The association between diet quality, plant-based diets, systemic inflammation, and mortality risk: findings from NHANES

This study used the publicly available data from the National Health and Nutrition Examination Survey (NHANES) [22] in the United States. The characteristics and design of the study have been described in detail previously [23]. Participants with missing data and those with CRP levels > 10 mg/L (indicating acute inflammation) [9, 24] were excluded from the study. A total of 27,915 participants from NHANES cycle 1999–2010 and 2015–2018 were included in the cross-sectional analysis to examine the association between HEI-2015, PDI, hPDI, uPDI, and PVD with the grade of inflammation (Fig. 1). A total of 11,939 participants from NHANES 1999–2008 were included in the longitudinal analysis to examine the association between HEI-2015, PDI, hPDI, uPDI, PVD, inflammation with all-cause, CVD, and cancer mortality risks. Participants with missing values for exposure and outcome variables, as well as covariates were excluded from the analysis.

Assessment of dietary intake data in NHANES has been described previously [23]. In brief, a repeated 24-h recall interview was used to collect the data. In cycle 1999–2002, one 24-h dietary recall was conducted in-person in the NHANES Mobile Examination Center. From cycle 2003, a second 24-h recall via a telephone interview was conducted approximately 3–10 days after, in addition to the first in-person recall. This study only used the first-day 24-h dietary recall data for several reasons. First, the methods used to obtain dietary intake data were different between days with face-to-face meeting for the first 24-h dietary recall followed by a telephone interview for the remaining dietary data. Second, to facilitate comparison with previous evidence, using the first-day dietary data, we chose to use the face-to-face dietary data. Finally, there was a significant number of missing observations for the second day dietary data. Macronutrient and micronutrient components of food were determined based on the U.S. Department of Agriculture (USDA) Food and Nutrient Database for Dietary Studies 1999–2008 [25]. Food and beverages were classified into 37 components of USDA Food Patterns by linking participant data to the USDA Food Patterns Equivalents Database [25]. The dietary sampling weights were used to account for the complex study design, missing dietary data, and post-stratification.

The overall diet quality of the population was assessed using the HEI-2015 score which reflects adherence to the 2015–2020 DGA [12, 26]. The HEI-2015 includes 13 food and nutrient groups, encompassing 9 adequacy (total fruit, whole fruit, total vegetables, greens and beans, whole grains, dairy, total protein foods, seafood and plant proteins, and fatty acids) and 4 moderation components (refined grains, sodium, added sugars, and saturated fats). The HEI-2015 assesses densities (per calorie) of consumed food groups and nutrients rather than absolute amounts and does not account for nutrients from dietary supplements. Scoring of HEI-2015 is provided in Supplementary Table 2. Briefly, a score of 0–5 was given for total fruits, whole fruits, total vegetables, greens and beans, total protein foods, seafood and plant proteins. A score of 0–10 was allocated to whole grains, dairy, fatty acids, and all moderation components. Scores were given based on energy density (excluding fatty acids) per 1000 cal, with a maximum total score of 100. HEI scores > 80 indicate a good quality diet, a score ranging from 51 to 80 indicates a diet that needs improvement, and a score < 51 reflects poor diet quality [27].

We constructed PVD using a method described previously [17]. In brief, it involves consumption of seven plant-sourced food groups (fruit, vegetables, nuts, cereals, legumes, olive oil, and potatoes) and five animal-sourced food groups (added animal fats, eggs, fish, dairy products, meats and meat products) (Supplementary Table 3). Participants consumption of those 12 food groups were divided into deciles and each decile was given a score 1–10. For each participant, positive scores were given to plant-sourced food groups while reverse scores were given to animal-sourced food groups. Scores of an individual from the 12 food groups were summed to obtain the indices. The scores for PVD range from 12 to 120 [17].

