Using Multiple Statistical Methods to Derive Dietary Patterns Associated with Cardiovascular Disease in Patients with Type 2 Diabetes: Results from a Multiethnic Population-Based Study.

Background: There are few reports on the relationship between dietary patterns and cardiovascular disease (CVD) risk in patients with type 2 diabetes (T2D). This study aimed to explore relationships between dietary patterns and CVD risk in the T2D population using multiple statistical analysis methods.
Methods: A total of 2,984 patients with T2D from the Xinjiang Multi-Ethnic Cohort, 555 of whom were suffering from CVD, were enrolled in this study. Participants' dietary intake was measured by the semiquantitative food frequency questionnaire (FFQ). Three statistical methods were used to construct dietary patterns, including principal component analysis (PCA) method, reduced-rank regressions (RRR) method, and partial least-squares regression (PLS) method. Then, the association between dietary patterns and CVD risk in T2D patients was analyzed by logistic regression. After excluding participants with CVD, the associations between dietary patterns and 10-year CVD risk scores were subsequently evaluated to reduce reverse causality.
Results: In this study, four dietary patterns were identified by three methods. Adjustment for confounding factors, subjects with the highest scores on the "high-protein and high-carbohydrate" patterns derived from PCA, RRR, and PLS had higher odds of CVD than those with the lowest scores (OR: 2.89, 95% CI: 2.11-3.96, P trend < 0.001; OR: 2.96, 95% CI: 2.17-4.03, P trend < 0.001; OR: 2.01, 95% CI: 1.50-2.70, P trend < 0.001, respectively). However, the dietary pattern of PCA-prudent was not significantly related to the odds of having CVD in T2D patients (adjusted ORQ4vsQ1: 0.93, 95% CI: 0.70-1.24, P trend =0.474). Interestingly, we also found significant associations between "high-protein and high-carbohydrate" patterns and the elevated predicted 10-year CVD risk in T2D patients (all P trend < 0.05).
Conclusion: The positive correlation between "high-protein and high-carbohydrate" patterns and CVD risk in T2D patients was robust across all three data-driven approaches. These findings may have public health significance, encouraging an emphasis on food choices in the usual diet and promoting nutritional interventions for patients with T2D to prevent CVD.

1. Introduction

Diabetes mellitus has impacted about 537 million individuals worldwide, more than 90% of whom suffer from T2D [1]. In China, the proportion of T2D reached 12.75% in 2018, affecting approximately 116 million adults [2, 3]. Unfortunately, those with T2D are at higher risk for concurrent fatal or nonfatal cardiovascular disease, leading to increased disability and mortality in T2D patients [4]. Compared with healthy people, T2D patients have a 2- to 4-fold increased risk of CVD and a 50% increased risk of cardiovascular mortality [5, 6]. Approximately 50.3% of all deaths in patients with T2D are due to CVD [7]. In addition, cardiovascular diseases place a tremendous financial burden on patients and healthcare systems. Compared with treating T2D patients alone, treating patients with CVD and T2D costs an additional $3,418 to $9,705 per year [8]. Therefore, reducing CVD risk factors is crucial to preventing or delaying the development of CVD in T2D patients.

As a potentially modifiable factor, people's diet has a crucial role in the early prevention of CVD, which has been incorporated into diabetes risk management criteria [9]. Evidence from the T2D population indicates that tree nuts [10], fiber [11], fruits, and vegetables [12] could reduce CVD risk, while higher levels of eggs [13] and saturated fat [14] increase the incidence of CVD. Although existing studies have confirmed the effectiveness of diet in preventing CVD in patients with T2D, they only concentrate on a certain food or a certain nutrient, which ignoring the interactions and synergies between foods or nutrients. Guidelines published by the American Diabetes Association recommend that healthy eating patterns, rather than a single nutrient or food, should be provided for people with diabetes to prevent diabetic complications [15].

