Dietary Inflammatory Patterns Are Associated With Serum TGs and Insulin in Adults: A Community-Based Study in Taiwan

Abstract

Background

Dietary patterns related to inflammation have become a focus of disease prevention but the patterns may vary among populations.

Objectives

The study was conducted to determine Taiwanese dietary inflammatory patterns and evaluate their associations with biomarkers of lipid and glucose.

Methods

Data were taken from 5664 community-dwelling individuals aged ≥55 y recruited in 2009–2013 in the Healthy Aging Longitudinal Study in Taiwan (HALST). Dietary data were obtained from an FFQ. An empirical dietary inflammatory pattern (EDIP) was derived from reduced rank regression models that explained the serum high-sensitivity CRP, plasma IL-6, and TNF receptor 1. Cross-sectional associations between dietary scores and biomarkers of total cholesterol (TC); HDL cholesterol; LDL cholesterol; TG; and ratios of TG/HDL cholesterol, TG/TC, fasting glucose, insulin, and HbA1c were analyzed via multiple linear regression and adjusted for major confounders. The false-discovery rate (FDR)-adjusted P < 0.05 was considered statistically significant. Abdominal obesity was defined as a waist circumference of ≥90 cm for men and ≥80 cm for women.

Results

Higher EDIP-HALST scores were associated with higher TG (per score increment: 1.62%, 95% CI: 0.58%, 2.76%; P_FDR = 0.01), TG/HDL cholesterol (2.01%, 95% CI: 0.67%, 3.37%; P_FDR = 0.01), and TG/TC (1.42%, 95% CI: 0.41%, 2.43%; P_FDR = 0.01) and nonlinearly associated with insulin, with those in the middle tertile had the highest serum insulin concentrations (means: 5.12 μIU/mL, 95% CI: 4.78, 5.78; P_FDR = 0.04) in men, but not in women. No heterogeneity was detected between sexes. The associations with TG (1.23%, 95% CI: 0.19, 2.23%; P_trend = 0.02), TG/HDL cholesterol (1.62%, 95% CI: 0.30%, 2.96%; P_trend = 0.02), and TG/TC (1.11%, 95% CI: 0.11%, 2.13%; P_trend = 0.03) were stronger in participants with abdominal obesity, but were borderline associated in participants with normal abdominal circumferences (all P_trend = 0.05).

Conclusions

Inflammatory diets, as measured via EDIP-HALST, are associated with serum TG concentration, particularly in participants with abdominal obesity. These findings may suggest that developing disease prevention strategies using dietary inflammatory patterns may be different by populations.

J Nutr 20xx;x:xx.

Introduction

Dietary patterns related to inflammation have become a focus of disease prevention but the patterns may vary among populations. Low-grade systemic inflammation usually originates from obesity-related metabolic dysfunction; together, disrupted metabolism and inflammation increase the risk of cardiometabolic diseases [1,2]. Circulating lipids and glucose are important intermediate biomarkers of cardiometabolic health. The diet significantly affects the risk of multiple chronic health conditions [3]. Because of the close relationship of nutrients originating from a single food and the combination of foods within dietary patterns, the importance of dietary patterns in research and public health is increasingly recognized [[4], [5], [6]]. Understanding how inflammatory-related dietary patterns affect cardiometabolic biomarkers would strengthen the application of dietary patterns in disease prevention.

Dietary patterns with abundant fruits, vegetables, whole grains, fish, and vegetable oils and that are low in saturated fat and processed meats, such as in Mediterranean or Nordic diets, have been shown to decrease inflammatory biomarkers and are associated with lower risk of metabolic syndrome and CVDs [7,8]. Several dietary indices have been created to evaluate the dietary inflammatory potential [9]. Shivappa et al. [10] reviewed thousands of articles and developed an algorithm, the dietary inflammatory index (DII), to categorize individuals’ diets from maximally anti-inflammatory to maximally proinflammatory. Tabung et al. [11] developed the empirical dietary inflammatory pattern (EDIP) based on the data-driven approach of reduced rank regression (RRR). RRR is a statistical method that identifies a linear function of predictors (for example, food groups) that explains the most variation in a set of response variables (for example, biomarkers) [12,13]. The DII is associated with higher risks of myocardial infarction, all-cause mortality, and overall and site-specific cancer incidences (colorectal, pancreatic, respiratory, and oral cancers) [14]. EDIP is also related to type 2 diabetes [15,16], CVDs [17], and cancers [[18], [19], [20]]. The major difference between these 2 DIIs is that the DII is calculated from nutrients, whereas the EDIP is calculated from food groups. In a community setting, food is more efficient in health communication with an ordinary lay person; hence, EDIP is preferred. However, similar to other a posteriori methods, the EDIP, which was derived from a data-driven approach, is specific to the original study population; thus, making the associations between the EDIP and inflammation or EDIP and disease risks less reproducible in other populations than in the populations that shared similar eating habits with the original study population [13,21,22].

