A Dietary Pattern Derived from Reduced Rank Regression and Fatty Acid Biomarkers Is Associated with Lower Risk of Type 2 Diabetes and Coronary Artery Disease in Chinese Adults

ABSTRACT

Background

Combinations of circulating fatty acids may affect the risk of type 2 diabetes (T2D) and coronary artery disease (CAD). No previous studies have identified a dietary pattern predicting fatty acid profiles using reduced rank regression (RRR) and evaluated its associations with the risk of T2D and CAD.

Objective

The aim of this study was to derive a dietary pattern to explain variation in plasma fatty acid concentrations using RRR and evaluate these in relation to risk of T2D and CAD.

Methods

We derived a dietary pattern using fatty acid concentrations from 711 controls of a nested case-control study in the Singapore Chinese Health Study using RRR with 36 food and beverages as predictors and 19 fatty acid biomarkers as responses. Dietary pattern scores were then calculated for the full cohort of men and women (mean age: 56 y). We followed up 45,411 and 58,065 participants for incident T2D and CAD mortality, respectively. Multivariable Cox regression models were used to estimate HRs and 95% CIs.

Results

We identified a dietary pattern high in soy, vegetables, fruits, tea, tomato products, bread, fish, margarine and dairy, and low in rice, red meat, coffee, alcohol, sugar-sweetened beverages, and eggs. This pattern predicted higher circulating n–3 (ω-3) PUFAs (18:3n–3, 20:3n–3, 20:5n–3), odd-chain fatty acids (15:0, 17:0), 18:2n–6 and 20:1, and lower 20:4n–6 and 16:1. During a mean follow-up of 11 y and 19 y, 5207 T2D and 3016 CAD mortality events, respectively, were identified. Higher dietary pattern scores were associated with a lower risk of T2D [multivariable-adjusted HR comparing extreme quintiles, 0.86 (95% CI: 0.79, 0.95); P-trend <0.001] and CAD mortality [HR, 0.76 (95% CI: 0.68, 0.86); P-trend <0.001].

Conclusions

Dietary patterns reflecting higher circulating n–3 PUFAs, odd-chain fatty acids, and linoleic acid may be associated with lower T2D and CAD risk in Chinese adults. This trial was registered at www.clinicaltrials.gov as NCT03356340.

Introduction

Several approaches are commonly used in nutritional epidemiology to derive dietary patterns. Principal component, factor, and cluster analysis are exploratory methods that derive dietary patterns based on correlations between food intake. These patterns reflect food consumption habits in a population, but not necessarily aspects of the diet that are important in disease development (1). Reduced rank regression (RRR) can be used to derive dietary patterns that maximally explain variation in a set of predetermined response variables, typically intermediates related to disease risk. Because an a priori hypothesis can be incorporated in the form of response variables, derived dietary patterns may be more readily linked to biological pathways relevant for disease etiology (2). Previous studies using RRR have used inflammatory biomarkers (3–5), serum lipids (6), nutrient intakes (1), or a combination of risk factors (7, 8) to generate dietary patterns predictive of the risk of type 2 diabetes (T2D) or coronary artery disease (CAD).

The intake of fatty acids and fatty acid biomarkers have been examined in relation to the risk of T2D and CAD in many studies (9–20), and the consumption of various food groups is known to influence variations of fatty acid biomarkers (21). However, to our knowledge, no research has been published that identifies food consumption patterns predicting circulating fatty acid profiles. The aim of this study was to derive a food consumption pattern predictive of fatty acid biomarkers using RRR analysis and evaluate its association with T2D and CAD mortality in an ethnic Chinese population. We hypothesize that RRR analysis will identify a dietary pattern associated with T2D and CAD risk that may reflect differences in plasma fatty acid concentrations.

Methods

Study population

The Singapore Chinese Health Study (SCHS) is a population-based, prospective cohort established between April 1993 and December 1998 that enrolled 27,959 ethnic Chinese men and 35,298 women aged 45–74 y. We recruited participants who lived in government-built housing estates, where 86% of the Singapore population resided during the enrollment period (22). All consenting participants were interviewed in person using structured questionnaires. The first and second rounds of follow-up interviews were conducted by telephone during the periods of 1999 to 2004 and 2006 to 2010, respectively. All participants gave informed consent and the Institutional Review Boards at the National University of Singapore and the University of Pittsburgh approved the study.

For the derivation of the dietary pattern using RRR, we used data from the controls (n = 711) of a case-control study of acute myocardial infarction nested in the SCHS (17). Briefly, these controls provided blood samples and did not have a history of heart disease or stroke at the time of blood collection based on self-reported diagnosis and data from hospital discharge databases.

For the analysis of dietary patterns and CAD mortality, we used the entire SCHS cohort but excluded participants who had self-reported physician-diagnosed cancer (n = 864), stroke (n = 947), or heart disease (n = 2598) at baseline, and individuals who had implausible daily energy intakes (<700 kcal/d or >3700 kcal/d for men; <600 kcal/d or >3000 kcal/d for women) (n = 1060), resulting in a final number of 58,065 participants for the analysis. Reasons for exclusion were not mutually exclusive.