Dietary pattern analysis is a comprehensive method to evaluate the antagonistic and synergistic effects of various foods on health outcomes, and its results are more helpful in translating into dietary recommendations [16, 17]. PCA or factor analysis was the common method in the process of dietary patterns analysis, which maximizes the explanation of variation in food group intake [18]. PLS and RRR are also increasingly applied in epidemiology as they integrate multivariate approaches with prior knowledge of underlying pathophysiological pathways to improve the prediction of diseases [19]. Previous studies have made different recommendations on the utility of these methods [19, 20], but their performance in the same population remains controversial. For instance, findings from an aging Australian population showed that the RRR was superior to the PCA and the PLS in determining dietary patterns associated with bone mass [21]. Conversely, two other studies on the general population found that the PCA and the PLS produced more disease-related dietary patterns [22, 23].

In previous studies, dietary patterns associated with CVD risk factors were obtained by the PCA and the RRR in patients with T2D [24, 25]. Nevertheless, to our knowledge, no study has applied PCA, RRR, and PLS to comprehensively study the relationships between dietary patterns and CVD in patients with T2D. Furthermore, almost all studies comparing the validity of different dietary pattern analysis methods have been conducted in European and American populations. Considering that dietary patterns are strongly influenced by regional, socio-demographic, economic, and cultural factors [26], data from Chinese populations, especially multi-ethnic populations with different dietary habits, are needed. Thus, the present study aimed to utilize three analysis approaches (PCA, RRR, and PLS) to identify dietary patterns associated with CVD among T2D patients in Xinjiang, China.

2. Materials and Methods

2.1. Study Design and Population

Patients were enrolled from Xinjiang Multi-Ethnic Cohort (XMEC). Simultaneously, this was a population-based study organized from April 2018 to May 2019. The ethical clearance for this study came from the Xinjiang Uyghur Autonomous Region Institute of Traditional Chinese Medicine (2018XE0108), and all patients signed written informed consent certificates. More details of the XMEC have been reported elsewhere [27].

In this study, we included a subsample of adults diagnosed with T2D (n = 3,759). T2D cases were identified as any one of the following criteria: [1] treatment with antidiabetic medications (insulin or oral hypoglycemic medication); [2] fasting plasma glucose (FPG) ≥7.0 mmol/L; and [3] self-reported past medical history was T2D. Those with missing data on serum lipids (n = 660) or other major covariates (n = 49) were excluded. After further exclusions of participants without dietary intake (n = 23) or with implausible total energy intake (<800 kcal/d or >6000 kcal/d for men; <600 kcal/d or >4000 kcal/d for women; n = 43) [28], a total of 2,984 eligible subjects with T2D were available for the analysis of dietary patterns and CVD prevalence. To reduce reverse causality, 555 participants diagnosed with CVD were excluded, leaving 2,429 to analyze the relationships between dietary patterns and 10-year CVD risk (Figure 1).

Figure 1

The flow chart illustrates the number of participants included in the analysis in the Xinjiang Multi-Ethnic Cohort study.

2.2. Assessment of Dietary Intake

The intensively trained interviewers assessed the habitual diet of participants using a semiquantitative food frequency questionnaire (FFQ), which has been presented previously [27]. The FFQ included 127 food items and was used to assess the frequency (daily, 4∼6 times/week, 1∼3 times/week, 1∼3 times/month, and never or rarely) and quantity of consumption (in units or specified portion size) for the past year. In this study, 127 foods were divided into 23 food groups according to similar nutritional composition or culinary use (relevant information can be found in Supplemental Table 1).

2.3. Assessment of Dietary Patterns

The dietary patterns were determined by three data-driven patterns, including PCA, RRR, and PLS. As for the PCA, the optimal structure was achieved by using orthogonal rotation, and the Kaiser–Meyer–Olkin (KMO) test was performed to measure sample adequacy. The number of dietary patterns was identified by three aspects, including the eigenvalues >1.5, scree plot, and interpretability of these factors [29]. In RRR and PLS analysis, the response variables were biomarkers of serum lipids (log-transformed values), including triglycerides (TG), high-density lipoprotein cholesterol (HDL-C), and low-density lipoprotein cholesterol (LDL-C). We chose TG, HDL-C, and LDL-C because the dietary effect on CVD risk in T2D patients is partly mediated by pathways that alter the serum lipid profiles [15, 25, 30]. For RRR and PLS, we only reserved the first dietary pattern for the subsequent analyses as they interpreted the greatest proportion of change in the food groups. When naming the patterns, food groups with an absolute factor load ≥0.20 were accounted for in all three methods [31]. Participants' dietary pattern scores were obtained by adding up the standardized daily consumption of all food groups and multiplying by their weighted factor loadings.

