An application of partial least squares for identifying dietary patterns in bone health

Osteoporosis is a global health issue that affects millions of individuals around the world, with fragility and fracture resulting from reduced bone mass or increased micro-architectural deterioration of the bone tissue [1]. Nutritional intake has long been considered to have an important role in bone mineral density (BMD) and maintenance through various mechanisms, including alterations in the endocrine system or in bone metabolism and structure [2, 3]. Traditionally, individual foods or nutrients were singled out, and most research has focused on the roles of certain vitamins or minerals. This approach with single nutrients or foods, while valuable, is problematic due to its inherent reductionist nature and lack of dietary context [4, 5].

Dietary patterns provide a more complete perspective to health and disease, as these patterns would more closely mirror what occurs in the real world where individuals eat meals of combined foods and nutrients. The combination of multiple nutrients utilized in a dietary patterns analysis may therefore be more powerful in detecting effects, as the influence of a single nutrient may be too small and the resulting recommendations too difficult to follow [6]. Examining dietary patterns could have a greater public health importance, as evaluating the overall combinations of food intake would be more practical and easily translatable to the public and has been advocated by the US Dietary Guidelines Advisory Committee [7].

The relationship between dietary patterns and bone health has been previously explored through data reduction methods such as dietary indices like the Mediterranean diet [8], factor and principal component analysis (PCA) [9, 10], reduced rank regression (RRR) [11] and cluster analysis [12]. However, though data-reduction methods such as PCA and RRR that are more commonly used, they may not be ideal in selecting dietary patterns in relation to disease because of how they reduce dietary data. PCA derives a set of uncorrelated factors (dietary patterns) characterized by the different foods through explanation of variation in the food intake with no consideration for the outcome measure [4]. RRR is similar to PCA in deriving uncorrelated factors, but has a different goal: dietary patterns are derived to account for the variation not in food intake, but in a set of response variables which are nutrients or biomarkers thought to be important to the health or disease outcome [13].

Partial least squares (PLS) is a mix between PCA and RRR, where extracted dietary patterns account for variation in both dietary intake and the intermediary response variables related to the health or disease outcome [13]. Dietary patterns derived by maximizing the variance in both food groups (the “predictors”) and the outcome-related nutrients or biomarkers (the “response variables”) may be more suitable as it would allow for directed data reduction of food groups through specific nutrients or biomarkers of interest. Common in bioinformatics and chemometrics research, PLS has not been widely utilized in epidemiology, and not in relation with bone health. It is hypothesized to be a more useful method of deriving dietary patterns because it considers the potential biochemical pathways by which the dietary patterns may influence health [14].

Our objective was to identify dietary patterns through the novel use of partial least-squares analysis using nutrients hypothesized to influence bone health as the intermediary response variables. We then assessed and evaluated the relationship between these derived dietary patterns and BMD in a population of women in northeast Scotland.

Participant characteristics are shown in Table 1. Our study participants, on average, were 66 years old (SD 2.2) and overweight, consumed 10.1 MJ/day and had a LS BMD of 1.09 (0.17) and 0.93 g/cm² (0.12) for FN BMD.

Five factors were retained from the PLS analysis. These five factors explained 25% of the variation in the food groups and 77% of the variation in the response variables (Table 2). The main food group loadings for factors 2 and 4 are shown in Fig. 1. Factor loadings, which are the correlations between the derived factors and the food groups, were considered important when values were ≥0.2 using absolute values [13, 23]. A higher factor loading for a food item indicates greater contribution in constructing the factor and is used to interpret the composition of the factor. Factor loadings for factor 2 were shown to be primarily characterized by high intakes of fluid dairy, potatoes, vegetables, fruit and vegetable juices, and wine, and low consumption of processed meats, cheese, cakes, puddings, confectionary, fizzy/carbonated drinks, and spirits. Factor 4 was characterized by high intakes of red and white meats, fruits, and wine, and low intakes of vegetables, sweet spreads, and fruit and vegetable juices.

In multivariable models, two of the factors were significantly associated with LS or FN BMD (Table 3). In adjusted analyses, a 1-unit increase in factor 2 was associated with a 0.012 g/cm² higher LS BMD (95% CI: 0.006, 0.01) and 0.006 g/cm² (95% CI: 0.002, 0.01) higher FN BMD, while a unit increase in factor 4 was associated with a 0.008 g/cm² (95% CI: 0.003, 0.01) increase in FN BMD. When examining the relationship between quartiles of factors 2 and 4 with BMD, LS and FN BMD increased significantly across quartiles of both factor 2 (LS BMD: P-trend = 0.0001; FN BMD: P-trend = 0.02) and factor 4 (LS BMD: P-trend = 0.02; FN BMD: P-trend = 0.006) (Fig. 2).