For the analysis of dietary patterns and T2D, we first excluded participants without data for both follow-up interviews, due to nonresponse or death (n = 8916). Participants who self-reported having physician-diagnosed diabetes (n = 5360), cancer (n = 1309), or heart disease (n = 1684) at baseline, and individuals who had implausible daily energy intakes (<700 kcal/d or >3700 kcal/d for men; <600 kcal/d or >3000 kcal/d for women) (n = 841) were further excluded resulting in a final number of 45,411 participants for the analysis.

Dietary assessment

Habitual diet over the past year was assessed using the interviewer-administered semiquantitative FFQ that contained 165 commonly consumed food and beverage items identified from a previous pilot study (22). Respondents reported intakes from 8 different frequencies (ranging from never or hardly ever to ≥2 times per day) and 3 options for serving size. The serving sizes, accompanied by color photographs, represented the 15^th, 50^th, and 85^th percentiles of the range of portion sizes obtained from the pilot study data. The food items were then aggregated into 26 food groups based on similarity of nutrient composition and culinary usage and included fresh red meat, organ red meat, preserved red meat, poultry, fish, shellfish, fresh eggs, preserved eggs, soy, nonsoy legumes, light green vegetables, dark green vegetables, yellow vegetables, white potatoes, tomato products, preserved vegetables, all fruits, noodles, rice, bread, cereals, biscuits and crackers, dairy, desserts, nuts and seeds, and sugar and candy.

The intake of energy and nutrients was calculated based on the FFQ using the Singapore Food Composition Table developed specifically for this cohort (22). The FFQ was subsequently validated against a series of 24-h dietary recalls among a randomly chosen subset of 810 participants (22). Similar distributions were obtained, with the difference in mean energy and nutrient intakes measured by the FFQ and 24-h recall within <10%.

T2D assessment

Incidences of T2D were assessed by asking participants the following question, “Have you been told by a doctor that you have diabetes (high blood sugar)?” If participants answered yes, we proceeded with the question “Please also tell me the age at which you were first diagnosed?” Cases were identified if they reported developing physician-diagnosed diabetes at any time between the baseline interview and the 2 follow-up telephone interviews.

We assessed the validity of using self-reported information on physician-diagnosed diabetes in this population among a subcohort (n = 1631) that reported incident diabetes at the first follow-up interview. The details of the validation have been published (23). Using standard protocols, we calculated the positive predictive value (at 98.8%) of self-reported diabetes in this cohort (23). We also measured glycated hemoglobin (HbA_1c) in 2625 randomly selected participants who reported having no diabetes and calculated the negative predictive value to be 94% using the HbA_1c diagnostic cut-off of ≥6.5% (23, 24).

CAD mortality assessment

Deaths were identified through record linkage with the Singapore Registry of Births and Deaths. For the current analysis, we updated mortality data until 31 December, 2016. As of 2015, only 49 participants were known to be lost to follow-up due to migration out of Singapore or for other reasons, suggesting that vital statistics in follow-up are virtually complete. Underlying causes of death were coded according to the International Classification of Diseases, 9th revision; we used codes 410–414 for CAD deaths.

Covariate assessment

Other covariates including cigarette smoking, alcohol intake, education level, medical history, physical activity, and height and weight were also assessed with the baseline questionnaire. BMI was calculated by taking the weight (kg) of a participant divided by the square of his/her height (m²). For participants with missing weight and/or height, BMI was calculated using imputed weight and/or height derived from linear regression (25). The respondents were also asked about their engagement in moderate and vigorous physical activities.

Measurement of plasma fatty acids

Plasma fatty acids were measured using a GC-tandem MS conducted on an Agilent 7890GC system equipped with a 7000B QQQ triple quadruple mass detector (Agilent) and an autosample injector. Total fatty acids including both free and esterified (triglycerides, phospholipids, cholesterol esters) fractions were measured. Nineteen plasma fatty acids were measured: 15:0, 16:0, 16:1, 17:0, 17:1, 18:0, 18:1n–9t, 18:1n–9, linoleic acid (LA; 18:2n–6), 18:3n–6, α-linolenic acid (ALA; 18:3n–3), 20:0, 20:1, 20:2, 20:3n–6, arachidonic acid (AA; 20:4n–6), 20:3n–3, EPA (20:5n–3), and DHA (22:6n–3). All unsaturated fatty acids except those ending with a ‘t’ are in the cis configuration. The within-batch CVs ranged from 3.7% to 8.1% whereas the between-batch CVs ranged from 7.8% to 16.8%.

Statistical analyses

The data analyses consisted of 2 major steps. First, we applied RRR with the goal of identifying dietary patterns that reflected plasma fatty acid profiles in a subsample of 711 controls from a previous case-control study on CAD (17). Next, based on the derived dietary pattern, we calculated dietary pattern scores for the full cohort and related the scores to the risk of T2D and CAD mortality.