Dietary patterns, which are usually referred to the quantities, proportions, types, or combinations of different foods and drinks in diets, and how often they are consumed [3,23], could vary among populations with different cultures and eating habits, as does inflammatory dietary patterns [24]. This study was conducted to construct a dietary inflammatory score based on the dietary data from the Healthy Aging Longitudinal Study in Taiwan (HALST) and to explore the association of the dietary score with a panel of biomarkers of lipid and glucose. We also examined whether obesity modifies the association between inflammatory dietary patterns and these biomarkers.

Methods

The HALST

The HALST is a prospective study of community-dwelling older adults in which 5664 volunteers were recruited across Taiwan between 2009 and 2013. The cohort has been documented previously [25]. Briefly, a sample of eligible residents (≥55 y) living within the catchment area of 7 collaborative hospitals was recruited for the study (n = 22,563). Among them, 6985 (31%) subjects agreed to participate. Participants with any of the following conditions were further excluded: highly contagious infectious diseases, diagnosed dementia, severe illness (based on the interviewers’ judgment of whether the participant was too ill to complete the interview), being bed-ridden, severe mental disorders, mutism, hearing impairment, blindness, or other conditions such as living in a long-term care facility or being hospitalized. Interviewers were trained to conduct face-to-face interviews using participants’ primary languages, which include Mandarin, Taiwanese, and Hakka.

All participants signed written informed consent. The Institutional Review Board at the National Health Research Institutes and the collaborative hospitals approved the study. All methods were performed in accordance with the relevant guidelines and regulations.

Measurements of biomarkers

Biological samples were collected according to a standard protocol. Briefly, fasting blood was collected and analyzed at a certified clinical laboratory. All blood samples were centrifuged, aliquoted, and stored in a −80°C freezer at the National Health Research Institutes after routine standardization and calibration tests.

Inflammatory markers

We used an immunoturbidimetric assay with the ADVIA 1800 Chemistry System to measure serum levels of high-sensitivity CRP; the inter- and intraassay coefficients of variation (CVs) were 10.8% and 5.6%, respectively. Plasma levels of TNF receptor 1 (TNFR1) were measured using an immunoassay kit (Human sTNF RI/TNFRSF1A Immunoassay, R&D Systems Inc.); the inter- and intraassay CVs were 7.4%–11.0% and 9.6%, respectively. Plasma levels of IL-6 were measured using a high-sensitivity ELISA (Quantikine HS ELISA Human IL-6 Immunoassay, R&D Systems) per the manufacturer’s instructions; the inter- and intraassay CVs were 10.3%–15.3% and 16.6%, respectively.

An overall inflammatory marker score was created by summing the z-scores of the above-mentioned log-transformed inflammatory biomarkers: z-score (log high-sensitivity CRP) + z-score (logTNFR1) + z-score (logIL-6) [11,26].

Lipid profiles

Serum lipids, including total cholesterol (TC), LDL cholesterol, HDL cholesterol, and TGs, were assayed in a certified laboratory. TC was measured using an enzymatic method on the ADVIA 1800 system; HDL cholesterol and LDL cholesterol were measured using elimination/catalase methods on the ADVIA 1800 system; and TGs were determined using the enzymatic GPO method on the ADVIA 1800 system. TG/HDL cholesterol and TG/TC ratios were also estimated.

Glycemia measures

Serum levels of insulin were measured via a chemiluminescence method on Beckman Access II-Double Antibody; glucose was measured using a hexokinase method on the ADVIA 1800 system. HbA1c levels were measured using high-performance liquid chromatography on the Tosoh G8 HPLC Analyzer.

FFQ

All participants were asked to complete a 72-item FFQ to estimate their average daily dietary intake over the past year. The original FFQ was developed for Chinese Americans and adapted for another validation study in the Taiwanese population [27,28]. Briefly, validation of the FFQ was conducted against a 1-d recall. Correlation coefficients of nutrient intakes from both methods ranged from 0.2 for total fat to 0.7 for calcium. In addition, agreements in quartile distributions between the 2 methods suggested that 50% agreed in the highest quartile, whereas only 10% might be misclassified [27,28]. Average consumptions (g/d) were estimated as the sum of the product of eating frequency, portion size, and energy and nutrient contents in each food referenced from the Food Composition Tables in Taiwan (March 2017). Energy intakes >5000 kcal/d for men, >4500 kcal/d for women, or <500 kcal/d for both men and women were deleted from the data [27,28] because the intakes were considered to be over- or under-reported (n = 79).

