Dietary pattern research

A whole-diet approach is an important tool in nutrition research to capture the inherently complex nature of food intake. Energy intake in human subjects has been shown to be relatively stable, and this implies that changes in one food or nutrient will be associated with changes in another one, also known as substitution effects^{(Reference Willett13,Reference Willett, Howe and Kushi14)} . Therefore, making inferences about a particular food or nutrient in observational studies is problematic and residual confounding will likely affect the observed associations with health outcomes. DP approaches can potentially overcome these issues, while capturing any synergistic effects and the cumulative exposure to different foods and nutrients^{(Reference Schulze and Hoffmann9)}, which in turn may result in stronger associations with health outcomes than what has been observed with single nutrients or foods^{(Reference Zhao, Li and Gao8,Reference Hoffmann, Schulze and Schienkiewitz10)} .

In observational studies, there are two main approaches used to derive DPs using self-reported measures of dietary intakes: a priori or pre-defined DP, and a posteriori or data-driven/exploratory DP^{(Reference Zhao, Li and Gao8,Reference Schulze, Martínez-González and Fung15)} . Pre-defined DP approaches include specific diets (i.e. vegetarian or vegan) as well as dietary indices or scores, which are used as a proxy for adherence to a particular diet (i.e. Mediterranean dietary score, healthy eating index). Exploratory approaches are data-driven because they use different statistical techniques to derive particular combinations of food groups from dietary data collected in a specific population. The most widely used exploratory approaches include factor and principal component analysis, partial least squares and RRR.

Both a priori and a posteriori approaches aim to reduce a large amount of information from complex diets in order to investigate diet–disease associations, but all these methods have particular pros and cons that have been summarised elsewhere^{(Reference Zhao, Li and Gao8,Reference Schulze and Hoffmann9,Reference Schulze, Martínez-González and Fung15)} . In any case, it is critical to choose the most appropriate approach that aligns with a well-developed research question, as this can influence the observed associations with disease risk^{(Reference Zhao, Li and Gao8,Reference Hoffmann, Schulze and Schienkiewitz10,Reference Hoffmann, Boeing and Boffetta16)} .

Dietary patterns derived by reduced rank regression

RRR has emerged as a useful hybrid approach which uses a priori knowledge about nutrients or biomarkers with an established link with disease, to help identify data-driven combinations of foods in a particular population^{(Reference Hoffmann, Schulze and Schienkiewitz10)}. In common with all the other exploratory approaches, the RRR method aims to extract successive linear combinations of predictor variables (i.e. food groups). However, the main goal of RRR is to explain the maximum variability in response variables (i.e. nutrients of interest with established associations with disease risk); while the other exploratory approaches aim to maximise the variability in the predictors (principal component analysis) or a balance between predictor and response variation (partial least squares)^{(Reference Hoffmann, Schulze and Schienkiewitz10)}.

The RRR model can include one or more response variables, and a number of uncorrelated factors (DPs) will be derived which is equal to the number of specified response variables in the model. The variation in each of the response variables that is explained by each DP gives an indication of the likely ‘strength’ between each DP and each response variable. Typically, the DPs that explain the maximum variability in all the response variables collectively are retained for analysis. Each predictor variable in the model (i.e. food group) will also receive a factor loading, so that a larger value (positive or negative) indicates that the food group makes a greater contribution to the DP. Participants in a study population can then obtain an overall score (z score) for each of the derived DPs based on the factor loading and their intake of each food group, representing the degree to which their dietary intake adheres to a DP relative to other participants.

A key strength of the RRR approach is that the derived DPs use a priori knowledge of established associations between biomarkers or nutrients of interest and disease, hence it is particularly appealing if the aim is to identify DPs in association with disease endpoints. The emerging DPs are food-based, meaning they are easier to interpret and communicate, which helps inform public health advice. However, RRR-derived DPs share limitations in common with the rest of approaches, for example, subjective decisions are made about the number of DPs that will be retained for analysis, and it is unclear which food group(s) may drive the associations. In particular, RRR relies on the availability of response variables with solid evidence of their associations with disease outcomes; and by choosing a specific set of nutrient response variables, there will be some variability explained by other nutrients which may also be involved in the disease pathway of interest^{(Reference Hoffmann, Schulze and Schienkiewitz10,Reference Schulze, Martínez-González and Fung15)} .