In the RRR analysis, the predictor variables were data on food intake and the response variables were plasma fatty acid concentrations. We used 36 food and beverage variables: 26 food groups that were originally part of the cohort nutrition database, 5 specific food items (bread spreads—butter, margarine, and cooking oils—corn oil, peanut oil, soybean oil) that were substantial contributors to dietary fat, and 5 beverages (tea, coffee, sugar-sweetened beverages, fruit juices, and alcohol). Before the RRR process, all food and nutrient intakes were adjusted for total energy intake using the residual method (26) and all 19 measured fatty acids were expressed as a percentage of plasma total fatty acids. As RRR could produce multiple dietary pattern scores explaining variations of the 19 fatty acid biomarkers, we retained only the first pattern for subsequent analyses as this pattern contributed to a larger proportion of explained variance than other patterns. We calculated dietary pattern scores for the full cohort as the linear combination of all 36 weighted standardized component food groups/items with their loadings as coefficients. Excluding controls with a history of diabetes (n = 50) did not appreciably change the derived dietary patterns.

We used Pearson's correlation coefficients to assess the relation between dietary pattern scores and fatty acid biomarkers. Distributions of demographic, lifestyle behaviors, and dietary characteristics were compared across quintiles of dietary pattern scores.

For the T2D analysis, we computed the person-years for each participant from the year of recruitment to the year of reported T2D diagnosis or the year of the last completed follow-up interview for individuals who did not report diabetes diagnosis, whichever came first. For CAD mortality, we computed the person-years for each participant from the year of recruitment to the year of death, year of the last completed follow-up interview, or 31 December 2016, whichever came first. Cox proportional hazards regression was used to compute HRs and 95% CIs for risk of T2D and CAD mortality across quintiles of dietary pattern scores with the lowest quintile as the referent group. To examine linear trends, median values of the quintiles of each dietary pattern score were entered as a continuous variable in the Cox proportional hazards model. In the multivariable analyses, we considered the following potential confounding variables in our age-adjusted model and lifestyle-adjusted model: age at recruitment (y), dialect group (Hokkien or Cantonese), year of interview (1993–1995, 1996–1998), cigarette smoking [never smoker, former smoker, current smoker (1–12 cigarettes/d), current smoker (>13 cigarettes/d)], education level (none, primary, secondary, ‘A’ levels, and above), physical activity (none, <2 h/wk of vigorous or <4 h/wk of moderate, ≥2 h/wk of vigorous or ≥4 h/wk moderate), BMI (kg/m²), and total energy intake (quintiles). The multivariable-adjusted model additionally included self-reported history of physician-diagnosed hypertension (yes or no) and for CAD mortality as the outcome only, self-reported history of physician-diagnosed diabetes (yes or no). There was no evidence for the violation of the proportional hazards assumption using Schoenfeld's residuals (Ps >0.050). We tested for effect modification by age (median age: ≥54 and <54 y), sex, overweight status (using the Asian criteria of ≥23 or <23) and in CAD mortality analysis, history of diabetes by including an interaction term (product term between the potential effect modifier and the dietary pattern score modeled as median values of quintiles of intake) in the multivariable-adjusted model.

We used SAS statistical software (SAS Studio 3.71; SAS Institute Inc.) to derive the dietary patterns; Stata Software version 14 (StataCorp LP) was used for all other statistical analyses. Two-sided P values <0.050 were considered statistically significant.

Results

The dietary pattern identified with the RRR method was positively correlated with the n–3 PUFAs ALA, 20:3n–3, and EPA, odd-chain saturated fatty acids (15:0, 17:0), the major n–6 PUFAs LA and 20:1, and negatively correlated with the n–6 PUFAs AA and 16:1 (Table 1). Correlations between food groups and circulating fatty acids are shown in Supplemental Table 1. This dietary pattern represented a diet higher in soy, vegetables, fruits, tea, tomato products, bread, fish, margarine, and dairy products, and lower in rice, red meat, coffee, alcohol, sugar-sweetened beverages, and eggs. The identified food consumption pattern explained 1.7% of the variation in fatty acid profiles that were assessed on average 6 y later. For individual plasma fatty acids, the explained variation ranged from 0.0% for 16:0 to 7.0% for ALA (Supplemental Table 2).

TABLE 1.