2.4. Ascertainment of Cardiovascular Disease

CVD was defined as the complex between nonfatal coronary heart disease (CHD) and nonfatal stroke. Among them, CHD includes coronary artery bypass graft surgery and nonfatal myocardial infarction. Patients with a self-reported diagnosis of CVD were asked to allow physicians access to their medical records. Nonfatal myocardial infarction and nonfatal stroke were identified according to the criteria of the World Health Organization [32] and the National Survey of Stroke [33].

2.5. Estimation of 10-Year Cardiovascular Disease Risk

The 10-year CVD predicted risk was measured by the China-PAR equation, a validated tool with good performance in the Chinese population [34, 35]. The risk factors included sex, age, treated or untreated systolic blood pressure, total cholesterol (TC), waist circumference (WC), HDL-C, current smoker, geographic region, diabetes status, urbanization, and family history of CVD [34]. Participants with a 10-year CVD risk score ≥10% were considered high-risk groups for future CVD events [36].

2.6. Assessment of Covariates

Trained investigators conducted all anthropometric measurements following standardized protocols. The participants were required to wear light clothes without shoes. The weight and height were measured normally (SK-X80, China), and the values were recorded to the closest 0.1 kg and 0.1 cm. The body mass index (BMI) was estimated by dividing weight (kg) by the square of height (m). After the participants exhaled normally, their WC was measured midway between the bottom ribs and pelvic bones, and the measurement accuracy was 0.1 cm. After 5 minutes of rest, participants measured their blood pressure twice in a standard way and recorded the average value of the two measurements as the final blood pressure.

Data on additional variables, including age (years), sex (male/female), region (Urumqi/Huo Cheng/Mo Yu), race (Han/Hui/Uyghur/Kazakh/other), education (elementary school or below/middle school/high school or beyond), marital status (married/widowed or divorced), current smoking (yes/no), hypertension (yes/no), family history of CVD (yes/no), and family history of diabetes (yes/no) were obtained by questionnaires. The International Physical Activity Questionnaire-Short Form was used to assess the physical activities in the past year, which could be divided into three categories: low level (<600 MET-minutes per week), moderate level (600∼3000 MET-minutes/week), and high level of physical activity (>3000 MET-minutes/week) [37].

2.7. Laboratory Assessment

Blood specimens were collected after 12 hours of overnight fasting, and the samples were temporarily stored at −20°C. Then, blood samples were analyzed at the local village health centers of the three survey sites. Fasting plasma glucose (FPG), TG, TC, HDL-C, and LDL-C were analyzed using automated biochemical analyzers (HITACHI 7600, Tokyo, Japan).

2.8. Statistical Analysis

Categorical variables were summarized as numbers (%), and continuous variables were represented as means and standard deviations or medians and interquartile ranges. The chi-squared tests, t-tests, or Mann–Whitney U tests were used to compare the baseline characteristics of participants with or without CVD. Linear regression analysis of dietary pattern scores and response variables (HDL-C, LDL-C, and TG) was performed to obtain the variation in response variables interpreted by dietary patterns. Pearson correlation tests were adopted to estimate the relationships between dietary pattern scores and response variables (log-transformed values).

The relationship between dietary patterns and CVD prevalence in T2D patients was explored by multivariable logistic regression. The first model was a crude model with dietary patterns as the only predictor. The second model was adjusted for energy intake, age, sex, region, race, education, current smoking status, marital status, hypertension, physical activity, body mass index, and waist circumference. The dietary pattern score used ordinal categorical variables and a likelihood ratio test to test the linear trends across quartiles. In addition, the association between each 1 SD increase in dietary pattern and the odds ratio of having a CVD was also calculated.