Development of the EDIP-HALST and its performance in associations with inflammation

The goal of using the EDIP was to create a score using food groups to assess the overall inflammatory potential of whole diets [11]. The major problem with the EDIP is its lack of reproducibility in different populations [13,21]. To address this issue, we arbitrarily used the data from participants at 4 of the 7 HALST sites (n = 3079) as the exploratory sample, and the data from the remaining participants (n = 2505) as the confirmatory sample. Supplemental Table 1 shows the baseline characteristics of the exploratory and confirmatory samples. In brief, the distributions of age and sex were similar in both samples; however, the inflammatory marker score was higher and the DII score based on Shivappa et al. [10] was lower in the exploratory sample. To reduce the influence of medications, we excluded participants who regularly used analgesics (n = 529) from the score development. Seventy-two food items were grouped into 27 food groups (Supplemental Table 2) according to their nutrient compositions. The mean intakes of these 27 food groups were calculated from the FFQ and adjusted for energy using the residual method [29]. Z-standardized energy-adjusted intakes of the food groups were then applied to RRR models to identify a dietary pattern most associated with 3 circulating inflammatory biomarkers: high-sensitivity CRP, TNFR1, and IL-6. Three scores were considered and the performance was compared: linear weighted [11], loading weighted, and unweighted [30,31]. For the linear weighted scores, the first factor obtained from the RRR was examined via stepwise linear regression analysis (P > 0.10 for entry and removal from the model) to identify the most important food groups contributing to this pattern. Regression coefficients from the final step of the stepwise linear regression analysis were used as the component weights. The weighted scores were calculated by summing the z-standardized intakes of the food groups with RRR loading >|0.20| weighted with the loading, and the unweighted score was the unweighted sum [31].

In the confirmatory phase, coefficients of the linear weighted, loading weighted, and unweighted scores obtained from the exploratory sample (Supplemental Table 3) were applied to the confirmatory sample, and then we calculated the linear associations among the dietary scores and the construct validators of the scores, that is, high-sensitivity CRP, TNFR1, IL-6, and the overall inflammatory marker score. The RRR models from our data suggested that higher intake of dark and light green vegetables and whole grains were anti-inflammatory, and nuts and Chinese bao were proinflammatory in men, whereas higher intakes of fruits, soymilk, and whole grains were anti-inflammatory, and refined grains were proinflammatory in women. More food groups were identified when fitting the RRR factor scores to the food groups using the stepwise linear regression and were used to calculate the linear weighted scores.

Besides, the DII score according to Shivappa et al. [10,32] was also calculated as a comparison. To evaluate the performance of the dietary scores, we performed multivariable-adjusted linear regression analyses. The dependent variables were the log-transformed biomarker concentrations, and the independent variables were the tertiles of the dietary scores modeled as ordinal variables. The adjusted means were then back-transformed to estimate the concentrations in each tertile of the dietary score. The models were adjusted for age at recruitment (continuous), center, education (low literacy, primary school, middle school and above), smoking status (never, former, and current), and physical activity at leisure time and at work (sex-specific quartile) [[33], [34], [35]]. In addition, previous studies have identified that BMI may be a mediator between inflammatory diets and biomarkers [11]. Therefore, sensitivity analyses were conducted by excluding or including BMI as a covariate in the model, stratifying by BMI (<27 or ≥27) [36], or restricting the analysis to those with optimal BMI in older people (18.5 ≤ BMI < 30) [37]. To analyze the linear trends of the serum inflammatory markers across the dietary scores, we used the continuous dietary scores adjusted to the covariates and interpreted the P value as the P-trend. Briefly, the linear weighted scores showed a statistically significant P-trend with the 3 markers examined in both men and women; thus, the score was used for the following analyses, termed the EDIP-HALST (Supplemental Table 4).

Measurement of covariates

A number of covariates, including age, sex, education level (low literacy, primary school, or more than primary school), smoking, drinking, physical activities, height, weight (for calculating BMI), waist circumference, and treatments for hypertension, dyslipidemia, and diabetes, were selected as potential confounders based on their known associations with inflammation, metabolic disorders, and CVDs from prior literature [33,35]. The study center was included as a surrogate for urbanization. Most of the covariates were self-reported, except height, weight, and waist circumference, which were measured during the physical examination by the trained interviewers.

Statistical analysis

Participants’ characteristics were expressed as counts (%) and compared using chi-square tests for categorical variables or described as means ± SDs and compared via the Kruskal-Wallis test for continuous variables.

Data from the exploratory and confirmatory samples were pooled to analyze the associations between EDIP-HALST scores and the biomarkers of interest. Tertiles of the EDIP-HALST scores were categorized according to the distribution by sex. The associations between dietary scores and biomarkers were analyzed using multiple linear regression. To reduce the influences of drugs, we excluded diabetes treatments (n = 925) when analyzing the associations with glycemic measures and dyslipidemia treatments (n = 815) when analyzing data for lipids. Sex differences were noted in previous studies on dietary patterns, metabolic parameters, and obesity [38,39]. Thus, the analyses were separated by sex and adjusted for age at recruitment (continuous), center, education (low literate, elementary school, middle school and above), smoking status (never, former, and current), alcohol consumption status (never, former, and current), physical activity (exercise and work-related combined), BMI (<18.5, 18.5 to <25, 25 to <30, and ≥30 kg/m²), and treatments for hypertension (yes or no), and dyslipidemia (yes or no) for analyses on glycemia measures, or diabetes (yes or no) for analyses on lipids. Heterogeneity between sexes was examined by the likelihood ratio test. P-trend was obtained by modeling the median value of each tertile as continuous variables. A nonlinear relationship between dietary scores and biomarkers was tested by adding a square term of the dietary scores into the model. Because no sex differences were detected in the current study, men and women were combined to further explore the potential modification effect of obesity. In Taiwan, abdominal obesity is defined as a waist circumference ≥90 cm for men and ≥80 cm for women [36].