Pearson correlations between dietary pattern scores and food groups/items and fatty acid biomarkers using the controls of a nested case-control study in the Singapore Chinese Health Study (n = 711)¹

	Dietary pattern
Food groups2
Positive loading
Soy	0.45**
Dark green vegetables	0.44**
Light green vegetables	0.43**
Fruits	0.40**
Tea	0.39**
Tomato products	0.29**
Bread	0.28**
Preserved vegetables	0.28**
Fish	0.27**
Yellow vegetables	0.26**
Margarine	0.26**
Dairy	0.22**
Negative loading
Rice	−0.40**
Fresh red meat	−0.34**
Coffee	−0.32*
Alcohol	−0.27**
Sugar-sweetened beverages	−0.25**
Organ red meat	−0.22**
Fresh eggs	−0.20**
Fatty acid biomarkers3
Positive correlation
18:3n–3 (ALA)	0.26**
15:0	0.18*
20:3n–3	0.17**
17:0	0.14**
20:5n–3 (EPA)	0.13**
18:2n–6 (LA)	0.12**
20:1	0.11**
Negative correlation
20:4n–6 (AA)	−0.16**
16:1	−0.15**

¹*Significant at P ≤0.050; **significant at P ≤0.010. The fatty acids 16:0, 17:1, 18:0, 18:1n–9t, 18:1n–9, 18:3n–6, 20:0, 20:2, 20:3n–6, 22:6n–3 were not correlated with the dietary pattern. AA, arachidonic acid; ALA, α-linolenic acid; LA, linoleic acid.

²For simplicity, only food groups with correlation coefficients ≥0.20 with the dietary patterns are displayed.

³All unsaturated fatty acids are in the cis configuration, except C18:1n–9t ending with a letter ‘t’ indicating trans.

We observed similar trends for baseline characteristics across quintiles of dietary pattern scores (Table 2 and Supplemental Table 3). Participants with higher scores on the fatty acid dietary pattern were more likely to be Cantonese, highly educated, and have a history of diabetes and hypertension, and were less likely to be smokers. Participants with higher scores consumed more fish, soy, vegetables, fruits, bread, dairy, and tea, and less red meat, eggs, noodles, rice, coffee, sugar-sweetened beverages, and alcohol. The dietary pattern score was associated with higher intakes of fat, particularly n–6 and n–3 PUFAs.

TABLE 2.

Baseline characteristics of participants according to quintiles of dietary pattern scores for the analytical cohort of coronary artery disease mortality in the Singapore Chinese Health Study (n = 58,065)¹

	Dietary pattern scores
	Quintile 1 (low)	Quintile 2	Quintile 3	Quintile 4	Quintile 5 (high)
Median score	−1.46	−0.62	−0.06	0.54	1.55
n	11,613	11,613	11,613	11,613	11,613
Age at interview, y	55.8 ± 7.8	56.5 ± 8.0	56.4 ± 8.0	56.3 ± 8.0	55.8 ± 7.8
Women, [n (%)]	4225 (36.4)	6573 (56.6)	7175 (61.8)	7445 (64.1)	7200 (62.0)
Cantonese, [n (%)]	4970 (42.8)	5133 (44.2)	5226 (45.0)	5522 (47.6)	6013 (51.8)
Higher education level2, [n (%)]	2739 (23.6)	2823 (24.3)	3105 (26.7)	3537 (30.5)	4439 (38.2)
Current smokers, [n (%)]	4348 (37.5)	2447 (21.1)	1879 (16.2)	1453 (12.5)	1235 (10.6)
Higher physical activity3, [n (%)]	1454 (12.5)	1065 (9.2)	1177 (10.1)	1281 (11.0)	1575 (13.6)
BMI, kg/m2	22.9 ± 3.3	23.1 ± 3.2	23.1 ± 3.3	23.2 ± 3.2	23.2 ± 3.3
History of hypertension, [n (%)]	2040 (17.6)	2456 (21.2)	2531 (21.8)	2760 (23.8)	2887 (24.9)
History of diabetes, [n (%)]	575 (5.0)	810 (7.0)	932 (8.0)	1078 (9.3)	1201 (10.3)
Energy intake, kcal/d	1706 ± 584	1436 ± 497	1436 ± 478	1491 ± 463	1681 ± 503
Food intake4, g/d
Red meat	38.8 ± 23.1	33.1 ± 16.4	30.4 ± 15.1	27.4 ± 14.8	21.6 ± 15.6
Poultry	20.1 ± 16.6	21.0 ± 14.1	20.7 ± 14.1	20.7 ± 15.6	19.5 ± 18.4
Fish	40.1 ± 22.0	46.4 ± 20.6	49.7 ± 22.3	52.8 ± 24.3	59.6 ± 32.5
Shellfish	6.10 ± 5.46	5.62 ± 4.31	5.26 ± 4.13	4.79 ± 4.00	4.19 ± 4.49
Eggs	16.8 ± 18.2	13.5 ± 11.2	12.8 ± 10.9	11.4 ± 9.7	9.7 ± 10.6
Soy	72 ± 60	96 ± 57	108 ± 63	122 ± 73	156 ± 112
Nonsoy legumes	2.92 ± 5.41	3.19 ± 4.10	3.20 ± 3.92	3.15 ± 3.78	3.16 ± 4.26
Vegetables	75 ± 36	97 ± 35	107 ± 39	119 ± 45	154 ± 71
Fruits	112 ± 116	162 ± 116	184 ± 121	205 ± 135	246 ± 171
Noodles	57.7 ± 45.2	57.0 ± 38.7	55.4 ± 37.6	52.4 ± 37.8	48.3 ± 40.8
Rice	497 ± 185	458 ± 139	419 ± 129	387 ± 125	323 ± 136
Bread	19.3 ± 19.1	28.1 ± 19.6	34.4 ± 21.5	39.5 ± 23.2	44.5 ± 27.8
Dairy	31 ± 66	51 ± 79	69 ± 99	90 ± 120	115 ± 146
Beverage intake
Tea, cups/d	0.25 ± 0.53	0.29 ± 0.58	0.39 ± 0.70	0.56 ± 0.89	1.21 ± 1.58
Coffee, cups/d	2.11 ± 1.42	1.53 ± 1.16	1.28 ± 1.08	1.09 ± 1.00	0.92 ± 0.97
Sugar-sweetened beverage, glasses/d	0.203 ± 0.516	0.077 ± 0.230	0.056 ± 0.188	0.047 ± 0.162	0.039 ± 0.148
Fruit juice, glasses/d	0.064 ± 0.191	0.058 ± 0.162	0.068 ± 0.204	0.080 ± 0.221	0.082 ± 0.317
Alcohol intake, g/d	5.4 ± 14.5	1.3 ± 5.2	0.9 ± 4.1	0.6 ± 3.2	0.7 ± 3.0
Total fat intake, % of total energy intake	22.8 ± 6.2	23.8 ± 5.3	25.0 ± 5.1	26.1 ± 5.0	27.8 ± 5.0
Saturated fat	8.53 ± 2.64	8.67 ± 2.45	8.92 ± 2.42	9.03 ± 2.46	9.19 ± 2.55
Monounsaturated fat	8.01 ± 2.31	8.20 ± 2.01	8.49 ± 1.93	8.70 ± 1.92	9.05 ± 1.91
Polyunsaturated fat	3.94 ± 1.26	4.45 ± 1.33	4.97 ± 1.48	5.54 ± 1.69	6.55 ± 2.19
EPA and DHA	0.158 ± 0.076	0.181 ± 0.083	0.192 ± 0.089	0.199 ± 0.092	0.214 ± 0.106
α-linolenic acid	0.242 ± 0.070	0.281 ± 0.082	0.318 ± 0.097	0.362 ± 0.130	0.434 ± 0.196
n–6 fatty acids	3.49 ± 1.17	3.96 ± 1.26	4.43 ± 1.42	4.96 ± 1.61	5.87 ± 2.07

