided (e.g., fruits and vegetables) than the food items provided as part of the Mediterranean diet intervention. Third, we used a combination of candidate biomarkers of intake, concentration, and function that are not only affected by dietary intakes, but also bioavailability, endogenous synthesis, genetic variation, homeostatic control, and nutrient metabolism [27,70,71]. Thus, changes in nutritional biomarkers in the MedLey trial may have represented a metabolic response to a healthy diet, rather than a specific biomarker profile indicative of adherence to the Mediterranean diet. Fourth, we were unable to assess whether participants in the intervention group reached equilibrium in concentrations of nutritional biomarkers and values of the biomarker score. Fifth, the trial experienced a moderate degree of drop-outs between the randomisation and the end-of-trial assessment (20% in each arm, including missing biomarker data). Both the unknown equilibrium status and the loss to follow-up [72] may have biased the derivation of the RCT score and evaluation of its performance as a classifier. The biomarker score had limited utility in discriminating between extreme categories of self-reported adherence to the Mediterranean diet; however, the measurement error due to subjective reporting and potentially insufficient construct validity of the self-reported Mediterranean diet score may have meant the C-statistics were underestimated. Sensitivity and specificity of the biomarker score, as well as external validity, remain unknown at present and require evaluation of the score in external trials of interventions with the Mediterranean diet and other dietary patterns. We also note that the MedLey trial used canned fish and legumes as a pragmatic, nontraditional approach to implementation of these components of the Mediterranean diet; however, we would not expect this method of preservation to have a materially differential effect on the biomarkers used compared to cooked fresh fish and beans cooked from dry.

Limitations in the observational research in EPIC-InterAct included residual confounding as a potential source of bias with an overall unclear direction and magnitude. It was unlikely to fully account for the inverse association given its large effect size and the effect sizes of risk factors for T2D previously reported in the literature [73,74]. However, the above mentioned potentially limited specificity may have contributed to positive residual confounding. There may have been degradation of the nutritional biomarkers despite the storage in liquid nitrogen, but we would not expect any bias in the relative risk estimate assuming the degradation rates were non-differential by the incident case status or biomarker level. The random measurement error in the nutritional biomarkers would be expected to bias the association towards the null when considered in isolation from other sources of error, but the context of multivariable statistical modelling precludes any inference on the direction of the potential bias [4]. The differential misclassification in the ascertainment of the outcome also may have biased the results in an unknown direction. We can speculate that higher adherence to the Mediterranean diet may have been associated with greater health-consciousness and healthcare-seeking behaviours, higher likelihood of T2D diagnosis, and an underestimation of the inverse association. We standardised the measures of the Mediterranean diet to improve interpretability of the results for the biomarker score over its original scale and to allow for a comparison of the effect size with that of the self-reported Mediterranean diet. This approach conditioned the effect sizes on the underlying distributions of the exposure variables in EPIC-InterAct which may have limited our quantitative and comparative interpretation [75]. Similarly, our modelling of the PAF was limited by using a relative scale of the biomarker score to quantify its potential impact on population-level prevention of T2D. Public health relevance of this result was further constrained by the fact that the current evidence-base does not enable a direct interpretation of how to achieve the 10-percentile change in the score. Our analysis was not adjusted for use of medications due to incomplete availability of data on these covariates across recruitment centres (<30% participants). However, sensitivity analyses adjusted for use of medications in subsets of study participants showed no appreciable differences compared with the primary results, reducing the potential for residual confounding by medication use. The use of dietary supplements was available as a binary variable that allowed only for a crude assessment of effect modification. Finally, reliability of the externally derived biomarker score applied to biomarker assays used in EPIC-InterAct remains unknown.

Implications of this research

The analysis of dietary patterns aims to evaluate the cumulative impact of dietary exposures on disease risk to inform development of dietary guidelines [1]. High-quality evidence on dietary patterns and the primary prevention of T2D is lacking, partly because of limitations of dietary assessment methods [3]. Self-reported tools have been used almost exclusively in the published studies on this topic [3,76]. We have developed a method of objective assessment of the Mediterranean diet via nutritional biomarkers in order to investigate the association of this dietary pattern and T2D. Our approach of using objective biomarkers yielded substantially greater magnitude and stronger inverse associations than for the Mediterranean diet assessed with subjective methods, also consequently addressing the limitation of small effect sizes often observed in nutritional epidemiology [77]. Our modelling indicates that a modest, 10-percentile increase in the objectively assessed Mediterranean diet could potentially avert 11% of new T2D cases. In comparative analysis, this effect size was approximately 40% of the estimated effect of decreasing mean BMI by 10 percentiles as an example of a well-established, causal risk factor for T2D with a large effect size [74]. Of note, the inverse associations of the biomarker score with T2D in the current study were independent from measured adiposity, with no evidence of effect modification by BMI. Therefore, our finding suggests that a sizeable decrease in the incidence of T2D could potentially be achieved through improved adherence to the Mediterranean diet even in the absence of modification of body weight and across the spectrum of adiposity in the population.

Our integrative approach of using a trial and an observational study with nutritional biomarkers strengthens the evidence for the utility of biomarkers in research on dietary patterns. Further investigations are warranted for a broader use of biomarkers to monitor dietary adherence in research, clinical settings, and potentially of preventive interventions in the real world. Our primary finding was consistent with interventional evidence from a Mediterranean (Spanish) population [65], and it lends itself to consideration for developing dietary guidelines, public health policy, and personalised dietary advice.