For multiple comparisons, the false-discovery rate (FDR)-adjusted P< 0.05 was considered statistically significant. All analyses were performed using SAS 9.4.

Results

Table 1 presents participants’ characteristics by EDIP-HALST score tertiles. Overall, men and women at the higher EDIP-HALST tertiles were older, less educated, and less physically active.

TABLE 1.

Participant characteristics at baseline in tertiles of EDIP-HALST score

	Men							Women							Total
	T1 (n = 874)		T2 (n = 875)		T3 (n = 875)		P value	T1 (n = 985)		T2 (n = 986)		T3 (n = 986)		P value	T1 (n = 1859)		T2 (n = 1861)		T3 (n = 1861)		P value
	Mean	SD	Mean	SD	Mean	SD		Mean	SD	Mean	SD	Mean	SD		Mean	SD	Mean	SD	Mean	SD
Age (y)	67.65	7.66	70.51	8.50	71.79	8.76	<0.01	68.17	7.86	69.49	7.94	70.42	3.06	<0.01	67.92	7.77	69.97	8.22	71.07	8.42	<0.01
BMI (kg/m2)	24.64	3.16	24.52	3.27	24.33	3.48	0.07	24.27	3.57	24.76	3.69	24.70	3.72	<0.01	24.44	3.39	24.65	3.50	24.52	3.61	0.17
Abdominal circumference (cm)	87.12	9.09	87.29	8.94	87.30	9.83	0.92	84.49	10.93	86.22	11.04	86.16	11.01	<0.01	85.73	10.18	86.73	10.11	86.70	10.48	0.00
Energy intake (kcal/d)	2368.15	769.34	2099.19	701.87	2370.78	849.71	<0.01	1826.71	642.18	1625.76	585.04	1929.95	621.21	<0.01	2081.27	754.70	1848.36	684.55	2137.22	769.46	<0.01
Education levels	n	%	n	%	n	%		n	%	n	%	n	%		n	%	n	%	n	%
Low literacy	15	1.7	32	3.7	41	4.7	<0.01	68	6.9	181	18.4	275	27.9	<0.01	83	4.5	213	11.5	316	17.0	<0.01
Primary school	276	31.6	381	43.6	385	44.0		414	42.2	516	52.3	508	51.5		690	37.2	897	48.2	893	48.0
More than primary school	582	66.7	461	52.7	449	51.3		500	50.9	289	29.3	203	20.6		1082	58.3	750	40.3	652	35.0
Missing	1		1		0			3		0		0			4		1		0
Smoking
Never	396	45.3	356	40.7	366	41.8	0.02	972	98.7	960	97.4	957	97.1	0.06	1368	73.6	1316	70.7	1323	71.1	0.02
Former	246	28.1	309	35.3	300	34.3		4	0.4	11	1.1	7	0.7		250	13.4	320	17.2	307	16.5
Current	232	26.5	210	24.0	209	23.9		9	0.9	15	1.5	22	2.2		241	13.0	225	12.1	231	12.4
Alcohol consumption
Never	315	36.0	342	39.0	369	42.2	<0.01	755	76.6	811	82.3	816	82.8	<0.01	1070	57.6	1153	62.0	1185	63.7	<0.01
Former	120	13.7	185	21.1	182	20.8		32	3.2	36	3.7	27	2.7		152	8.2	221	11.9	209	11.2
Current	439	50.2	348	39.8	324	37.0		198	20.1	139	14.1	143	14.5		637	34.3	487	26.2	467	25.1
Physical activity1
Low	240	27.6	293	33.6	336	38.7	<0.01	261	26.6	342	35.2	374	38.2	<0.01	501	27.1	635	34.4	710	38.4	<0.01
Median	310	35.6	289	33.1	273	31.4		362	36.9	318	32.7	297	30.4		672	36.3	607	32.9	570	30.9
High	321	36.9	290	33.3	260	29.9		357	36.4	312	32.1	307	31.4		678	36.6	602	32.7	567	30.7
Missing	3		3		6			5		14		8			8		17		14
Treatment for diabetes	160	18.3	137	15.7	160	18.3	0.24	147	14.9	162	16.4	159	16.1	0.63	307	16.5	299	16.1	319	17.1	0.68
Treatment for hypertension	340	38.9	350	40.0	389	44.5	0.04	391	39.7	423	42.9	412	41.8	0.34	731	39.3	773	41.5	801	43.0	0.07
Treatment for dyslipidemia	115	13.2	109	12.5	107	12.2	0.83	167	17.0	173	17.6	144	14.6	0.18	282	15.2	282	15.2	251	13.5	0.25

Abbreviations: EDIP, empirical dietary inflammatory pattern; HALST, the Healthy Aging Longitudinal Study in Taiwan.