¹Mean ± SD (all such values).

²Secondary school and above.

³≥2 h/wk of vigorous or 4 h/wk of moderate physical activity.

⁴Energy-adjusted values using the residual method (19).

We followed up 45,411 participants for a mean of 11 y (494,741 person-years of follow-up) and observed 5207 cases of incident T2D. In the age-adjusted model, a higher dietary pattern score was associated with a lower risk of T2D in women but not in men (Table 3). However, after further adjustments for potential confounders, similar associations were observed for men and women [multivariable-adjusted HR comparing extreme quintiles for men and women combined, 0.86 (95% CI: 0.79, 0.95); P-trend = 0.001; P-interaction for sex = 0.115]. No evidence for effect modification by age or overweight status was observed (all P-interactions >0.050).

TABLE 3.

HRs (95% CIs)¹ of type 2 diabetes according to quintiles of dietary pattern scores of participants in the Singapore Chinese Health Study (n = 45,411)

	Quintiles of score
	1 (low)	2	3	4	5 (high)	P-trend
Men, n	5695	3848	3371	3180	3315
Cases/person-years	642/61,135	451/41,256	382/36,488	347/34,421	373/35,948	––
Median score	−1.58	−0.64	−0.07	0.55	1.56	––
Age-adjusted HR	1.00	1.04 (0.92, 1.18)	1.00 (0.88, 1.13)	0.96 (0.84, 1.10)	0.99 (0.87, 1.12)	0.607
Lifestyle-adjusted HR2	1.00	1.02 (0.90, 1.15)	0.94 (0.83, 1.08)	0.94 (0.82, 1.08)	0.94 (0.82, 1.08)	0.228
Multivariable-adjusted HR3	1.00	1.01 (0.90, 1.15)	0.95 (0.83, 1.08)	0.93 (0.81, 1.06)	0.93 (0.81, 1.06)	0.157
Women, n	3388	5234	5711	5902	5767
Cases/person-years	419/37,289	664/57,219	699/62,436	638/65,217	592/63,332	––
Median score	−1.32	−0.60	−0.05	0.54	1.54	––
Age-adjusted HR	1.00	1.03 (0.91, 1.16)	1.00 (0.89, 1.13)	0.88 (0.77, 0.99)	0.84 (0.74, 0.96)	<0.001
Lifestyle-adjusted HR	1.00	1.01 (0.89, 1.14)	1.02 (0.90, 1.15)	0.89 (0.78, 1.00)	0.86 (0.75, 0.97)	0.001
Multivariable-adjusted HR	1.00	1.00 (0.88, 1.13)	1.00 (0.88, 1.13)	0.86 (0.76, 0.97)	0.81 (0.72, 0.93)	<0.001
Men and women combined4, n	9083	9082	9082	9082	9082
Cases/person-years	1061/98,424	1115/98,475	1081/98,924	985/99,638	965/99,280	––
Median score	−1.47	−0.62	−0.06	0.54	1.55	––
Age-adjusted HR	1.00	1.04 (0.96, 1.13)	1.00 (0.92, 1.09)	0.91 (0.83, 0.99)	0.90 (0.82, 0.98)	0.001
Lifestyle-adjusted HR	1.00	1.01 (0.93, 1.10)	1.00 (0.91, 1.09)	0.91 (0.83, 0.99)	0.89 (0.81, 0.98)	0.001
Multivariable-adjusted HR	1.00	1.01 (0.93, 1.10)	0.99 (0.91, 1.08)	0.89 (0.81, 0.97)	0.86 (0.79, 0.95)	<0.001