Dietary patterns emerging from longitudinal cohort studies

The state of the evidence with regards to RRR-derived DPs and non-communicable diseases such as CVD and diabetes is gradually growing. A literature search on PubMed and Embase up to May 2022 returned seventeen studies using data from longitudinal cohorts looking at prospective associations between RRR-derived DPs with CVD risk and/or all-cause mortality^{(Reference Gao, Jebb and Aveyard11,Reference Hoffmann, Boeing and Boffetta16–Reference Nazari, Mokhayeri and Mansournia31)} , and nineteen studies investigating associations with the risk of developing diabetes^{(Reference Gao, Jebb and Aveyard12,Reference Batis, Mendez and Gordon-Larsen32–Reference Song and Lee49)} . The vast majority of these studies have been published since 2015, and have included samples from a variety of populations including the USA (n 8^{(Reference Héroux, Janssen and Lam20,Reference Li, Chen and Zhang24,Reference Mazidi, Katsiki and Mikhailidis25,Reference Nazari, Mokhayeri and Mansournia31,Reference Jacobs, Kroeger and Schulze39,Reference Johnson, Vanderlinden and DeFelice40,Reference Liese, Weis and Schulz42,Reference Schulze, Hoffmann and Manson45,Reference Shin, Lee and Song46)} ); Europe (n 11^{(Reference Hoffmann, Boeing and Boffetta16,Reference Biesbroek, Van Der A and Brosens18,Reference Cai, Osté and Gomes-Neto19,Reference Johns, Lindroos and Jebb21,Reference Leermakers, van den Hooven and Franco23,Reference Meyer, Döring and Herder26,Reference Drogan, Hoffmann and Schulz29,Reference Drake, Sonestedt and Ericson35–Reference Heidemann, Hoffmann and Spranger37,Reference Sluijs, Beulens and Van Der47,48)} ), Asia (n 6^{(Reference Lee, An and Lee22,Reference Seah, Ong and Koh27,Reference Shi and Ganji28,Reference Batis, Mendez and Gordon-Larsen32,Reference Batis, Mendez and Sotres-Alvarez33,Reference Lee, Son and Lee41,Reference Song and Lee49)} ), UK (n 4^{(Reference Gao, Jebb and Aveyard11,Reference Gao, Jebb and Aveyard12,Reference McNaughton, Mishra and Brunner30,Reference Brayner, Kaur and Keske34,Reference McNaughton, Mishra and Brunner43,Reference Pastorino, Richards and Pierce44)} ) or others (n 2^{(Reference Appannah, Pot and Huang17,Reference Imamura, Lichtenstein and Dallal38)} ). Population specificity is important here, because even when the choice of response variables (e.g. nutrients) is the same, the emerging DPs can vary substantially between different populations, owing to cultural differences in food intakes.