¹Leisure time and work-related combined and were categorized by sex-specific tertile.

Table 2 shows the association between EDIP-HALST scores and serum lipid biomarkers for participants not undergoing dyslipidemia treatments. In men, higher scores were associated with higher serum TG (per score increment: 1.62%, 95% CI: 0.58%, 2.76%, P_trend < 0.01, FDR P = 0.01), TG/HDL cholesterol (2.01%, 95% CI: 0.67%, 3.37%, P_trend < 0.01, FDR P = 0.01), and TG/TC (1.42%, 95% CI: 0.41%, 2.43%, P_trend < 0.01, FDR P = 0.01). However, no association was observed in women.

TABLE 2.

Multivariable-adjusted¹ mean concentration (95% CI) of serum lipid biomarkers in tertiles of EDIP-HALST scores among participants who did not take dyslipidemia medications

	T1		T2		T3		Per unit increment			P-trend2	FDR P value
	Mean1	95% CI	Mean1	95% CI	Mean1	95% CI	Mean1	95% CI
Men, n	759		766		768
Total cholesterol (mg/dL)	177	(173, 182)	179	(175, 184)	180	(175, 184)	0.44	(−0.29, 1.18)		0.26	0.29
HDLC (mg/dL)	46.2	(44.8, 47.7)	45.0	(43.7, 46.4)	45.4	(44.1, 46.8)	−0.39%	(−0.87%, 0.09%)		0.11	0.16
LDLC (mg/dL)	109	(105, 114)	110	(106, 114)	112	(107, 116)	0.40	(−0.27, 1.06)		0.29	0.29
TGs (mg/dL)	97.2	(90.9, 104)	104	(97.5, 111)	105	(98.5, 112)	1.62%	(0.58%, 2.67%)		<0.01	0.01
TGs/HDLC	2.10	(1.93, 2.29)	2.32	(2.14, 2.52)	2.31	(2.13, 2.50)	2.01%	(0.68%, 3.37%)		<0.01	0.01
TGs/total cholesterol	0.557	(0.522, 0.594)	0.595	(0.558, 0.633)	0.596	(0.560, 0.634)	1.42%	(0.41%, 2.43%)		<0.01	0.01
Women, n	818		813		842
Total cholesterol (mg/dL)	191	(184, 199)	191	(183, 198)	190	(183, 198)	−0.22	(−1.02, 0.59)		0.58	0.58
HDLC (mg/dL)	51.6	(49, 54.2)	50.9	(48.4, 53.5)	50.7	(48.2, 53.3)	−0.39%	(−0.90%, 0.14%)		0.15	0.22
LDLC (mg/dL)	117	(110, 124)	117	(110, 124)	116	(109, 123)	−0.30	(−1.04, 0.44)		0.41	0.49
TGs (mg/dL)	110	(99.1, 122)	113	(102, 125)	115	(104, 128)	0.97%	(−0.13%, 2.08%)		0.08	0.16
TGs/HDLC	2.14	(1.86, 2.45)	2.21	(1.93, 2.53)	2.27	(1.99, 2.60)	1.36%	(−0.06%, 2.80%)		0.06	0.16
TGs/total cholesterol	0.586	(0.529, 0.650)	0.601	(0.543, 0.665)	0.616	(0.557, 0.682)	1.09%	(0.01%, 2.17%)		0.04	0.16

Abbreviations: EDIP, empirical dietary inflammatory pattern; FDR, false-discovery rate; HALST, the Healthy Aging Longitudinal Study in Taiwan.

¹Means were adjusted for age at recruitment (continuous), center, education (low literacy, primary school, middle school and above), smoking (never, former, and current), alcohol consumption (never, former, and current), physical activity at leisure time and at work (sex-specific quartile), BMI (<18.5, 18.5 to <25, 25 to <30, and ≥30 kg/m²), treatment for hypertension (yes and no), and diabetes (yes and no). The dependent variables were log-transformed before fitting a multiple linear regression, except for the total cholesterol and LDLC. The adjusted means were then calculated by exponentiating the adjusted natural-log transformed means. The adjusted means per unit increment were calculated by exponentiating the adjusted natural-log transformed means minus 1 and multiplied by 100, thus should be interpreted as “relative change.”

²The P value for the linear trend was calculated by modeling the median values of each category as continuous variables and adjusted for all covariates listed in footnote 1.

