Obesity and associated chronic cardiometabolic health conditions are major public health concerns. In the UK, 26% of adults are classified as obese [1
] and nearly a quarter of all deaths are caused by cardiovascular diseases [2
]. In the US, cardiovascular disease is the leading cause of death and its prevalence is expected to increase to 45% of adults in the next 15 years [3
]. Diet is an established modifiable risk factor for chronic diseases. Decades of nutrition research have identified numerous dietary factors as relevant to human health, and numerous clinical trials have targeted such factors with the goal of preventing and treating chronic diseases, including obesity, metabolic disorders, and cardiovascular disease.
Nutrition research has traditionally focused on single nutrients, overall diet quality, and/or composition and their associations on cardiometabolic health. For example, Key et al. [4
] examined the risk of ischemic heart disease associated with a range protein sources, including eggs, meat, dairy, and fish in more than 400,000 individuals participating in the European Prospective Investigation into Cancer and Nutrition (EPIC) cohort study. They observed an increased risk of heart disease for individuals reporting higher processed and red meat consumption, but not fish, poultry, or milk. Another prospective UK cohort study examined breakfast energy intake (estimated from a 7-day food diary) and showed that individuals with the highest breakfast energy intake levels had the lowest body mass index (BMI), despite higher overall energy intake, and gained less weight over a two-year follow-up period [5
]. During the last decade, nutritional epidemiology has also moved towards the study of dietary patterns and combinations of different nutrients. For example, in a prospective analysis of 75,020 women in the Nurses’ Health Study and 42,865 men of the Health Professionals Follow-Up Study, AlEssa et al. [6
] found that higher carbohydrate to cereal fiber ratio as well as higher starch to cereal fiber ratio are associated with an higher risk of coronary heart disease. Their findings highlight the importance of studying nutrient intakes in tandem as these may provide more information on the quality of food consumed.
More recently, timing, frequency, and regularity of food intake have emerged as novel risk factors for cardiometabolic health [7
]. For example, Ma et al. [12
] conducted a cross-sectional analysis in 499 participants within the SEASONS (Seasonal Variation of Blood Cholesterol) Study using 24-h dietary recalls and body weight measurements. Their results indicated that those individuals reporting ≥ 4 eating occasions per day had a 45% lower risk of obesity compared to individuals that ate less frequently (≤ 3 times per day, OR 0.55, 95%CI = 0.33–0.91), independent of daily energy intake. Pot et al. [13
] have proposed a novel method of assessing the inconsistency of day to day energy intake. In their study of 1768 individuals who completed 5-day food diaries, they quantified energy intake during predefined meals (breakfast, lunch, and dinner), between meal intake, and daily intakes compared to a 5-day mean energy intake. The results showed that higher irregularity scores of energy intake, particularly for breakfast and between meal intake, were associated with an increased risk of metabolic syndrome (OR 1.34, 95% CI = 0.99–1.81 and OR 1.36, 95% CI = 1.01–1.85, respectively). The findings of this study provide compelling evidence to consider regularity of energy intake when examining eating patterns. However, to this date it is unclear how regularity of additional nutrients (proteins, carbohydrates, and fats) may affect cardiometabolic health and how regularity may interact with the other dimensions of eating patterns.
From this body of evidence, individual dietary intake dimensions (composition, frequency, regularity) are novel predictors of cardiometabolic disease risk, but the concurrent study of multiple of these dimensions is so far under-developed. Understanding diet exposures that need to be considered in relation to cardiometabolic disease could provide critical insights to novel design strategies for targeted dietary interventions to address the global obesity and cardiovascular disease pandemics. It could also help address the challenge of multiple comparisons, which has been raised in the past [14
]. The goal of this study was to undertake a data-driven approach to identify non-redundant, minimally collinear diet exposures from 7-day food diaries, which we then used to probe associations with cardiometabolic risk (CMR) in the Airwave Health Monitoring Study (AHMS). This exploratory, proof-of-principle study was designed to evaluate the potential of data-driven methods to address challenges that arise when modeling large numbers of potentially highly correlated measures.
