Application of a new dietary pattern analysis method in nutritional epidemiology

Subjects in this study were participants in the National Health and Nutrition Examination Survey (NHANES) from 2005 to 2006. As part of the nationwide survey, which involves about 5000 persons each year, participants completed the Household Adult Questionnaire and the Dietary Food Frequency Questionnaire. Additionally, participants were invited to a mobile examination center, where blood and urine specimens were obtained, and a number of tests and measurements—including body measurements and blood pressure testing—were performed [23].

A number of restrictive criteria were applied to the data. Participants in our final data set were at least 20 years old; were not pregnant or lactating at the time of data collection; were not currently suffering from heart disease; and were not diabetic. Moreover, our data set included only participants who had completed the Food Frequency Questionnaire. The resulting sample included 2609 participants with a mean age of 49.56 and a mean BMI of 28.45 kg/m². 54% of the participants in our sample were female. Laboratory measures of biomarkers for cardiovascular disease (triglycerides, LDL cholesterol, HDL cholesterol, and total cholesterol) were obtained from NHANES.

Participants in the NHANES study completed a full Food Frequency Questionnaire (FFQ), which collects quantitative information about how many times foods had been consumed in the past month. FFQ has been validated to be a reasonable tool to assess relationships between dietary habits and cardiovascular risk factors [6, 24]. Following the recommendation in the literature [6], the individual food items and beverages measured in FFQ were combined into 35 food groups that represented all major food groups consumed by the US population. FFQ raw scores were available from NHANES website.

The NHANES dataset also includes significant laboratory test results. Among these were blood tests, which were conducted via venipuncture, with necessary precautions taken to ensure participants could safely complete the test. Blood levels of triglycerides, LDL cholesterol, HDL cholesterol, and total cholesterol were available from NHANES.

PCA yielded 10 principal components (PCs) that together accounted for 65% of the variation in the data set. The top 10 PCs were selected based on the scree plot, eigenvalues, and parallel analysis, as well as consideration of content and theoretical consistency. To interpret the dietary patterns represented by these PCs, we focused on the top and bottom 10th percentile of the PC loadings for each food category (Table 1). Both the magnitude and direction of the loadings were interpreted. For example, the first principal component (PC1) has high positive loadings of organ meats (e.g., liver and other organ meats), fish, chili peppers, and gravy, indicating that participants who followed the dietary pattern represented by PC1 typically consumed greater amounts of these foods. The second principal component (PC2) represents a healthy dietary pattern with high consumption of fruits and vegetables (green, dark-yellow, and other) and low consumption of liquor, soft drinks, beer, and gravy. In contrast, the third principal component (PC3) characterizes a relatively unhealthy dietary pattern with high consumption of high-fat dairy, red meat, refined grains, sweets and low consumption of meal replacements, organ meats, liquor, and wine. Additionally, PC4 identifies a large number of liquids with positive loadings of juice, fruit juices, and soft drinks as well as negative loadings of coffee and tea. PC4 also has large negative loadings of non-dairy cream and sweets. PC5 has high positive loadings of vegetables, salad dressing, and beer as well as large negative loadings of low-fat dairy, sweets, whole grains, and margarine. PC 6 represents a dietary pattern with high consumption of low-fat dairy and alcoholic beverages (e.g., liquor, beer, wine). The remaining principal components can be interpreted in a similar way based on the sign and absolute value of the PC loadings.

The 10 selected principal components were included in the multiple linear regression model to predict the levels of triglycerides, LDL cholesterol, HDL cholesterol, and total cholesterol. The regression model was first estimated using the training set and then evaluated using the test set. During the training stage, the regression model was statistically significant for triglycerides (F(18, 760) = 6.04, p < .001), LDL cholesterol (F(18, 736) = 3.85, p < .001), HDL cholesterol (F(18, 1617) = 6.04, p < .001), and total cholesterol (F(18, 1617) = 7.21, p < .001). To avoid model overfitting, model performance was evaluated using the independent test set. The correlation coefficient between the observed outcome and the predicted outcome for the test data was 0.43, 0.17, 0.50, and 0.21 for triglycerides, LDL cholesterol, HDL cholesterol, and total cholesterol respectively (Table 2). Additionally, adjusted R², the modified version of R² that has been adjusted for the number of predictors in the model, was calculated for each regression model using the test data (adjusted R² = 0.163, 0.005, 0.235, and 0.024 for triglycerides, LDL cholesterol, HDL cholesterol, and total cholesterol respectively).