PDI, hPDI, and uPDI were constructed following a previously described method [18]. In brief, food items were condensed into 18 food groups based on nutrient and culinary similarities and classified under 3 larger categories: animal-based foods, healthy and less healthy plant-based foods (Supplementary Table 3). Healthy and less healthy plant-based foods were separated based on the existing knowledge on the association between plant food components with intermediate conditions and health outcomes [18]. Margarine was excluded from the indices as the fatty acids content has changed from high levels of trans fat to unsaturated fats [18]. Alcoholic beverages were also excluded from the indices given their unclear association with different health outcomes. However, alcohol was adjusted for in the analysis [18].

Food groups were divided into deciles and each decile was given a score 1–10. For PDI, a score of 10 was given to participants for each plant food group, if they were above the highest decile of consumption, a score of 9 if they were above the second highest decile but below the highest decile, and so on. A score of 1 was allocated for consumption below the lowest decile. For each animal food group, reverse scores were given to the participants. For hPDI, positive scores were allocated to healthy plant food groups while reverse scores were given to less healthy plant and animal food groups. Conversely, positive scores were given to less healthy plant food groups, and reverse scores to healthy plant and animal food groups, for uPDI. Scores for the 18 food groups were summed for each participant to obtain the indices. The possible score range was 18 to 180.

The CRP (high sensitivity) values were examined using an immunoturbidimetric system from the collected blood samples. The detailed method for each cycle has been described previously [28]. For the cross-sectional analysis, the CRP values were categorized into low (< 3 mg/L) and severe (≥ 3 mg/L) [29]. For the longitudinal analysis, the CRP values were categorized into low (< 1 mg/L), moderate (1–3 mg/L), and high (> 3 mg/L) inflammation [9].

Ascertainment of mortality for the included participants were obtained from National Center for Health Statistics Public-Use Linked Mortality Files with a probabilistic matching algorithm to the National Death Index [30]. We considered the mortality data up to 2015 for participants in the NHANES cycle 1999–2008 to allow sufficient lag time to determine mortality rates [31], with a mortality status equal to 0 considered to be alive through to the end of 2015.

We selected covariates based on previous literature and summarized in a directed acyclic diagram (Supplementary Fig. 1). Those covariates include sociodemographic factors (age, sex, race, marital status, education, family poverty to income ratio), behavioral factors (smoking, physical activity, alcohol intake), chronic conditions (CVD, cancer, arthritis, diabetes, hypertension, cholesterol), and body mass index (BMI). Measurement of the covariates has been described in the NHANES survey methods and analytics guidelines [23] and NHANES Laboratory Data [28]. A summary of assessment methods and type of the covariates data is provided in Supplementary Table 4.

A complete case analysis was used. HEI-2015, PDI, hPDI, uPDI, and PVD were categorized into tertiles based on participant intake scores. Categorization was undertaken to facilitate interpretation of association estimates in relevance to clinical settings and public health practice and to allow comparison with other studies.

Descriptive analysis of covariates was performed across tertiles of dietary indices. For continuous and normally distributed variables, mean, median and standard deviation were calculated. For categorical variables, proportions were used. Significant differences across dietary indices tertiles were determined using ANOVA and Chi-square tests. Multivariable logistic regression was used to determine the odds ratio of the association between dietary measures and the grade of inflammation. Cox proportional hazard models were used to examine the association between the dietary indices and all-cause, CVD and cancer mortality.

Four models were developed for the cross-sectional and longitudinal analysis. Model one was adjusted for sociodemographic factors. Model two was additionally adjusted for behavioral factors. Model three was additionally adjusted for chronic conditions. Model four was additionally adjusted for BMI and total energy intake (PBD and PVD models only). The p value for trend was determined using tertiles as a continuous variable.

We developed additional Cox proportional hazard models, with exposure variables that combine dietary measure tertiles and inflammation levels to determine the joint association of HEI-2015, systemic inflammation and mortality. We derived a joint variable that combined the HEI-2015 score and inflammation levels. The participants in the highest tertile of HEI-2015, the low inflammation group, served as the reference.