¹Estimates are HRs (95% CIs).

²Lifestyle-adjusted model controlled for age, father's dialect, year of interview, cigarette smoking, education level, physical activity, BMI, and energy intake.

³In addition to the covariables used in the lifestyle-adjusted model, adjusted for history of hypertension.

⁴ P-interaction for sex in the age-adjusted model, lifestyle-adjusted model, and multivariable-adjusted model, respectively, were 0.023, 0.203, and 0.115. All models were additionally adjusted for sex.

We followed up 58,065 participants for a mean of 19 y (1,077,170 person-years of follow-up) and identified 3016 CAD mortality events. A higher dietary pattern score was associated with a lower risk of CAD mortality in all statistical models [multivariable-adjusted HR, 0.76 (95% CI: 0.68, 0.86); P-trend <0.001] (Table 4). No evidence for effect modification by sex, age, and overweight status was observed (all P-interactions >0.050). When we stratified the participants by baseline history of diabetes, we found similar associations in both strata [HR for participants without diabetes, 0.74 (95% CI: 0.64, 0.85); P-trend <0.001; HR for participants with diabetes, 0.82 (95% CI: 0.63, 1.06); P-trend = 0.159; P-interaction for diabetes = 0.263].

TABLE 4.

HRs (95% CIs)¹ of coronary artery disease mortality according to quintiles of dietary pattern scores of participants in the Singapore Chinese Health Study (n = 58,065)

	Quintiles of score
	1 (low)	2	3	4	5 (high)	P-trend
Men, n	7388	5040	4438	4168	4413
Cases/person-years	492/129,374	365/88,290	290/78,644	286/74,668	290/79,367	—
Median score	−1.58	−0.64	−0.07	0.54	1.57	—
Age-adjusted HR	1.00	0.99 (0.86, 1.13)	0.85 (0.73, 0.98)	0.84 (0.73, 0.97)	0.83 (0.72, 0.96)	0.002
Lifestyle-adjusted HR2	1.00	1.01 (0.88, 1.16)	0.88 (0.76, 1.02)	0.93 (0.79, 1.08)	0.96 (0.82, 1.11)	0.331
Multivariable-adjusted HR3	1.00	0.96 (0.83, 1.10)	0.83 (0.71, 0.96)	0.85 (0.73, 0.99)	0.83 (0.71, 0.96)	0.004
Women, n	4225	6573	7175	7445	7200
Cases/person-years	204/79,687	309/125,221	265/138,018	287/143,965	228/139,867	—
Median score	−1.32	−0.60	−0.05	0.53	1.54	—
Age-adjusted HR	1.00	0.95 (0.80, 1.14)	0.75 (0.62, 0.90)	0.81 (0.67, 0.96)	0.73 (0.60, 0.88)	<0.001
Lifestyle-adjusted HR	1.00	0.97 (0.81, 1.16)	0.79 (0.65, 0.95)	0.89 (0.74, 1.07)	0.85 (0.70, 1.03)	0.063
Multivariable-adjusted HR	1.00	0.91 (0.76, 1.09)	0.70 (0.58, 0.84)	0.75 (0.63, 0.90)	0.69 (0.56, 0.84)	<0.001
Men and women combined4, n	11,613	11,613	11,613	11,613	11,613
Cases/person-years	696/209,061	674/213,511	555/216,662	573/218,633	518/219,234	—
Median score	−1.46	−0.62	−0.06	0.54	1.55	—
Age-adjusted HR	1.00	0.97 (0.87, 1.08)	0.80 (0.71, 0.89)	0.82 (0.73, 0.92)	0.77 (0.69, 0.87)	<0.001
Lifestyle-adjusted HR	1.00	0.99 (0.89, 1.11)	0.84 (0.75, 0.94)	0.91 (0.81, 1.02)	0.90 (0.80, 1.02)	0.037
Multivariable-adjusted HR	1.00	0.94 (0.84, 1.05)	0.77 (0.69, 0.87)	0.81 (0.72, 0.91)	0.76 (0.68, 0.86)	<0.001

¹Estimates are HRs (95% CIs).

²Lifestyle-adjusted model controlled for age, father's dialect, year of interview, cigarette smoking, education level, physical activity, BMI, and energy intake.