Until very recently, evidence with regards to RRR-derived DPs in association with CVD and all-cause mortality risk in the UK population was limited to one small study^{(Reference McNaughton, Mishra and Brunner30)}, while the evidence with regards to diabetes was constrained to two small cohort studies^{(Reference McNaughton, Mishra and Brunner43,Reference Pastorino, Richards and Pierce44)} . Here we focus on the use of the UK Biobank study to illustrate the application of the RRR approach to derive DPs and summarise the prospective associations between RRR-derived DPs and CVD^{(Reference Gao, Jebb and Aveyard11)}, type 2 diabetes^{(Reference Gao, Jebb and Aveyard12)} and all-cause mortality^{(Reference Gao, Jebb and Aveyard11)} in the context of previously conducted studies. We have also investigated associations with cardiometabolic risk factors (e.g. BMI, blood pressure or lipid biomarkers) to help understand potentially underlying mechanisms. The UK Biobank is a prospective cohort that recruited over 500 000 adults between 2006 and 2010, aged 40–69 years at baseline. Together with important information on demographic, lifestyle factors, medical history, physical measures (height and weight) and blood and urine samples, detailed dietary intake measures were collected through an online questionnaire (Oxford WebQ) administered via email on up to five times during approximately 1⋅5–2 years^{(Reference Collins50,Reference Sudlow, Gallacher and Allen51)} . The WebQ is a validated 24-h dietary assessment tool which collected quantitative information on two hundred and six widely consumed foods and thirty-two beverages^{(Reference Galante, Adamska and Young52–Reference Liu, Young and Crowe54)}. We developed a food grouping system to group the number of reported foods and beverages into fifty major food groups before including them into the RRR model^{(Reference Piernas, Perez-Cornago and Gao55,Reference Perez-Cornago, Pollard and Young56)} , as this would also facilitate obtaining meaningful and relevant DPs. Four nutrient response variables were included: energy density (kJ/g), saturated fat (% total energy), free sugars (% total energy) and fibre density (g/MJ). These nutrient variables have established associations with the development of obesity, type 2 diabetes and CVD^{(Reference Macdonald57–Reference Threapleton, Greenwood and Evans66)}, which underscores the importance of a well-supported hypothesis underlying the proposed research. From the initial baseline cohort, a subsample of participants that provided at least two 24-h questionnaires (n ~117 000^{(Reference Gao, Jebb and Aveyard11)}; n ~120 000^{(Reference Gao, Jebb and Aveyard12)}) and free of disease at baseline, were included and the mean dietary intake across multiple assessments was calculated to derive DPs.

With four response variables, the RRR model extracted four DPs together explaining a total of 77 % of variability in the response variables. The top two DPs explaining 43 % (DP1) and 20 % (DP2) of variability were retained for analyses. DP1 showed high factor loadings for chocolate and confectionery, butter and other spreads, low-fibre bread, table sugars, grain-based desserts, sugar-sweetened beverages (SSBs) among others; and low factor loadings for fresh fruit, vegetables and high-fibre breakfast cereals. DP1 showed strong positive correlations with energy density, free sugars and saturated fat, but negative correlations with fibre. Conversely, DP2 showed high factor loadings for SSBs, fruit juice, table sugars and preserves, chocolate confectionery; and low factor loadings for high-fat cheese, butter and other animal fat spreads. DP2 showed a strong positive correlation with free sugars, but a negative correlation with saturated fat, while energy density or fibre were not correlated^{(Reference Gao, Jebb and Aveyard11,Reference Gao, Jebb and Aveyard12)} .

There is wide heterogeneity in previous studies regarding the choice of response variables in the RRR model, and most studies included combinations of nutrients^{(Reference Hoffmann, Boeing and Boffetta16,Reference Johns, Lindroos and Jebb21,Reference Brayner, Kaur and Keske34,Reference Pastorino, Richards and Pierce44,Reference Huisman, Soedamah-Muthu and Vermeulen67)} (% energy from monounsaturated fats and polyunsaturated fats; protein and total carbohydrates), as well as biomarkers^{(Reference Mazidi, Katsiki and Mikhailidis25,Reference Seah, Ong and Koh27,Reference Batis, Mendez and Gordon-Larsen32)} or disease risk factors (BMI, waist circumference, blood lipids or blood pressure)^{(Reference McNaughton, Mishra and Brunner30,Reference Drake, Sonestedt and Ericson35,Reference Imamura, Lichtenstein and Dallal38,Reference Song and Lee49)} . The majority of these previous studies obtained dietary measurements from food frequency questionnaires with fewer using either 24 h recalls or food records/diaries. The resulting DPs yielded combinations of food groups which were population-specific, although most studies identified at least one unhealthy or ‘Western’ DP characterised by high intakes of soft drinks, alcohol, refined carbohydrates, added sugars, fats, meat (including processed) and low fruit/vegetable intakes.