In this study assessing the use of a novel data-reduction technique to derive dietary patterns in relation to bone health in post-menopausal women, we observed two dietary patterns that were positively associated with LS and FN BMD. These two dietary patterns, or factors, are described by foods that are considered components of a “healthy” diet: intakes of fruits, vegetables, milk and wine, and low intakes of processed meats, cheese, cakes and sweets, fizzy/carbonated beverages and spirits.

Similar to other studies, we found that consumption of alcohol, in the form of wine, was positively associated with BMD, which has been observed in other studies [24]. Moderate alcohol consumption may be beneficial by raising serum estradiol levels [25] and stimulating secretion of calcitonin [26], which has been associated with decreased rate of vertebral fractures and increased bone mass. Wine, rather than spirits, may be valuable to bone health because it is a source of boron, which supports bone growth and maintenance, possibly through enhanced angiogenesis or signal transduction [27, 28].

Interestingly, vegetables and fruit and vegetable juices were negatively correlated with factor 4, one of the two dietary patterns positively associated with BMD. This perhaps suggests that other elements of this dietary factor, such as fruit intake, may have offset the negative loading of vegetables and fruit and vegetable juices to result in the overall beneficial relationship with BMD. A similar explanation was proposed for the loading of a partially hydrogenated soybean oil onto a “vegetable” dietary pattern associated with a decrease in risk of myocardial infarction [29]. The authors hypothesized that the potentially atherogenic effects of partially hydrogenated fatty acids may have been counteracted by the beneficial elements in the vegetable dietary pattern.

Dairy intake also differed in how it loaded on factor 2, which was characterized by high intakes of fluid dairy but low intakes of cheese. Włodarek et al. [30] observed a similar situation, with positive correlation between milk intake and BMD, but negative correlation with rennet and cottage cheeses. Calcium may be the major nutrient attributed to dairy products in relation to bone health but its absorption is promoted, and therefore influenced, by lactose [31]. The authors hypothesized that, as lactose content is higher in milk than in cheeses, the increased calcium absorption led to the positive association with BMD. But, as calcium intake in our population was above the reference nutrient intake (RNI) in the UK of 700 mg/day for adults [20], we postulate that our finding may also be a result of how our food groups were constructed.

We separated dairy products into three groups: milk drinks (including dried, condensed, soy), yogurt and cream products (including full-fat, low-fat, skimmed), and cheese products (including full-fat, low-fat, hard, soft) in order to differentiate between the dairy products. There is no consensus for how to group foods when conducting a dietary pattern analysis; studies subjectively cluster foods based on similarities in food and nutrient intake, culinary preference, and logic [32]. How other studies chose to categorize dairy products may be why positive associations between dairy and bone health were observed when all dairy products were grouped together [33], while inconsistent relationships were found when dairy products were in multiple sub-groups [9].

This is the first study to apply PLS in determining dietary patterns and their relation to bone; there are two other studies applying a statistical method that takes nutrient or biomarker response variables into consideration when identifying dietary patterns, through the use of RRR [11, 34]. In a study of Australian adolescents by van den Hooven et al. [11], patterns were characterized based on whether the response variables (protein, calcium and potassium) were high or low. Two patterns were extracted: pattern 1 with high protein, calcium and potassium, and pattern 2 with high protein, low calcium and potassium. Only the first pattern was positively associated with BMD and bone mineral content; the foods that loaded on this pattern were comparable to a healthy dietary pattern characterized by high intakes of low-fat dairy, whole grains, vegetables, fish, fruit, legumes and low intakes of confectionary, chips/crisps, sweets and processed meats. The study by Ward et al. [34] utilized food diaries collected from age 36 years to 60–64 years to determine RRR dietary patterns constructed with protein, calcium, and potassium as response variables, and how dietary patterns may track through time. The first dietary pattern extracted when participants were 36 years old was positively associated with dietary intakes of fruits and vegetables and low-fat dairy (milk and yogurt), while negatively associated with intakes of refined grain products, processed and sugary foods, and alcohol. Adherence to this dietary pattern at each subsequent visit was calculated for each individual, and the trajectories showed that an adherence to the dietary pattern was positively associated with bone health at age 60–64 years. These two studies utilizing another dietary data reduction technique also showed that a healthy, nutrient-dense diet is beneficial for bone health in both young and older participants.