³In addition to the covariables used in the lifestyle-adjusted model, adjusted for history of hypertension, and history of diabetes.

⁴ P-interaction for sex in the age-adjusted model, lifestyle-adjusted model, and multivariable-adjusted model, respectively, were 0.338, 0.395, and 0.230. All models were additionally adjusted for sex.

Discussion

We derived a dietary pattern that reflected plasma fatty acid profiles using the RRR method. This dietary pattern predicted higher circulating n–3 PUFAs (ALA, 20:3n–3, EPA), odd-chain fatty acids (15:0, 17:0), LA, and 20:1, and lower AA and 16:1, and was characterized by higher intakes of soy, vegetables, fruits, tea, tomato products, bread, fish, margarine, and dairy products, along with lower intakes of rice, red meat, coffee, alcohol, sugar-sweetened beverages, and eggs. A higher score on this dietary pattern was associated with a substantially lower risk of T2D and CAD mortality.

A beneficial effect of the plasma fatty acid profile associated with the dietary pattern in our study is plausible. In previous studies, higher circulating ALA and EPA, from soy and fish, respectively (21), were associated with a lower CAD risk (12, 14–17). Diets high in n–3 PUFAs are cardioprotective (27–29) and may improve blood pressure, platelet aggregation response, and heart rate (30–33). The association between n–3 PUFAs and T2D risk is less consistent; n–3 PUFAs may suppress inflammation, enhance glucagon-like peptide 1 secretion, and improve insulin sensitivity by activating peroxisome proliferator-activated receptor-γ (34, 35). However, randomized controlled trials have not consistently demonstrated beneficial effects of n–3 supplementation on insulin sensitivity (36). In meta-analyses of observational studies, higher circulating ALA and higher ALA intake was associated with a modestly lower risk of T2D (37, 38); however, circulating EPA and DHA have not been consistently associated with T2D risk (37). An inverse association was reported between EPA and DHA intake and T2D in Asian populations (38, 39), but a direct association was reported in Western populations (37–39). However, in our Asian cohort, neither fish (40) nor marine n–3 intake (41) was significantly associated with T2D risk and these would not have substantially contributed to the inverse association between our dietary pattern and T2D risk. Higher concentrations of the major n–6 PUFA, LA, have been associated with a lower risk of T2D (11) and CAD (9, 10) in cohort studies, although a recent pooling analysis suggests that the association for CAD may be more modest than previously believed (20). Replacing saturated fat with dietary PUFAs significantly lowered CAD events in clinical trials, possibly through lowering serum cholesterol (42–44) and increasing insulin sensitivity (45).

Higher concentrations of the odd-chain saturated fatty acids 15:0 and 17:0 have been associated with lower T2D risk (13, 46) and higher 17:0 with lower CAD risk (12, 47) in other cohort studies. Dairy products, which were correlated with the dietary pattern, are likely to have contributed to the higher odd-chain fatty acid concentrations (48–50). Odd-chain fatty acids have lower melting points than even-chain homologs and may increase membrane fluidity (51, 52), which has been associated with a lower risk of T2D and CAD (53–56).

Our dietary pattern was linked to higher concentrations of 20:1, possibly due to vegetables and fish, which loaded highly on the pattern (57). Contrary to other very-long chain MUFAs 22:1 and 24:1, 20:1 was not associated with incident congestive heart failure in 2 US cohorts (57), and the implications of high circulating 20:1 on T2D remain largely unknown. Higher plasma 20:1 has been directly correlated with adiponectin concentrations (58) and 20:1 has been shown to act as a peroxisome proliferator-activated receptor-γ agonist in vitro (59).

Although our dietary pattern was derived to maximally explain variation in plasma fatty acids, the pattern was also characterized by foods that are not likely to substantially affect circulating fatty acids. This illustrates the intercorrelated nature of food intake and the dietary pattern may also have health benefits due to food components that affect nonfatty-acid-related pathways. For instance, isoflavones from soy may lower the risk of T2D (60–62). Vegetables and fruits are sources of micronutrients and phytochemicals that may be protective against CAD (63–65) including potassium, which has been reported to lower blood pressure in clinical trials (66, 67). Red meat intake, which was inversely correlated with our dietary pattern, may promote CAD through heme iron, saturated fatty acids, L-carnitine and for processed meat, nitrate preservatives (68–70). Heme iron is a strong pro-oxidant which may damage insulin-producing pancreatic cells through reactive oxygen species (40, 71–75). A lower intake of sugar-sweetened beverages may have also contributed to the inverse association between our dietary pattern scores and T2D and CAD risk (23, 76–79). In addition to individual foods and their components, synergy between different dietary components across foods may have contributed to the putative benefits of the identified dietary patterns for cardiometabolic health.