Associations with health outcomes

CVD

Previous research studies have derived CVD-related DPs and their associations with several outcomes, mostly using response variables such CVD biomarkers or risk factors^{(Reference Biesbroek, Van Der A and Brosens18,Reference Héroux, Janssen and Lam20,Reference Meyer, Döring and Herder26,Reference Drogan, Hoffmann and Schulz29–Reference Nazari, Mokhayeri and Mansournia31,Reference Drake, Sonestedt and Ericson35,Reference Lamichhane, Liese and Urbina68)} as oppose to nutrients^{(Reference Gao, Jebb and Aveyard11,Reference Hoffmann, Boeing and Boffetta16,Reference Johns, Lindroos and Jebb21,Reference Mazidi, Katsiki and Mikhailidis25)} . In common with the DPs derived in the UK Biobank population, most of these previous DPs were characterised by intakes of high-fat/high-sugar, refined carbohydrates (i.e. low-fibre bread, table sugars, SSBs) and high saturated fat foods (i.e. butter and animal spreads, red and processed meats), but low intakes of fruit, vegetables or higher-fibre foods. Of the studies performed in US cohorts, one reported significant associations between DPs characterised by high meat and refined carbohydrates/potato products but low fruit intakes with CVD mortality^{(Reference Héroux, Janssen and Lam20)}; and another one reported a DP characterised by high intakes of fats, SSBs, meat, but low vegetables in association with a higher incidence of stroke^{(Reference Nazari, Mokhayeri and Mansournia31)}. In European populations, a study in Germany reported significant inverse associations with CVD mortality among those following a DP high in whole grain, fruits, cereals and vegetables but low intakes of processed meat, butter and cheese^{(Reference Drogan, Hoffmann and Schulz29)}. Another study among Swedish participants with obesity reported a significant association with total CVD in people following an energy-dense, high SFA, low-fibre DP^{(Reference Johns, Lindroos and Jebb21)}. A study in the UK Whitehall II study reported significant associations with incident CHD in association with a DP characterised by high intakes of white bread, fried potatoes, added sugars, processed meat and SSBs but low intakes of vegetables^{(Reference McNaughton, Mishra and Brunner30)}.

The associations between DPs obtained in the UK Biobank population and several health outcomes are summarised in Fig. 1^{(Reference Gao, Jebb and Aveyard11)}. People in the highest quintile of adherence to DP1 showed a 40 % higher risk of total cardiovascular events and 29 % higher risk of fatal CVD events compared to people in the first quintile. DP2 showed non-linear associations, although there were 14 % higher risk of total CVD and 18 % higher risk of fatal CVD among people in the highest quintile of DP2 compared to those in the lowest quintile. Adjustment for the BMI revealed slightly attenuated associations only for the associations observed with DP1, suggesting that the observed association was potentially mediated through excess weight. These observations are broadly consistent with previous evidence from single-nutrient studies or other DP studies using dietary quality scores^{(Reference Yang, Zhang and Gregg69–Reference Schwingshackl, Bogensberger and Hoffmann72)}.

Fig. 1. Risk of health outcomes associated with dietary pattern 1 (DP1) and dietary pattern 2 (DP2) in the UK Biobank study. Adjusted hazard ratios (HR) and 95 % CI, showing the risk of each DP for people in the fifth quintile (indicating the highest adherence to each DP). Estimates were obtained from Cox proportional hazard models using age during study as the underlying timescale. The models were stratified by sex and regions (England, Scotland and Wales) and adjusted for socio-demographic characteristics: ethnicity (white, others, missing), Townsend index of deprivation (quintiles one to five, with higher scores representing greater deprivation), education (higher degree [college or university degree, or professional qualifications], any school degree [A levels, AS levels, O levels, General Certificate of Secondary Educations or Certificate of Secondary Educations], vocational qualifications [National Vocational Qualifications, Higher National Diploma or Higher National Certificate], other [none of the above qualifications], missing); behavioural risk factors: smoking status (never, current, previous, missing), physical activity (low [<600 metabolic equivalent (MET)-minutes per week], moderate [≥600 and <3000 metabolic equivalent (MET)-minutes per week], high [≥3000 metabolic equivalent (MET)-minutes per week], missing), log-transformed energy intake; and health history/conditions: family history of diabetes, menopause in women, hypertension, CVD, high cholesterol.

Strengths and limitations

The evidence summarised here comes from longitudinal cohort studies, with wide variability in population characteristics and sample sizes. The UK Biobank is one of the largest contemporary cohorts of UK adults who provided detailed measures of dietary intake, which is essential to investigate the associations between food-based DPs and long-term disease risk. The follow-up times since completion of the dietary questionnaires (from 5–8 years) was sufficient to accumulate enough cases of CVD, diabetes and all-cause mortality and to achieve more precise estimates of association. Linkages with electronic health records also allowed for an objective ascertainment of cases.

However, previous research as well as recent studies by the authors in the UK Biobank share common limitations that are inherent to observational studies. The derived DPs are based on self-reported measures of dietary intakes, which can be over- and/or underreported by participants introducing biases, although people with implausible dietary intakes were excluded in the studies by the authors. Day-to-day variation in diet was partly accounted for by including at least two dietary questionnaires, but some degree of bias due to random variation will remain^{(Reference Carter, Lewington and Piernas84)}. In terms of investigating the associations between dietary exposures and longer-term health, it is clear that a lifetime exposure to a poor diet is a major determinant. However, most studies collected measurements of diet over a short time-period before disease manifestation and this may only reflect the most recent portion of this exposure. It is also unknown if DPs are stable over time or if changes preceding disease may affect the results. As with any observational research, caution is needed to avoid any causal inference and residual confounding (even after careful adjustment is made) cannot be ruled out. The cross-sectional associations investigating potential mechanisms through the associations with cardiometabolic risk factors measured at baseline may be affected by reversed causality, therefore careful interpretation is warranted. Finally, DPs identified in each study are culturally specific and relevant to each population, hence generalisability to other populations or countries is limited.

The main advantage of using RRR over other DP methods is that it allows an investigation of the cumulative effects of the whole diet, hence reducing confounding due to other dietary aspects, while also incorporating the mechanistic evidence through the integration of nutrient response variables^{(Reference Hoffmann, Schulze and Schienkiewitz10)}. The interpretation of the results is based on foods rather than nutrients, providing a more specific evidence base to inform the development of food-based dietary guidelines and making it easier to develop public health nutrition policies. However, the RRR method hinges on an imperfect evidence base to identify the underlying response variables and their association with disease endpoints. By focusing on a limited set of nutrients and foods which usually cannot explain 100 % of the nutrient response variance, it is possible that the emerging DPs contain other important nutrients which are involved in the biological pathway to disease but cannot be specified. For most chronic diseases there might be complex interactions in different metabolic pathways linking diet and disease, which may not be captured in a single RRR model with a limited number of response variables.

Implications and conclusions

In the context of previous evidence, recent DP research by the authors in the UK Biobank study supports current dietary recommendations for good health which emphasise reductions in saturated fat and free sugars, and increases in vegetables, fruits or grains, among others^(6,7) . But this new evidence is food-based, and points towards diets which are lower in chocolate and confectionery, butter, low-fibre bread, added sugars and SSBs, and higher in fresh fruit and vegetables and whole grain foods to reduce the risk of CVD, diabetes and limit premature mortality. This should also help avoid confusing messages to the public regarding specific nutrients (e.g. saturated fat and sugars), recognising that foods are mixtures of nutrients and there might be synergistic effects between them. By identifying patterns in population-specific cohorts the hope is that food-based dietary advice may help accelerate behaviour change because it helps target specific foods and their combinations which are culturally relevant.

In conclusion, a whole-diet approach to disease prevention is attracting increasing interest from researchers, practitioners and policymakers. RRR is a particularly useful tool to derive population-specific food-based DPs using a priori knowledge about biomarkers or nutrients of concern which mediate the associations between the emerging DPs and disease outcomes and to help provide culturally relevant information to underpin behaviour change strategies to reduce non-communicable diseases.