Other studies deriving non-PLS dietary patterns in older adults also found healthy or nutrient-dense dietary patterns to be associated with decreased bone resorption [10], decreased fracture risk [8, 35] or increased BMD [12], and dietary patterns with high loadings for energy-dense, processed foods were associated with lower BMD [9, 10, 12]. While some studies report null or contradictory relationships between healthy dietary patterns and bone health [36], adherence to dietary patterns which are nutrient-dense are beneficial for health outcomes beyond bone, including hypertension [37] and type 2 diabetes [38].

Strengths of our study include a large sample size and the use of bone-related nutrient biomarkers as response variables in the novel PLS procedure to construct dietary patterns. The benefit of using PLS, as opposed to other data reduction techniques, results from including knowledge about the intermediary response variables between the food groups and the health or disease outcome. Studies differ in what intermediary response variables to include. We included calcium and vitamin D, as they are necessary for calcium absorption and bone formation; in postmenopausal women, calcium intake was positively associated with FN BMD change and hypothesized to reduce bone loss [24] and there is good evidence to suggest that intakes of combined calcium and vitamin D are beneficial for BMD [39]. Vitamin C is an antioxidant that could reduce bone loss by counteracting oxidative stress that may reduce BMD, and dietary intakes have been associated with increased BMD [40]. Protein is necessary for many bone-related activities including growth factors and hormones that impact bone synthesis, break-down, and bone matrix structure [41]. Moderate alcohol intake has been positively associated with BMD and less bone loss by promoting secretion of calcitonin or increasing endogenous oestrogens [24]. Phosphorus is necessary for mineralization of the skeleton and inadequate levels result in impaired bone integrity and can lead to osteomalacia [42]. Magnesium may influence bone metabolism through its necessity as a co-factor in metabolism and enzymatic processes and, directly, it may decrease hydroxyapatite crystal size [43]. Zinc is necessary for collagen synthesis and osteoblastic activity, and a trial among postmenopausal women showed that supplementation resulted in a small increase in BMD over a 2-year period [43, 44]. Finally, potassium is hypothesized to benefit bone by producing an alkaline environment, reducing the need to recruit skeletal calcium salts to counteract the acids generated from acid-generating foods [3, 43].

The PLS approach in including response variables results in dietary factors that are underpinned by the underlying nutrients of interest, so the choice of the intermediary response variables will influence how dietary factors are constructed. While RRR is similar to PLS, dietary patterns identified by RRR are limited to those nutrients included as the intermediary variables and may miss out on dietary patterns that specify nutrient pathways not included as intermediary response variables. Therefore, the dietary patterns produced by RRR are more tailored to the health outcome in question, whereas the PLS procedure is a better tool for informing how to modify intake of dietary intake to elicit a nutrient response and impact health outcomes. While Hoffman et al. [29] concluded that RRR was better at identifying dietary patterns in relation to diabetes risk, DiBello et al. [29] found that PCA and PLS produced more dietary patterns associated with their outcome of myocardial infarction risk. Thus, our objective was methodologically driven by the use of this novel technique to assess how dietary patterns may be derived. The dietary patterns we identified support previous findings using other dietary reduction techniques, and this present analysis using a different data reduction technique confirms and adds to the evidence of the benefits of a healthy dietary pattern.

This study also had several limitations. Like other data-reduction methods, the assessment of dietary intake is subject to measurement error and residual confounding, even though our FFQ had been previously validated using weighed food records and biochemical markers of antioxidant status [5]. It is also possible that the dietary patterns observed were based on nutrients that were chosen as our response variable nutrients, where different nutrient response variables may elicit formation of different dietary patterns. Additionally, dietary patterns extracted in one population with data-reduction methods will not reproduce the same dietary factors with the same food loadings in another data set, and thus, our results are limited to postmenopausal women from northeast Scotland. Another limitation is the cross-sectional nature of our study, where causality cannot be assumed.

Nevertheless, multiple studies have reported general “healthy” dietary patterns that have similar food group loadings. The persistent relationships between similar dietary patterns with various outcome measures have led government advisory groups such as the US Dietary Guidelines Advisory Committee to recommend that the population achieve optimal health through a healthy diet, rather than focusing on specific foods or nutrients [7]. Earlier initiatives had also attempted to move from nutrient population goals to one that is foods-based [45].

In summary, dietary pattern analysis using the PLS method extracted dietary patterns rooted in a biological link between dietary intakes and the health outcome; this is important because foods are the modifiable aspect in this relationship. In this study, we observed that adherence to dietary patterns with higher intakes of nutrient-rich foods and lower intakes of energy-dense foods was associated with an increase in bone mineral density in a population of postmenopausal women. Our results show that PLS is an appropriate method to determine which dietary patterns are associated with BMD and supports previous findings using other data-reduction techniques on the relationships between diet and health.