Subgroup analysis based on obesity status was performed in both cross-sectional analysis on the association between HEI-2015, PDI, hPDI, uPDI, and PVD with systemic inflammation, and the joint association between HEI-2015, systemic inflammation, and all-cause mortality. Obesity status was determined based on BMI following the World Health Organization definition, with BMI ≥ 30 km/m² defined as obesity and BMI < 30 kg/m² defined as non-obesity [32]. Stratification by obesity status was determined based on a theoretical justification.

The statistical analyses were performed with STATA/SE version 17 (Stata, StataCorp LP, College Station, TX, USA). Data visualization was performed using RStudio version 2022.02.3 [33] and R package ggplot2 [24].

Characteristics of participants for the cross-sectional analysis are presented in Table 1 and Supplementary Table 5. Participants in higher tertiles of the HEI-2015, PDI, hPDI, and PVD were likely to be older, married or living with a partner, more physically active, non-/ex-smoker, and had a higher level of education. In contrast, participants in the higher tertiles of uPDI were likely to be younger, had a lower physical activity level, and had a BMI ≥ 30 kg/m². HEI-2015 scores of the participants ranged between 8 and 100 with a mean score of 50.4, suggesting a poor diet quality [27]. The mean follow-up year was 11.2 ± 3.2, 1149 deaths were attributable to all-causes, 222 deaths were caused by CVD, 263 deaths were caused by cancer, and the overall mortality rate was 5.3 per 1000 person-years.

The odds ratio of the association between dietary indices and inflammation in the fully adjusted model is presented in Fig. 2 (details for the other models are provided in Supplementary Table 6–8). HEI-2015 (OR_T3vsT1 = 0.76, 95% CI 0.69–0.84; p trend =  < 0.001), PDI (OR_T3vsT1 = 0.83, 95% CI 0.75–0.91; p trend =  < 0.001), hPDI (OR_T3vsT1 = 0.79, 95% CI 0.71–0.88; p trend =  < 0.001), and PVD (OR_T3vsT1 = 0.85, 95% CI 0.75–0.97; p trend = 0.02) were associated with lower systemic inflammation in the fully adjusted model, in all participants. In contrast, uPDI was associated with higher systemic inflammation (OR_T3vsT1 = 1.18, 95% CI 1.06–1.31; p trend = 0.03). The association between all dietary indices and systemic inflammation remained strong in participants with or without obesity with no marked changes in the effect size, except for uPDI in participants without obesity (OR_T3vsT1 = 1.14, 95% CI 1.00–1.29; p-trend = 0.03).

The association between HEI-2015, CRP, and mortality risks is presented in Table 2. In the fully adjusted model, the highest tertile of CRP was associated with all-cause mortality risk in all participants (OR_T3vsT1 = 1.25, 95% CI 1.03–1.53, p trend = 0.02). The highest tertile of HEI-2015 was also associated with reduced all-cause mortality risk in all participants. Association between severe inflammation and all-cause mortality was stronger in people with obesity compared to those without obesity. HEI-2015 was associated with reduced CVD and cancer mortality. No association was observed between CRP with CVD or cancer-associated mortality risk. We also explored the association between PBD indices and PVD with mortality; however, no association was found in the fully adjusted model for all participants (Supplementary Table 9). The joint associations between HEI-2015 and the grade of inflammation with all-cause, CVD, and cancer mortality are presented in Fig. 3 and Supplementary Table 10. No association was found for the joint associations, except for all-cause mortality risk in participants with severe inflammation in the second tertile of HEI-2015 (HR_T3vsT1 = 1.45, 95% CI 1.02–2.05) of the fully adjusted model. A greater reduction of mortality risk with an increase of HEI-2015 scores (tertile 1 to tertile 2) was observed in individuals with low and moderate inflammation. When comparing people with and without obesity, adherence to higher scores of HEI-2015 was not associated with decreased all-cause mortality in participants with obesity who had severe inflammation.