Previously reported results from the SCHS suggest that a combination of a high intake of unsweetened soy (80), temperate fruits (81), wholemeal bread (82), and dairy (83), and a low intake of red meat (40, 84) and sugar-sweetened beverages (23, 84) may have contributed to the inverse association between the dietary pattern and T2D. Although we previously reported an association between higher coffee consumption and lower risk of T2D (85), coffee loaded negatively on our dietary pattern. The lower consumption of coffee characteristic of the dietary pattern may have been compensated for by the more beneficial levels of the consumption of several other foods. The inverse association between our dietary pattern and CAD may result from the higher consumption of vegetables and fruits (86), fish (29), and lower consumption of red meat (86).

In a meta-analysis of prospective studies, a “prudent/healthy” or “mainly healthy” pattern was associated with a lower risk of T2D and CAD, respectively, and an “unhealthy/Western” or “mainly unhealthy” pattern with a higher risk (7, 87). Our dietary pattern shared several features with these “prudent/healthy” patterns including a higher intake of fish, legumes, vegetables, and fruits, and contrasted with the “unhealthy/Western” pattern with a lower intake of refined grains and red meat (7, 87). Unlike our dietary pattern, the patterns in that meta-analysis were derived from principle component analysis or factor analysis (7, 87). We identified additional foods that characterized our dietary pattern and are frequently consumed in Chinese populations, such as soy and tea. Indeed, RRR-derived dietary patterns can be population specific and irreproducible when applied to other populations (88, 89). This issue can be partly circumvented by calculating dietary pattern scores using the same weights and food items, or by using simplified scores (2, 88).

To our knowledge, no previous studies have identified dietary patterns using circulating fatty acids as response variables in RRR analysis. RRR based on nutrition biomarkers or serum metabolites can be expanded in the future as novel dietary biomarkers or metabolites are rapidly being identified through metabolomics studies (90, 91). For instance, in the EPIC-Potsdam study, the authors used RRR to derive dietary patterns that explain maximum variation in 127 metabolites from various classes including acylcarnitines, amino acids, and sphingomyelins (91). Using RRR with dietary biomarkers as response variables bypasses potential measurement errors relating to food composition data or bioavailability for different combinations of foods, as RRR simply identifies combinations of foods that maximally explain variation in these response variables (2). However, there are several considerations in the use of such RRR-based patterns. Our dietary pattern explained only 1.7% of the total variation in fatty acid biomarkers. This may be partly due to the long period between the dietary assessment and the collection of blood for biomarker measurement. We used total plasma fatty acids, which shows good reproducibility and correlation with selected dietary sources (92, 93). However, the choice of fatty acid biomarker may affect the result of RRR analyses using concentrations of fatty acids as response variables. Fatty acid compositions of different tissues and fractions may reflect pathways such as endogenous synthesis and conversion rather than dietary intake (48, 94–97). Indeed, our RRR method explained a larger proportion of fatty acids from dietary sources (e.g., 7.0% for ALA) than for fatty acids that can be endogenously derived (e.g., 0.0% for 16:0) (98).

The strengths of our study included the large sample size, the prospective design, and minimal loss to follow-up. Our study also has several potential limitations. First, dietary intakes were measured only at baseline, and changes in dietary intakes over time would have led to an attenuation of disease-risk estimates for the analysis of dietary patterns and disease outcomes. Measurement errors, resulting from self-reported food consumption, are inevitable but our validation study suggests that the assessment was reasonably accurate. Second, a substantial proportion of incident diabetes cases are likely to have been undiagnosed. However, this would only have affected the association between the dietary pattern and diabetes risk if the dietary pattern score was associated with the likelihood of being undiagnosed (99). Third, we did not use information on different cooking methods of foods in our RRR analysis. Fatty acids in food, particularly PUFAs, are susceptible to oxidation loss during cooking (100–103), although this has not been consistently demonstrated in experiments (104–106). Finally, we urge caution in generalizing our results to other ethnic groups.

We derived a dietary pattern that reflects plasma fatty acid concentrations using RRR analyses and this dietary pattern was associated with a lower risk of T2D and CAD in ethnic Chinese adults. We hypothesize these inverse associations were partly due to the higher circulating n–3 PUFAs, odd-chain fatty acids, and LA predicted by the dietary pattern. Our findings suggest that using fatty acid biomarkers or other serum metabolites as response variables in RRR may be a promising approach to identify dietary patterns related to chronic disease risk.

Notes

Supported by the National Institutes of Health, USA (R01 CA144034 and UM1 CA182876). JYHS is supported by the NUS Graduate School for Integrative Sciences and Engineering. W-PK is supported by the National Medical Research Council, Singapore (NMRC/CSA/0055/2013).

Author disclosures: JYHS, CNO, W-PK, J-MY, and RMvD, no conflicts of interest.

Supplemental Tables 1–3 are available from the “Supplementary data” link in the online posting of the article and from the same link in the online table of contents at https://academic.oup.com/jn/.

Abbreviations used: AA, arachidonic acid (20:4n–6); ALA, α-linolenic acid (18:3n–3); CAD, coronary artery disease; LA, linoleic acid (18:2n–6); RRR, reduced rank regression; SCHS, Singapore Chinese Health Study; T2D, type 2 diabetes.