In Table 3, EDIP-HALST scores were nonlinearly associated with serum insulin in men, with the middle tertile having the highest concentrations of serum insulin (adjusted means: 5.12 μIU/mL, 95% CI: 4.78, 5.78, P_quadratic = 0.01, FDR P = 0.04). Again, there was no association in women.

TABLE 3.

Multivariable-adjusted¹ mean concentration (95% CI) of serum biomarkers of glycemia measures in tertiles of EDIP-HALST scores among participants who did not take diabetic medications

	Men								Women
	T1 (n = 714)		T2 (n = 738)		T3 (n = 715)		P-quadratic2	FDR P value	T1 (n = 838)		T2 (n = 824)		T3 (n = 827)		P-quadratic2	FDR P value
	Mean	95% CI	Mean	95% CI	Mean	95% CI			Mean	95% CI	Mean	95% CI	Mean	95% CI
HbA1c (%)	5.91	(5.83, 6.00)	5.89	(5.81, 5.98)	5.88	(5.80, 5.96)	0.90	0.99	6.03	(5.90, 6.16)	6.01	(5.89, 6.14)	6.03	(5.90, 6.16)	0.49	0.49
Fasting glucose (mg/dL)	104	(102, 106)	103	(101, 105)	103	(101, 105)	0.99	0.99	104	(101, 107)	104	(101, 107)	104	(101, 108)	0.48	0.49
Insulin (μIU/mL)	4.70	(4.38, 5.05)	5.12	(4.78, 5.48)	4.86	(4.54, 5.20)	0.01	0.04	5.02	(4.49, 5.62)	5.26	(4.71, 5.87)	5.18	(4.64, 5.78)	0.25	0.49

Abbreviations: EDIP, empirical dietary inflammatory pattern; FDR, false-discovery rate; HALST, the Healthy Aging Longitudinal Study in Taiwan.

¹Means were adjusted for age at recruitment (continuous), center, education (low literacy, primary school, middle school and above), smoking (never, former, and current), alcohol consumption (never, former, and current), physical activity at leisure time and at work (sex-specific quartile), BMI (<18.5, 18.5–25, 25–30, and ≥30 kg/m²), treatment for dyslipidemia (yes and no), and hypertension (yes and no). The dependent variables were log-transformed before fitting a multiple linear regression. The adjusted means were then calculated by exponentiating the adjusted natural-log transformed means.

²The P value for the linear trend was calculated by modeling the median values of each category as a continuous variable and adjusted for all covariates listed in footnote 1.

Nevertheless, none of the examined associations showed heterogeneity between sexes; therefore, we combined men and women to examine the effect modification of abdominal obesity. Higher EDIP-HALST scores were associated with higher serum TG (1.23%, 95% CI: 0.19%, 2.23%, P_trend = 0.02), TG/HDL cholesterol (1.62%, 95% CI: 0.30%, 2.96%, P_trend = 0.02), and TG/TC (1.11%, 95% CI: 0.11%, 2.13%, P_trend = 0.03), particularly in participants with abdominal obesity (Table 4). Conversely, the associations were weaker in participants with nonobese abdominal circumferences.

TABLE 4.

Mean serum lipids by EDIP-HALST tertiles stratified by abdominal obesity¹ in men and women combined

	T1		T2		T3		Per unit increment			P-trend3	P_heterogeneity4 (obesity vs. normal)
	Mean2	95% CI	Mean2	Mean1	Mean1	95% CI	Mean1	95% CI
Normal, n	738		678		692
Total cholesterol (mg/dL)	191.33	(191, 187)	196	(189, 185)	193	(192, 188)	0.01	(−0.77, 0.80)		0.97	0.04
HDLC (mg/dL)	50.77	(50.8, 49.4)	52.2	(49.6, 48.2)	51	(49.9, 48.5)	−0.38%	(−0.90%, 0.14%)		0.15	0.55
LDLC (mg/dL)	116.43	(116, 113)	120	(114, 110)	117	(117, 113)	0.00	(−0.70, 0.71)		0.99	0.01
TGs (mg/dL)	99.37	(99.4, 93.6)	105	(102, 96.4)	108	(105, 98.9)	1.09%	(−0.02%, 2.20%)		0.05	0.74
TGs/HDLC	1.97	(1.97, 1.82)	2.12	(2.07, 1.92)	2.24	(2.10, 1.95)	1.43%	(−0.01%, 2.89%)		0.05	0.82
TGs/total cholesterol	0.53	(0.533, 0.504)	0.565	(0.556, 0.526)	0.588	(0.561, 0.531)	1.08%	(0.01%, 2.16%)		0.05	0.95
Abdominal obesity, n	744		776		756
Total cholesterol (mg/dL)	187.49	(187, 184)	191	(190, 187)	194	(189, 185)	0.29	(−0.45, 1.03)		0.44
HDLC (mg/dL)	46.92	(46.9, 45.9)	48	(46.4, 45.4)	47.5	(46.1, 45.1)	−0.36%	(−0.83%, 0.12%)		0.14
LDLC (mg/dL)	115.76	(116, 112)	119	(118, 115)	121	(116, 113)	0.21	(−0.48, 0.90)		0.55
TGs (mg/dL)	117.16	(117, 112)	123	(122, 117)	128	(124, 119)	1.23%	(0.19%, 2.27%)		0.02
TGs/HDLC	2.50	(2.50, 2.35)	2.66	(2.65, 2.49)	2.81	(2.70, 2.54)	1.62%	(0.30%, 2.96%)		0.02
TGs/total cholesterol	0.64	(0.637, 0.607)	0.668	(0.657, 0.627)	0.688	(0.672, 0.642)	1.11%	(0.11%, 2.13%)		0.03