Our results relating to diet quality and inflammation are consistent with previous cross-sectional studies using HEI-2015. The Mitchelstown cohort of the Cork and Kerry Diabetes and Health Diseases Study in Ireland reported that levels of inflammatory biomarkers (i.e., CRP, IL-6 and TNF-α) were lower in individuals with higher HEI-2015 scores[34]. A study of NHANES 2005–2010 reported the highest quintile of HEI-2015 was associated with a 17% lower level of hs-CRP in adults [29]. In our study using NHANES 1999–2010 and 2015–2018, we found higher scores of HEI-2015 were associated with a 16–24% reduction in hs-CRP. This finding is also in accordance with evidence using the previous HEI versions [35, 36], although some have reported no association in studies only using female participants [37, 38].

Our study is one among only a few that have explored the association between PBD and inflammation. We found that adherence to a healthy PDI was associated with a 15–21% reduction in hs-CRP. This is in accordance with the Nurse’s Health Study (NHS), where healthy PDIs were associated with a reduction in hs-CRP levels in cross-sectional (13.6%) and longitudinal (17.8%) analysis [39]. In addition, a lower hs-CRP level was associated with an increased hPDI score in Saudi Arabian [40] and American females with overweight/obesity [41]. We also showed that unhealthy PDI were associated with increased inflammation, in line with previous observational studies [39, 40, 42]. Furthermore, this study revealed an inverse association between overall PDI and inflammation which diverged from previous evidence, reporting no association between overall PDI and hs-CRP [39, 40, 42]. This could be due to the small sample size included in the previous studies. In addition, a study of the NHS cohort reported that BMI attenuated the association [39]. Consistent with this observation, our fully adjusted model demonstrated a similar effect of BMI, although it did not affect the direction of association.

We also found that the PVD was associated with a 12–15% reduction in hs-CRP, consistent with the results of the overall PDI. PVD has been associated with a lower risk of metabolic syndromes [17, 43] and cardiovascular-related risks [44] but no study has explored PVD with inflammation as an outcome, particularly in the US population. This result is expected given that PVD and overall PDI share similar food groupings and scoring methods [17, 18]. In support to our findings, a recent systematic review showed that studies using indices that emphasize intake of PBD (e.g., DASH, vegan/vegetarian pattern, HEI-2010, Mediterranean and paleolithic diet) were associated with lower levels of oxidative stress and inflammatory biomarkers, including hs-CRP [45].

When comparing between HEI-2015, overall PDI, and PVD in the association with hs-CRP, we found the association was stronger for HEI-2015, followed by the overall PDI and PVD. This suggests that HEI-2015 could be a better predictor of inflammation. HEI-2015 is constructed based on a combination of food groups and nutrient components [12]. This is different from PDI and PVD where both only consider food groups [17, 18]. In accordance to our findings, this may indicate that including nutrient components in dietary indices is necessary to examine the association between diet and inflammation.

Furthermore, subgroup analysis by obesity status revealed a greater reduction of hs-CRP in individuals with obesity compared to ones without obesity who adhere to high diet quality and overall PBD (HEI-2015:19–25% vs. 13–22%, PDI:15–22% vs 14–18%). Altogether, our cross-sectional results suggest that higher diet quality and higher intake of healthy PBD were associated with lower systemic inflammation with a greater effect in participants with obesity.

We showed that HEI-2015 and hs-CRP are independent predictors of mortality. Adherence to HEI-2015 was associated with a reduction in CVD mortality. Likewise, HEI-2015 was associated with all-cause and cancer mortality. These findings are in line with several cohort studies in the general US population [46, 47] and Spain [48].