In this study we systematically analyzed the association between non-redundant diet exposures and CMR in the AHMS, a large, on-going cohort study. We implemented dimension reduction techniques to select diet exposure variables among a large set of a priori defined candidate exposures, expected to be relevant for cardiometabolic health. Dimension reduction techniques enabled a data-driven selection of exposures among often highly correlated variables; we then used the diet exposures that were identified as the most representative of overall diet intake patterns. Out of 45 candidate variables derived from 7-day food diaries, seven diet exposures were ultimately identified as non-redundant diet exposures, of which energy-adjusted carbohydrate intake, fiber intake, and meal frequency were associated with CMR.
We found higher energy-adjusted intake of carbohydrates to be associated with lower CMR. Findings from previous studies also examining this association are somewhat mixed [35
]. For example, in a two-year study of 322 obese individuals assigned to a low fat, low carbohydrate or Mediterranean diet, the individuals on a low carbohydrate diet lost more weight and had improved lipids when compared to the those on a low-fat diet [36
]. However, the same study also observed benefits in individuals assigned to a Mediterranean diet, typically higher in carbohydrate intake, with individuals seeing increased weight loss and improved insulin and fasting glucose levels on this diet. These findings would rather suggest that diets, such as the Mediterranean Diet, which include higher intake of high quality carbohydrates, such as whole grains and increased fiber intake, may be beneficial for cardiometabolic disease risk factors [37
]. We also found higher carbohydrate intake to be associated with higher risk of impaired glucose control, arguably the most important risk factor for CMR. These findings are consistent with a recent meta-analysis of 10 different trials that found a 34% reduction in HbA1c in individuals after 1 year on a low carbohydrate diet [40
]. However, the link between higher carbohydrate intake and higher glucose levels is more established for carbohydrates with a higher glycemic load (i.e., low-quality carbohydrates) [41
]. We also observed that individuals with the highest fiber intake (on average 13.30 g/1000 kcal) had a lower likelihood of CMR, and upon adjustment for fiber intake the protective effect of carbohydrate intake on CMR was somewhat attenuated. This suggests that fiber intake may be a potential mediating factor in the association between carbohydrate intake and CMR. Therefore, further analysis of the quality of carbohydrates captured by the carbohydrate dimension we analyzed would be of future interest.
Another possible explanation for our findings is that those individuals with the lowest carbohydrate intake (consisting, on average, of 39% of daily energy intake) may be intentionally trying to reduce carbohydrate intake because they are aware of their CMR status. Indeed, the average carbohydrate intake in the lowest quartile falls below CDC recommendations for individuals with diabetes (45%) [42
] and within the range (10–40%) some studies have shown to be beneficial for metabolic syndrome [43
]. Additionally, Seidelmann et al. [46
] have suggested a U-shaped relationship for carbohydrate intake with minimal risk for all-cause mortality observed when carbohydrate intake constituted 50–55% of energy intake. In our study, individuals with the highest carbohydrate intakes—consisting on average of 56%—fall just above this range. In fact this value falls within the range for recommended carbohydrate intake (45–65%), so these highest carbohydrate eaters may be those that are following dietary guidelines [47
]. Our findings do not allow us to disentangle effects of disease status on diet from those of diet on metabolic disease status, given our cross-sectional study design, warranting prospective studies to address this question further.
Associations between meal frequency and CMR are similarly complex. Many previous studies have found that increased meal frequency is associated with increased energy intake [48
]. We observe in our study that those individuals that eat most frequently (>5 meals/day) also have the greatest energy intake. However, these individuals that ate most frequently also had the lowest energy intake, on average, per eating occasion. Eating more frequent, less caloric meals therefore may be beneficial to CMR. This has also been shown in a study conducted within the Norfolk cohort of EPIC (N
= 14,666), where greater meal frequency (>6 meals/day) improved LDL and cholesterol levels [52
]. Another study of 2696 individuals of the INTERnational study on Macro/micronutrients and blood pressure (INTERMAP) found increased eating frequency (>6 meals/day) to be associated with improved diet quality and lower BMI [53
]. Despite this, the alternative explanation of reverse causation could again be true in this case: those individuals that are eating fewer meals per day are aware of their risk status and therefore consciously trying to eat less overall.