After excluding participants with CVD, the correlations between dietary patterns and predicted high 10-year CVD risk in both models were assessed. Then, variables included in the CVD risk score were not adjusted to avoid overadjustment. The SAS 9.4 (SAS Institute Inc, Cary, NC, USA) and R version 4.0.3 were used for data analysis. When P < 0.05, it was statistically significant.

3. Results

3.1. Characteristics

The 2,984 eligible individuals were analyzed, with 555 CVD (18.6%) cases. Table 1 shows the characteristics of the T2D patients based on CVD status. In contrast to non-CVD, most T2D patients with CVD were female, Uyghur, widowed/divorced, and older; have hypertension; and have higher LDL-C, TG, BMI, and WC but lower HDL-C. In addition, CVD cases had CVD in their family more often, tended to live in Mo Yu, and had lower education and physical activity levels than controls.

Table 1

Characteristics of the participants with T2D by CVD status.

	Total (n = 2,984)	CVD (n = 555)	Non-CVD (n = 2,429)	P -value∗
Age (years)
Mean ± SD	61.77 ± 9.53	63.55 ± 8.38	61.36 ± 9.73	<0.001
Sex (%)
Male	1251 (41.9)	212 (38.2)	1039 (42.8)	0.049
Female	1733 (58.1)	343 (61.8)	1390 (57.2)
Region (%)
Urumqi	1857 (62.2)	308 (55.5)	1549 (63.8)	<0.001
Huo Cheng	438 (14.7)	78 (14.1)	360 (14.8)
Mo Yu	689 (23.1)	169 (30.4)	520 (21.4)
Race (%)
Han	1731 (58.0)	268 (48.3)	1463 (60.2)	<0.001
Hui	292 (9.8)	58 (10.5)	234 (9.7)
Uyghur	886 (29.7)	222 (40.0)	664 (27.3)
Kazakh	51 (1.7)	4 (0.7)	47 (1.9)
Other	24 (0.8)	3 (0.5)	21 (0.9)
BMI (kg/m2)
Mean ± SD	26.30 ± 3.95	26.74 ± 4.29	26.20 ± 3.87	0.007
Waist circumference (cm)
Mean ± SD	92.91 ± 10.98	95.44 ± 11.24	92.33 ± 10.84	<0.001
Education (%)
Primary school or below	1658 (55.6)	348 (62.7)	1310 (54.0)	<0.001
Middle school	813 (27.2)	138 (25.0)	675 (27.8)
High school or beyond	513 (17.2)	69 (12.4)	444 (18.3)
Marital status (%)
Married	2531 (84.8)	452 (81.4)	2079 (85.6)	0.014
Widowed/Divorced	453 (15.2)	103 (18.6)	350 (14.4)
Physical activity (%)
Low	2116 (70.9)	434 (78.2)	1682 (69.2)	<0.001
Moderate	804 (26.9)	115 (20.7)	689 (28.4)
High	64 (2.1)	6 (1.1)	58 (2.4)
Current smoker (%)
Yes	444 (14.9)	78 (14.1)	366 (15.1)	0.545
No	2540 (85.1)	477 (85.9)	2063 (84.9)
Family history of CVD (%)
Yes	382 (12.8)	102 (18.4)	280 (11.5)	<0.001
No	2602 (87.2)	453 (81.6)	2149 (88.5)
Family history of diabetes (%)
Yes	308 (10.3)	54 (9.7)	254 (10.5)	0.611
No	2676 (89.7)	501 (90.3)	2175 (89.5)
Hypertension (%)
Yes	1784 (59.8)	417 (75.1)	1367 (56.3)	<0.001
No	1200 (40.2)	138 (24.9)	1062 (43.7)
TC (mmol/L)
Median (interquartile range)	4.90 (4.11, 5.64)	4.84 (4.01, 5.86)	4.90 (4.13, 5.60)	0.619
TG (mmol/L)
Median (interquartile range)	1.50 (1.07, 2.16)	1.57 (1.13, 2.35)	1.50 (1.05, 2.13)	0.005
HDL-C (mmol/L)
Median (interquartile range)	1.36 (1.13, 1.64)	1.33 (1.13, 1.60)	1.36 (1.14, 1.65)	0.048
LDL-C (mmol/L)
Median (interquartile range)	2.65 (2.00, 3.22)	2.78 (2.13, 3.32)	2.63 (1.98, 3.20)	0.001
FPG (mmol/L)
Median (interquartile range)	7.60 (6.42, 9.26)	7.40 (6.10, 9.68)	7.62 (6.55, 9.20)	0.253

The P values were derived from the chi-squared test for categorical variables and Student's t-test or the Mann–Whitney U test for continuous variables.