The LASSO model was directly applied to the defined food categories to predict the levels of triglycerides, LDL cholesterol, HDL cholesterol, and total cholesterol. The model tuning parameter lambda was automatically selected using cross-validation to zero out the coefficients of categories not helpful for predicting the outcomes. For triglycerides, the model included 12 predictors with non-zero coefficients and yielded a correlation coefficient of 0.93 and adjusted R² of 0.861 for the independent test set (Table 2). For LDL, the model identified 18 predictors with non-zero coefficients and produced a correlation coefficient of 0.95 and adjusted R² of 0.899 for the test set. For predicting HDL and total cholesterol, the model selected 25 and 29 predictors. As shown in Table 2, the correlation coefficients calculated using the test set were 0.95 and 0.97 for HDL and total cholesterol respectively. The corresponding adjusted R² values were 0.890 and 0.935.

In the multiple regression with selected principal components, significant predictors for triglycerides were PC2 (b = − 0.037, t = − 2.99, p = 0.003), and PC3 (b = 0.039, t = 3.12, p = 0.002). This makes sense because PC2 represented a healthy dietary pattern with high consumption of fruits and vegetables and low consumption of liquor, soft drinks, beer, and gravy. Thus, the dietary pattern represented by PC2 was negatively associated with the level of triglycerides. PC3 characterized a relatively unhealthy dietary pattern with high consumption of high-fat dairy, red meat, refined grains, sweets and low consumption of meal replacements, organ meats, liquor and wine. Hence, the dietary pattern denoted by PC3 was positively associated with the level of triglycerides. Consistent with the findings from multiple regression analysis, the LASSO model suggested that increased consumption of meal replacements, organ meat, chili, dark-yellow vegetables, tea, and wine was negatively associated with the level of triglycerides.

Multiple regression analysis with selected principal components yielded significant predictors for LDL cholesterol, including PC4 (b = − 0.025, t = − 2.55, p = 0.011) and PC6 (b = − 0.032, t = − 2.76, p = 0.006). This suggests that both PC4 and PC6 negatively correlated with the level of LDL cholesterol. PC4 contained large negative loadings of non-dairy cream and sweets. PC6 represented a dietary pattern with high consumption of low-fat dairy and alcoholic beverages (e.g., liquor, beer, wine). Similar findings were suggested by the LASSO model. From Table 3, we can see that high consumption of low-fat dairy, meal replacements, organ meats, soups, chili, liquor, and wine predicted a decreased level of LDL cholesterol, while high consumption of high-fat dairy, red meat, refined grains, and sweets predicted an increased LDL cholesterol level.

Significant predictors for HDL cholesterol in the multiple regression model were PC2 (b = 0.008, t = 2.11, p = 0.035), PC3 (b = − 0.009, t = − 2.29, p = 0.022), PC4 (b = − 0.015, t = − 2.86, p = 0.004), PC5 (b = 0.029, t = 5.38, p < .001), and PC6 (b = 0.040, t = 6.43, p < .001). This suggests that the level of HDL cholesterol, known as “good” cholesterol, positively correlated with high consumption of fruits and vegetables (represented by PC2) and low consumption of high-fat dairy, sweets and red meats (characteristics of PC3). Additionally, increased consumption of vegetables, salad dressing and beer (denoted by PC5) as well as low-fat dairy and alcoholic beverages (PC6) predicted an increased HDL cholesterol level. PC4 represented a dietary pattern with high consumption of a number of liquids (e.g., juice, fruit juices and soft drinks) and low consumption of non-dairy cream and sweets. Hence, the dietary pattern characterized by PC4 was negatively associated with the level of both “good” and “bad” cholesterol levels. Results from the LASSO model suggest that increased consumption of vegetables and decreased consumption of a number of liquids (e.g., juice and soft drinks) predicted increases in the level of HDL cholesterol. Additionally, consumption of proteins such as organ meats, nuts, chili peppers, and eggs predicted decreases in HDL, while high-fat dairy and refined grains predicted increases in HDL level.

For total cholesterol levels, PC3 (b = 0.007, t = 2.41, p = 0.016), PC4 (b = − 0.013, t = − 3.07, p = 0.002), and PC5 (b = 0.016, t = 3.79, p < .001) were significant predictors in the multiple regression model. As described previously, PC3 represented a relatively unhealthy dietary pattern with high consumption of high-fat dairy, red meat, refined grains, and sweets. Thus, PC3 was positively associated with the level of total cholesterol. The dietary pattern characterized by high consumption of a number of liquids and low consumption of non-dairy cream and sweets (PC4) was negatively associated with the level of total cholesterol. Additionally, PC5, characterized by high consumption of vegetables, salad dressing and beer, was positively related to the level of total cholesterol. Results from the LASSO model suggested similar findings. The increased consumption of high-fat dairy, red meat, sweets, and certain vegetables (e.g., peppers, green beans) predicted an increase in total cholesterol, while higher consumption of non-dairy cream, meal replacements, organ meats, legumes, liquor and soft drinks predicted a decrease.