We found no association between higher intake and the quality of PBDs with all-cause mortality. Although this is surprising, given PBDs have been associated with lower risks of CVD, cancer, and type 2 diabetes [49‐51]. Previous evidence reporting the association between PBDs and mortality is inconsistent [21, 52, 53]. A study of NHANES III (1988–1994) reported no association between overall PDI and uPDI with all-cause and CVD mortality while a non-linear association was observed between hPDI and lower all-cause mortality risk [21]. Conversely, associations between overall PDI and hPDI with lower all-cause mortality risk, and uPDI with higher all-cause mortality risk, were found in a study using NHANES 1999–2014 data [52]. Our results for PVD also differ from previous studies which reported higher PVD scores with lower all-cause mortality risk in the US [20] and older Spanish [17, 54] population. These may be due to differences in methodological approaches (e.g., variation in dietary intake data used, inaccuracies in cause-of-death information) and the possibility that participants changed their dietary habits during the follow-up time. Further examination is warranted to confirm our findings.

The joint association suggests that a greater adherence to the dietary guidelines may reduce all-cause, CVD, and cancer mortality risk in all grades of inflammation. This protective effect is biologically plausible since a healthier diet, or high diet quality, is characterized by intake of vegetables, fruits, nuts, and legumes, which compose the healthy plant-based foods. These food groups contain anti-inflammatory and antioxidants components (e.g., dietary fibers, vitamins, polyphenols, and unsaturated fatty acids) which reduce inflammation [5‐7]. Conversely, poor diet or low diet quality, often characterized by animal-based foods (e.g., red meat, saturated fats) and unhealthy plant-based foods, was associated with increased inflammation, partly due to their pro-inflammatory properties [55, 56]. In obesity, all-cause mortality risk was lower with greater adherence to the dietary guidelines in low and moderate inflammation but not for severe inflammation. This indicates that a diet with anti-inflammatory properties alone might not be effective to reduce mortality risk in the obesity. This is likely due to the tenacious nature of low-grade systemic inflammation in obesity [57].

We also observed a weak or null association in the joint associations. This may be due to the scoring of protein intake in the index. In the last two versions, HEI-2010 [58] and HEI-2015 [47], protein intake is separated into total protein foods and seafood and plant proteins components to indicate consumption preference. However, in the earlier versions, HEI-1995 [59] and HEI-2005 [60], they were allocated to the meat (and beans) component. Thus, this may affect the strength of association. The alternate healthy eating index (AHEI) has been suggested as a better predictor of chronic diseases risk compared to HEI [61, 62]. However, the focus of this study was the association between adherence to the dietary guidelines, inflammation, and mortality. As AHEI may not reflect adherence to the dietary guidelines, the HEI was chosen. In addition, it is also plausible that different CRP cutoff levels used to classify inflammation status could influence the strength of the association. Future studies are warranted to confirm the suitability of HEI and the use of AHEI to assess the joint association between systemic inflammation and mortality.

The study strengths include the large number of participants, representative of the general US adult population, with a mean of 11.2-year follow-up. The study is one among the few that has used the general adult population data to examine the association between diet quality, PBDs, inflammation, and mortality in prospective cohort design [47, 52, 63, 64]. We also only included participants who completed a face-to-face 24-h dietary recall given there was a significant number of missing observations for the second day dietary data; and to facilitate comparison with previous evidence.

There are some limitations which should be considered. Given dietary intake data were only obtained at baseline with repetitive assessment unavailable with the NHANES data, it is possible that participants changed their diets during the follow-up period which may impact on the associations. The percentage of missing data that were excluded in this study was quite high (e.g., > 10% for hs-CRP values). This might impact on the strength of association. The use of 24-h dietary recall may not fully reflect habitual dietary intake of the participants since it is unable to account day-to-day variation. Furthermore, the grade of inflammation was assessed solely based on hs-CRP levels. Nonetheless, CRP is widely used as a clinical biomarker of inflammation [65], is a predictor of CVD risk [9], and is associated with increased risk of all-cause and cancer mortality [2]. We were not able to determine causality given the nature of observational studies, there are potential residual confounders that may affect the association. In addition, caution should be taken in interpreting results with wide confidence intervals, which could be due to a sample size limitation in subgroups.