Abbreviations: EDIP, empirical dietary inflammatory pattern; HALST, the Healthy Aging Longitudinal Study in Taiwan.

¹Waist circumference ≥90 cm for men and ≥80 cm for women.

²Means were adjusted for age at recruitment (continuous), sex, center, education (low literacy, primary school, middle school and above), smoking (never, former, and current), alcohol consumption (never, former, and current), physical activity at leisure time and at work (sex-specific quartile), treatment for hypertension (yes and no), and diabetes (yes and no). The dependent variables were log-transformed before fitting a multiple linear regression, except for the total cholesterol and LDLC. The adjusted means were then calculated by exponentiating the adjusted natural-log transformed means. The adjusted means per unit increment were calculated by exponentiating the adjusted natural-log transformed means minus 1 and multiplied by 100, thus should be interpreted as “relative change.”

³The P value for the linear trend was calculated by modeling the median values of each category as a continuous variable and adjusted for all covariates listed in footnote 1.

⁴Heterogeneity was examined using the likelihood ratio test.

Discussion

In this study, we used RRR, stepwise linear regression, a combination of biostatistical approaches, and inflammatory biomarkers from blood samples to evaluate complex patterns of food intake with inflammatory potential and created the EDIP-HALST score. We revealed that higher EDIP-HALST scores reflected dietary patterns related to higher chronic systemic inflammation and higher serum TG, particularly in participants with abdominal obesity. EDIP-HALST scores showed a nonlinear relationship with serum insulin, with the middle tertile of EDIP-HALST having the highest means of insulin concentrations.

The original version of the EDIP has shown promising results in United States populations [11,40,41), but its associations with inflammatory biomarkers could not be replicated in a Brazilian population; the EDIP components for Sao Paulo, Brazil were processed meat, fruits and vegetables, and rice and beans instead [21]. Supplemental Table 5 summarizes the data-driven inflammatory dietary patterns from different countries. Similar but different inflammatory dietary patterns across studies were noted. For example, red meat was proinflammatory in Sweden [42], the United Kingdom [43], and the United States [11]; however, the correlation/loading of red meat was low in Korea [44] and in our study. We suspected that the reason of the low correlation/loading in Taiwan might be because quantity of red meat consumed in our population was lower than that in other countries. The total red meat intake was 1.10 servings (1 serving ≈ 84 g) per day in the Health Professional Follow-up Study and 0.92 servings per day in the Nurse’s Health study [45]. These 2 cohorts were used to develop and validate the original EDIP score [11]. The mean red meat intake, including beef and pork, in the HALST study was only 50.5 g for men and 34.5 g for women. Moreover, even within Western countries, legumes seemed to be anti-inflammatory in Sweden [42] but proinflammatory in the United Kingdom [43]. The authors commented that the baked beans were usually consumed with processed meat and other fried food in English-style breakfast; thus, the inflammatory dietary pattern score was not generalizable [43]. Given that the Taiwanese are a rice-eating population, selection of whole-grain rice, more fruits and vegetables, and soymilk, and avoiding refined grain rice were all associated with reduced inflammation in both men and women (Supplemental Table 3). The proinflammatory properties of Chinese bao and pancakes may be related to the culinary culture, including the added ingredients, oil, and condiments. Nuts were associated with higher inflammation in men, possibly because of seasonings such as salt and the amount eaten.

A Taiwanese study [46] identified that dietary patterns associated with high-sensitivity CRP and neutrophil-to-lymphocyte ratios included lower intake of seafood, grains, vegetables, and fruits and higher intake of meat, eggs, preserved/processed foods, and sugary drinks. Similar to our findings, that study also showed that dietary patterns were associated with an increased risk of dyslipidemia in men, but not in women. Researchers in Korea [47] found that a pattern of “whole grain and soybean products” was related to decreased risks of hypercholesterolemia and hypertriglyceridemia after a 14-y follow-up. More interestingly, the association was more evident in the genetically defined high-risk population with dyslipidemia. Most healthy dietary patterns, for example, Mediterranean or DASH diets, as well as the DII and the original version of the EDIP, are associated with healthier lipid profiles [[48], [49], [50], [51]]. To lower TG, reduce CVD risks, and prevent coronary plaque progression, intakes of fruits and vegetables, whole grains, low-fat dairy, and fish or other ω-3 fatty acids should be increased and saturated fats and refined carbohydrates should be minimized [52]. This guidance also fits the dietary patterns of lower EDIP-HALST scores.