3.2. Dietary Pattern Analysis

Factor loadings derived from the three analysis methods are presented in Figure 2. Two dietary patterns determined by PCA jointly accounted for 20.04% of the total variance in food intake. The first dietary pattern (PCA-prudent dietary) was rich in whole grains, poultry, seafood, eggs, vegetables, root crops, fruits, soybean products, milk, low-fat dairy products, pickles, dried vegetables, dried fruit, nuts, and tea, which accounted for 12.30% of the total variance in food intakes and 0.50% in response variables (Table 2). The second dietary pattern (PCA-high-protein and high-carbohydrate), which explained 7.74% of the total variance in food intakes and 7.80% in response variables (Table 2), was specified as high consumption of refined grains, red meat, poultry, eggs, organ meat, and high fat dairy products, and low consumption of whole grains, vegetables, fruits, and tea (Figure 2).

Figure 2

Principal component analysis, reduced-rank regression, and partial least-squares regression were used to study the factor load of the food group in the dietary pattern. Food groups with positive impacts on dietary patterns are shown in red, while food groups with negative impacts on dietary patterns are shown in blue.

Table 2

Response variables explained by dietary patterns obtained by three statistical methods and the proportion of changes in the food group.

Dietary patterns	Explained variation in food groups (%)	Explained variation in response variables (%)
		Triglycerides	High-density lipoprotein cholesterol	Low-density lipoprotein cholesterol	Total
PCA-prudent pattern	12.30	0.50	4.50	2.80	0.50
PCA-high-protein and high-carbohydrate patterns	7.74	6.00	7.40	13.2	7.80
RRR-high-protein and high-carbohydrate patterns	7.64	14.05	7.98	15.22	12.42
PLS-high-protein and high-carbohydrate patterns	8.96	11.27	8.29	15.14	11.57

The RRR-derived pattern (RRR-high-protein and high-carbohydrate) included high consumption of refined grains, red meat, poultry, and low consumption of whole grains, vegetables, fruits, and tea (Figure 2). The RRR-high-protein and high-carbohydrate pattern explained 12.42% of the variation in three intermediate variables (14.05%, 7.98%, and 15.22% of the variation in TG, HDL-C, and LDL-C, respectively) and 7.64% in food groups (Table 2). The PLS-derived pattern (PLS-high-protein and high-carbohydrate) was a diet rich in refined grains, red meat, and poultry and low in whole grains, vegetables, fruits, root crops, nuts, and tea (Figure 2). The PLS-high-protein and high-carbohydrate patterns explained 11.57% of the variation in intermediate variables (11.27%, 8.29%, and 15.14% of the variation in TG, HDL-C, and LDL-C, respectively) and 8.96% in dietary components (Table 2).

Figure 3 shows the correlations between four dietary patterns and intermediate variables. PCA-prudent dietary pattern was slightly related to higher HDL-C (r = 0.18), lower LDL-C (r = -0.17), and TG (r = -0.09). The “high-protein and high-carbohydrate” patterns determined by PCA, PLS, and RRR were inversely correlated with HDL-C and positively correlated with TG and LDL-C.

Figure 3

Correlations between dietary patterns and response variables were derived from the principal component analysis, partial least-squares, and reduced-rank regressions. PCA1, principal component analysis-prudent dietary pattern; PCA2, principal component analysis-high-protein and high-carbohydrate patterns.

3.3. Dietary Patterns and Participant Characteristics

The characteristics of the subjects according to four dietary patterns are indicated in Supplemental Tables 2 and 3. Subjects with higher scores in PCA-prudent diet pattern were older, Han, married, more physically active, more educated, lived in Urumqi, had higher HDL-C, but lower TC, TG, and LDL-C, and had a higher prevalence of CVD in their families. By contrast, the group with the highest scores in the PCA-high-protein and high-carbohydrate patterns tended to be male, current smokers, have higher BMI, WC, TC, TG, LDL-C, and FPG but lower HDL-C. In addition, individuals who had higher scores on the RRR-derived pattern or PLS-derived pattern were generally more likely to be younger, Uyghur, less educated, live in Mo Yu, and have higher BMI, WC, TC, TG, and LDL-C, but lower HDL-C.

3.4. Association Analysis between Diet (Patients with T2D) and Cardiovascular Disease

Figure 4 shows associations of the dietary patterns with the CVD across the three methods. Adjustment for confounding factors, subjects with the highest scores on “high-protein and high-carbohydrate” patterns derived from PCA, RRR, and PLS had higher odds of CVD than those with the lowest scores (OR: 2.89, 95%, CI: 2.11–3.96, P < 0.001; OR: 2.96, 95% CI: 2.17–4.03, P < 0.001; OR: 2.01, 95% CI: 1.50–2.70, P < 0.001; respectively). However, PCA-prudent dietary pattern was not significantly related to the odds of having CVD (adjusted ORQ4vsQ1: 0.93, 95% CI: 0.70–1.24, P=0.474). Furthermore, each 1 SD increased in the “high-protein and high-carbohydrate” patterns derived from PCA, RRR, and PLS was related to 32% (adjusted OR: 1.32, 95% CI: 1.20–1.46), 53% (adjusted OR: 1.53, 95% CI: 1.37–1.71), and 39% (adjusted OR: 1.39, 95% CI: 1.24–1.55) increased in the odds of CVD, respectively (Supplemental Table 4).

Figure 4

Associations between dietary pattern scores and the presence of cardiovascular disease in participants with T2D. Values are odds ratios and 95% confidence intervals presented in quartiles of the dietary patterns. The first model was crude, and the second model was adjusted for energy intake, age, sex, region, race, education, current smoking status, marital status, hypertension, physical activity, body mass index, and waist circumference. P trend was calculated by using the quartiles of dietary pattern scores as continuous variables in the models.

3.5. Associations between Dietary Patterns and 10-Year Risk of Cardiovascular Disease in T2D Patients

The “high-protein and high-carbohydrate” patterns derived from PCA (P=0.013), RRR (P=0.001), and PLS (P=0.001) were associated with higher predicted 10-year CVD risk, and these associations remained consistent after adjustment covariates (all P < 0.05) (Table 3). Adjustment for confounding factors, each 1 SD increased in the “high-protein and high-carbohydrate” patterns from PCA, RRR, and PLS were 12% (OR: 1.12, 95% CI: 1.02–1.23), 14% (OR: 1.14, 95% CI: 1.04–1.25), and 13% (OR: 1.13, 95% CI: 1.03–1.23) increased in 10-year elevated CVD risk, respectively (Table 3). However, the PCA-prudent pattern was not significantly related to the predicted 10-year ASCVD risk (adjusted ORQ4vsQ1: 0.78, 95% CI: 0.60–1.02, P=0.155) (Table 3).

Table 3

Odds ratio and 95% confidence interval, by quartile, of estimated 10-year CVD risk by dietary pattern among participants with T2D.

	Quartile of dietary pattern score (n = 2,429)				P-trend	OR for per 1 SDincrease (95% CI)
	Q1	Q2	Q3	Q4
	OR (95% CI)	OR (95% CI)	OR (95% CI)	OR (95% CI)
PCA-prudent dietary pattern
Model 1	1.00	0.92 (0.73, 1.16)	1.04 (0.82, 1.31)	0.84 (0.67, 1.05)	0.262	0.96 (0.88, 1.04)
Model 2	1.00	0.87 (0.67, 1.12)	1.00 (0.77, 1.29)	0.78 (0.60, 1.02)	0.155	0.95 (0.86, 1.04)
PCA-high-protein and high-carbohydrate patterns
Model 1	1.00	1.10 (0.88, 1.37)	1.16 (0.92, 1.45)	1.34 (1.06, 1.69)	0.013	1.13 (1.04, 1.23)
Model 2	1.00	1.08 (0.84, 1.38)	1.22 (0.94, 1.57)	1.30 (1.00, 1.70)	0.033	1.12 (1.02, 1.23)
RRR-high-protein and high-carbohydrate patterns
Model 1	1.00	1.14 (0.92, 1.42)	1.36 (1.08, 1.70)	1.44 (1.14, 1.81)	0.001	1.13 (1.04, 1.23)
Model 2	1.00	1.27 (0.99, 1.63)	1.47 (1.13, 1.89)	1.54 (1.17, 2.01)	0.001	1.14 (1.04, 1.25)
PLS-high-protein and high-carbohydrate patterns
Model 1	1.00	1.05 (0.84, 1.31)	1.24 (0.99, 1.55)	1.42 (1.12, 1.79)	0.001	1.11 (1.03, 1.21)
Model 2	1.00	1.20 (0.94, 1.54)	1.41 (1.09, 1.82)	1.45 (1.11, 1.91)	0.003	1.13 (1.03, 1.23)

Q, quartile; OR, odds ratio; CI, confidence interval; SD, standard deviation. Model 1: logistic regression model without adjustment. Model 2: logistic regression model with adjustment of energy intake, region, race, education, marital status, hypertension, physical activity, and body mass index. P trend was derived by using the quartiles of dietary pattern scores as continuous variables.

4. Discussion

Given the strengths and limitations of each analysis method, this study used three different methods (PCA, RRR, and PLS) to derive dietary patterns, thus better outlining the associations between dietary patterns and CVD risk in T2D patients. Each data-driven method did not identify exactly the same dietary patterns, but some similar labels emerged between the three methods. All three methods extracted the “high-protein and high-carbohydrate” patterns that were related to higher odds of CVD prevalence. The PCA-prudent dietary pattern was not significantly correlated with CVD prevalence. Among baseline subjects without CVD, we found that the “high-protein and high-carbohydrate” patterns derived from the three methods were also associated with the higher predicted 10-year CVD risk, while the PCA-prudent pattern was not.

Evidence on the relationship between dietary patterns and cardiovascular disease risk from the general population is well established [38-40]. However, a few studies specifically investigate dietary patterns and CVD in patients with T2D, particularly in the multi-ethnic Chinese populations. This study suggested that the PCA-prudent pattern was rich in whole grains, poultry, seafood, eggs, vegetables, root crops, fruits, soybean products, milk, low-fat dairy products, pickles, dried vegetables, dried fruit, and nuts. Although previous studies have demonstrated that many food groups in the PCA-prudent pattern were correlated with a lower risk of CVD in the general population [9, 23, 41], we investigated that the PCA-prudent pattern was not significantly associated with CVD prevalence and 10-year CVD risk score in T2D patients. Several possible reasons may explain the nonsignificant relationship between the PCA-prudent pattern and CVD risk for T2D patients. First, the health effects of dietary patterns are a combination of multiple foods rather than simply adding up the effects of different food groups. Second, the low correlations between the PCA-prudent pattern and blood lipids (TG, HDL-C, and LDL-C) may partially explain these insignificant correlations, as existing evidence suggests that abnormal lipid metabolism in diabetic populations is the cause of CVD [42]. Third, dietary patterns obtained by PCA tended to reflect the actual eating habits of the target population but maybe weakly correlated with health outcomes because behaviorally relevant patterns do not necessarily predict target disease risk [43]. Thus, we found that the PCA-prudent pattern explained the highest variation in food groups (12.3%) but was not significantly related to CVD risk for T2D patients.

In this study, the “high-protein and high-carbohydrate” patterns obtained from PCA, RRR, and PLS were significantly correlated with CVD risk in T2D patients. Meanwhile, the “high-protein and high-carbohydrate” patterns derived from the three methods were rich in refined grains, red meat, and poultry and low in whole grains, fruits, vegetables, and tea. There is growing evidence, indicating that high-carbohydrate intake is related to an increased risk of cardiovascular disease morbidity and mortality [44]. Furthermore, we found that “high-protein and high-carbohydrate patterns” were mainly animal proteins, which were positively associated with a higher risk of CVD [45]. Compared with plant proteins, animal proteins have relatively high levels of dietary amino acids and aromatic amino acids, both of which are related to a high risk of CVD [46].

While there are differences among the three data-driven methods (PCA, RRR, and PLS), we found that dietary patterns identified by the three methods appear similar, suggesting that these methods can identify major underlying dietary components. Consistent with our findings, research on middle-aged women found that western dietary patterns from PCA, PLS, and RRR increased the risk of atherosclerosis [17]. However, another multi-ethnic cohort study from the United States showed that the RRR-derived dietary pattern was related to the risk of atherosclerosis, yet the dietary pattern determined by PCA was not [18]. Different populations, health outcomes, and intermediate response variables may explain this inconsistency. This study selected serum lipids (TG, HDL-C, and LDL-C) as intermediate-risk biomarkers between dietary patterns and CVD. The “high-protein and high-carbohydrate” patterns obtained by PCA, PLS, and RRR were inversely correlated with HDL-C and positively with TG and LDL-C, demonstrating that serum lipids might mediate the connection between diet and CVD risk in T2D patients.

Among the three “high-protein and high-carbohydrate” patterns, we have proved that the patterns obtained by the RRR and PLS methods have stronger associations with intermediate variables. Similarly, other studies suggested that RRR- and PLS-derived patterns accounted for more changes in intermediate variables than the PCA-derived patterns [47, 48]. Compared with PCA, the main advantage of RRR and PLS is that establishing dietary patterns incorporates intermediate variables, thus allowing the assessment of underlying etiologic mechanisms and informing early intervention in specific pathways. Furthermore, we found that the pattern determined by PLS explained changes in intermediate variables similarly to the RRR, but explained more changes in dietary intake than the RRR. PLS represents a trade-off between PCA and RRR by using a covariance matrix between predictors (dietary intake) and responses, allowing for a balance between the two objectives of explaining changes in the predictors and explaining changes in the responses [20].

This study first introduced the relationship between diet pattern and cardiovascular disease risk in T2D patients. Our investigation found that the PCA-prudent pattern was not significantly correlated with CVD prevalence and 10-year CVD risk score in T2D patients. On this basis, it explains the insignificant relationship between PCA-prudent pattern and CVD risk in T2D patients. The “high-protein and high-carbohydrate” patterns obtained by PCA, PLS, and RRR were negatively correlated with HDL-C and positively correlated with TG and LDL-C.

The main innovation of this study is to use three methods to determine the CVD-related dietary patterns of T2D patients in the Chinese population. All three methods extracted similar dietary patterns associated with higher CVD odds in T2D patients. Moreover, the samples included several ethnic groups and had important information on their demographics and physical health, allowing for adjusted analyses of several variables traditionally related to CVD risk. At the same time, this study has some limitations. First, our study used self-reported FFQ to collect dietary data, which is prone to reporting bias. Nonetheless, FFQ is still widely used to measure habitual eating and is a reasonable, repeatable, and effective tool to evaluate overall dietary consumption [49]. Second, given that the design of this study was cross-sectional, the results from this study did not allow making inferences about the causal connection between diet and CVD risk in T2D patients. Therefore, our study needs to be demonstrated in future prospective studies. Third, despite controlling for several confounders, residual confounding from other variables may have affected the results.

5. Conclusion

In conclusion, the “high-protein and high-carbohydrate” patterns determined by PCA, PLS, and RRR were significantly related to CVD risk for patients with T2D. The common features of the “high-protein and high-carbohydrate” patterns were rich in refined grains, red meat, and poultry and low in whole grains, fruits, vegetables, and tea. These findings may have public health significance, encouraging an emphasis on food choices in the usual diet and promoting nutritional interventions for patients with T2D to prevent CVD. At the same time, we found that all three approaches provided valuable insights when investigating the relationships between dietary patterns and CVD risk in T2D patients, and the results complemented each other. Nonetheless, future investigations are still needed to verify the usefulness of PCA, RRR, and PLS in different response variables, disease outcomes, and populations.