Our data-driven approach should be considered hypothesis-generating and provides novel paths towards addressing the multi-dimensionality in exposure information. It is noteworthy that dimension reduction method results are dependent on the underlying data structure and may be of limited generalizability. Data formats, study sample size, and interpretability will have to be considered when choosing specific dimension reduction methods, including k-means clustering, for future research studies. Our findings highlight the benefits of such dimension reduction approaches when handling high-dimensional data, including nutritional exposure data. The application to nutrient intake patterns is novel and future studies embedded in distinct cohort studies and populations will be useful to delineate generalizability and uncover differences in dietary intake patterns. In addition, future studies are needed to understand the usefulness of our approach in guiding prevention and intervention strategies. Our data clearly highlight the potential of systematically considering all potentially relevant exposures, as well as the need to consider diet exposure dimensions like frequency, temporality, and irregularity in dietary intake. Further, these data-driven approaches may pave the way for future studies that would systematically expand dietary factors to all nutrients, including vitamins, minerals, and other dimensions of dietary intake, including diet quality. Our data-driven approach allows to include multiple dietary exposures in analyses concurrently and may help to understand independent contributions to disease risk among a host of candidate variables. This in turn may help inform dietary intervention strategies.
Our study has several limitations. First, participants self-reported dietary intake. Bias in reporting is common for all dietary recording methods and in this AHMS cohort, there is a strong and significant bias towards under-reporting in individuals with a higher BMI [18
]. However, our primary results were not attenuated upon restriction to those individuals with higher BMI. Further, dietary recording was limited to 7 days and may not reflect habitual intake. The balance of data collection and increased participant burden needs to be considered in collection periods greater than 7 days. Worth noting, however, is that the 7-day diary captures weekday/weekend variation in intake and includes information on other aspects of diet. An additional limitation of this present study includes limited reports for shift work timing. Shift work can greatly influence diurnal eating patterns, particularly the temporal distribution of certain macronutrient intake [54
]. However, preliminary findings indicate that in the AHMS cohort, duration of weekly working hours (which we have adjusted for in this present study) has been found to be associated with shift work, where those individuals with the highest weekly working duration are the most likely to participate in shift work. Finally, we did not have precise timing information for eating occasions. However, we were able to approximate temporal distribution of intake by using the pre-specified categories for eating occasions included in the food diaries. We also derived one variable, frequency of having both a late night and early morning snack, in order to estimate fasting period, although this did not emerge as a relevant dimension in our analysis. This allows for the possibility for future analyses of diurnal eating patterns and cardiometabolic health, as the temporal aspect of eating patterns is becoming an established risk factor for metabolic diseases [27
The present study also has several strengths. The use of 7-day food diaries enabled for a detailed weekly dietary intake to be measured. Using this high-quality dietary intake data, we were able to rigorously and systematically analyze multiple aspects of diet exposures using dimension reduction techniques to prioritize non-redundant diet exposures that may most strongly influence CMR. Our approach demonstrates the power of data-driven approaches for diet exposure selections when a host of correlated diet exposures are hypothesized to influence cardiometabolic health. Similar techniques have been used in previous studies, where dimension reduction techniques are used to cluster individuals based upon their dietary habits [25
] to see how they may be differentially affected by chronic disease risk. Our approach takes a multi-dimensional approach by allowing us to quantify several dimensions of dietary intake that include temporality, regularity, and macronutrient composition. Understanding which of these dimensions may represent independent clusters of dietary intakes may then allow for multi-target interventions of specific dietary aspects.