When stratified by abdominal obesity, a surrogate for body fat accumulation and a state of chronic inflammation, the associations between EDIP-HALST scores and TG, TG/HDL cholesterol, and TG/TC were more evident in participants with abdominal obesity and weaker in participants without abdominal obesity. This suggests a direct relationship between dietary inflammatory patterns in TG levels, which can be exaggerated when another stressor (for example, obesity) is present. Another possible explanation would be that the background inflammation (for example, obesity) was not strong enough that the dietary challenge was not sufficient to induce inflammatory responses in this population [24]. Because dyslipidemia is a well-established risk factor for major chronic diseases (for example, CVDs), EDIP-HALST scores might be a promising dietary strategy for disease prevention.

Previous studies have suggested that increased TG can increase free fatty acids for oxidation and thus can increase inflammation while decreasing insulin sensitivity. Blood glucose and free fatty acids can further stimulate the pancreas to produce more insulin. β-Cell dysfunction can result from increased insulin resistance [53,54]. TG/HDL cholesterol is a marker of insulin resistance [55]. However, the association between EDIP-HALST scores and insulin was not linear as it is with TG/HDL cholesterol, and the dietary score was poorly associated with fasting glucose in our study. We previously reported that, compared with participants whose HbA1c levels were 5.5%–6%, participants with HbA1c levels <5.5% or ≥7.0% had higher sIL-6r and TNFR1 levels [56]. The EDIP-HALST score was well correlated with circulating TNFR1 and IL-6 levels (Supplemental Table 4), although it was unrelated to HbA1c. Insulin concentrations generally depend on β-cell responses to glucose. Whether proinflammatory diets affect β-cell health warrants further investigation.

Previous studies suggested sex differences between dietary patterns and metabolic parameters and obesity [38,39]. In our study, the association between EDIP-HALST scores and TGs and insulin was significant in men but not statistically significant in women. The prevalence of smoking in our study was higher in men than in women (25% in men vs. 2% in women). Compared with women, men might experience more inflammatory stress; thus, this association is more common in men. However, we adjusted for smoking, alcohol consumption, physical activity, and BMI in our model, and no heterogeneity was detected in the association between EDIP-HALST scores and TG and insulin between sexes. More in-depth investigations on this issue are advised.

This study had some limitations. First, the FFQ included response bias and measurement errors [57]. However, the FFQ was found to have good repeatability and reasonable validity for most, but not all, nutrients in older adults [58]. Nevertheless, inflammatory dietary patterns may have been misclassified because of misreporting; thus, attenuated associations are likely. Second, the EDIP-HALST scores were only validated within our own cohort. The response rate of the original cohort was only 31%. The generalizability of the study results to the other population is needed to be evaluated carefully. These scores should be evaluated with a wider spectrum of other inflammatory biomarkers and validated in other Taiwanese populations. Third, residual confounding may have occurred, although we have adjusted for well-known documented risk factors. Finally, the study’s cross-sectional design makes it difficult to draw a conclusion on temporal direction or infer causality. However, we hypothesized that a more inflammatory diet may disrupt lipid homeostasis and insulin resistance and elevate the inflammatory environment, which may further predispose patients to adverse health outcomes such as CVDs and cancer.

The major strengths of this study include the relatively large population of community-dwelling participants and the detailed data regarding lifestyles, such as education level, smoking and alcohol consumption behaviors, comorbidities, and medications. This enabled us to explore associations independent of these factors. Both sexes were equally represented in the cohort, allowing examination of gender differences between men and women. In addition, the cohort provided a wide range of biomarkers, including lipids, glucose homeostasis, and clinical health parameters.

In summary, our results showed that inflammatory diets, as measured by the EDIP-HALST, were associated with TG, particularly in participants with abdominal obesity. The nonlinear association between EDIP-HALST scores and insulin deserves further investigation. These findings suggest that developing disease prevention strategies using dietary inflammatory patterns may be different by populations.

Author contributions

The authors’ responsibilities were as follows – S-CC, I-CW, and CAH: designed research; C-WC and H-LC: conducted research; S-CC, CAH, and C-CH: provided essential materials; S-CC and H-TC: analyzed data; all authors: wrote paper; S-CC: had primary responsibility for the final content; and all authors: read and approved the final manuscript.

Funding

This work was supported by the National Health Research Institutes in Taiwan (grant numbers PH-110-SP-01, PH-110-PP-12).

Author disclosures

The authors report no conflicts of interest.

Appendix A. Supplementary data

The following is the Supplementary data to this article: