Simple Nutrient-Based Rules vs. a Nutritionally Rich Plant-Centered Diet in Prediction of Future Coronary Heart Disease and Stroke: Prospective Observational Study in the US

To better understand nutrition paradigm shift from nutrients to foods and dietary patterns, we compared associations of a nutrient-based blood cholesterol-lowering diet vs. a food-based plant-centered diet with risk of coronary heart disease (CHD) and stroke. Participants were 4701 adults aged 18–30 years and free of cardiovascular disease at baseline, followed for clinical events from 1985 and 86 to 2018. A plant-centered diet was represented by higher A Priori Diet Quality Score (APDQS). A blood cholesterol-lowering diet was represented by lower Keys Score. Proportional hazards regression was used to calculate hazard ratios (HR). Higher APDQS showed a nutrient-dense composition that is low in saturated fat but high in fiber, vitamins and minerals. Keys Score and APDQS changes were each inversely associated with concurrent plasma low-density lipoprotein cholesterol (LDL-C) change. Over follow-up, 116 CHD and 80 stroke events occurred. LDL-C predicted CHD, but not stroke. APDQS, but not Keys Score, predicted lower risk of CHD and of stroke. Adjusted HRs (95% CIs) for each 1-SD higher APDQS were 0.73 (0.55–0.96) for CHD and 0.70 (0.50–0.99) for stroke. Neither low dietary fat nor low dietary carbohydrate predicted these events. Our findings support the ongoing shift in diet messages for cardiovascular prevention.


Introduction
Over recent decades, there has been a shift in focus from nutrient-based messages to recommendations about food groups and diet patterns [1]. The low fat recommendation, particularly, has been one of the most influential dietary messages ever delivered [2]. Low fat promulgated in the 1980 Dietary Guidelines for Americans (DGA) was adopted, which was influenced by the diet-heart hypothesis (Keys) [2,3]. In that hypothesis, especially

Study Design and Subjects
The Coronary Artery Risk Development in Young Adults (CARDIA) cohort study was initiated to examine the development and progression of CVD risk [13]. CARDIA enrolled 5115 Black and White adults from four US cities, aged 18-30 years in 1985-1986 (exam year 0 [Y0]), and able to walk on a treadmill when recruited. Participants in each field center included a balanced proportion by age, sex, race, and education. Participants have been contacted biannually for 32 years and had nine clinical examinations. Institutional review boards at all study sites approved the study, and all participants gave informed consent. Exclusion criteria for the current analyses included missing data at Y0 for LDL-C; implausible energy intakes (<800 or >8000 kcal/d for men and <600 or >6000 kcal/d for women); having CVD, diabetes or hypertension or receiving treatment for those conditions at Y0; missing covariates. No one took lipid-lowering medications until Y5 in this sample. After exclusion, the final sample for CVD outcomes analyses included 4701 participants, and the plasma lipids analyses included 3495 for 7-year change, 2360 for 13-year change, and 2824 for 20-year change.

Diet Assessment
Diet data were collected using an interviewer-administered diet history at Y0, Y7, and Y20. The reliability and validity of the questionnaire were established previously [14]. Trained interviewers asked the participants about food consumption over the previous month among 100 food categories and recorded open-ended responses of specific types of foods and bever-ages mentioned, their frequency of consumption, their unit or serving sizes, and preparation methods. Total energy and nutrient intake were estimated based on the Nutrition Data System for Research (NDSR, University of Minnesota, Minneapolis, MN, USA) [15].
The NDSR summarized foods into 166 food groups (the same in each exam), which CARDIA then collapsed into 46 food groups for the purpose of creating APDQS, a hypothesisdriven index. APDQS has been validated with varying degrees of predictive ability for obesity, diabetes, kidney function decline, myocardial infarction (MI), and mortality [10][11][12]16,17]. We calculated the diet quality score of plant-centered diets using APDQS. The 46 food groups were classified into beneficial (n = 20), adverse (n = 13), and neutral (n = 13) groups, according to their presumed influence on CVD. Beneficially rated food groups were ranked into quintiles and given positive values (0 [lowest quintile] to 4 [highest quintile]), whereas adversely rated food groups were ranked into their own quintiles and given reverse scores (4 [lowest quintile] to 0 [highest quintile]). Neutrally rated food groups were assigned scores of zero. The total APDQS range was 0-132. In the ADPQS, alcohol was not initially designed to have a U-shaped effect on CVDs. However, the mean alcohol intake in the CARDIA population was very low, and accordingly the cut-off point for the highest quintile of alcoholic beverage intake used to calculate the APDQS was very low (beer 1.10, wine 0.38, liquor 0.45 drinks/day in the year of examination 0 [Y0]). Therefore, we argued that the APDQS tends to regard light to moderate alcohol consumption as beneficial. The mean (SD) for beer, wine, and liquor in the highest quintile was 2.7 (2.2), 1.0 (1.1), and 1.3 (1.6) drinks/day, respectively.

Ascertainment of CVD Events
The first occurrences of CHD or stroke were identified via annual follow-ups and subsequent medical record reviews. CHD included MI, non-MI acute coronary syndrome, and atherosclerotic CHD. Deaths were identified from biannual contact with family members and linkage to the National Death Index. When appropriate, the death certificate, autopsy report, and hospital records were requested with next-of-kin consent. The underlying nonfatal diagnosis or cause of death was adjudicated by two physicians or by committee consensus.

Fasting Plasma Lipid Measurements
Venous blood was drawn after a 12-h fast and sent to a central laboratory. Total cholesterol and HDL-C concentrations were measured using enzymatic reactions [19]. HDL-C was measured after dextran sulfate-magnesium precipitation of other lipoproteins [20], and LDL-C was calculated using the Friedewald equation [21]. Non-HDL-C was calculated as total cholesterol minus HDL-C.

Assessment of Covariates
All covariates were collected at baseline (Y0) and updated at each of the eight follow-up examinations. Information on age, race, educational level, smoking status, medical history, medication use was collected via a self-reported standardized questionnaire and a review of medication bottles. Physical activity levels were assessed through a validated intervieweradministered physical activity history questionnaire which inquired about the frequency of 13 physical activities, with the intensity-weighted summation over all activities expressed as exercise units [22]. Pack-years of smoking were calculated by multiplying the number of packs of cigarettes smoked per day by the duration of smoking in years. Weights and heights directly measured by trained staff were used to calculate body mass index (BMI; kg/m 2 ).

Statistical Analyses
We used linear regression models to investigate the association between changes in APDQS and Keys Score (per 1 SD increase in each) over 7 (Y7-Y0), 13 (Y20-Y7), and 20 (Y20-Y0) years, and the dependent variables of concurrent changes in LDL-C and non-HDL-C. Y0 was the baseline for the 7-and 20-year changes, and Y7 for the 13-year changes. Models were adjusted for baseline LDL-C or non-HDL-C, baseline age, sex, race (White or Black), total energy intake (baseline and change), maximal educational attainment, parental history of CVD (yes or no), smoking pack-years (baseline and change), physical activity levels (baseline and change), time-varying use of lipid-lowering medications (yes or no), and BMI (baseline and change). Analyses were repeated after excluding individuals who had ever used lipid-lowering medications during each of the change periods.
We used Cox proportional hazards regression to estimate hazard ratios (HRs) and 95% confidence intervals (CIs) for CHD and stroke. These were evaluated separately for time-varying APDQS, Keys Score, LDL-C, and non-HDL-C, coded both as quintiles and as continuous variables. Y0 exposure variables predicted events from Y0 to Y7, the average of Y0 and Y7 exposure variables predicted events from after Y7 to Y20, and the average of Y0, Y7, and Y20 exposure variables predicted events from after Y20 to Y32. Note that if both Y7 and Y20 diet was missing, Y0 data was the predictor over all follow-up. The proportional hazards assumption was tested by including interaction terms between the exposures and the log of follow-up time in each model, and hazards were found to be proportional. Person-years were calculated from the date of baseline examination to the date of first reported CVD outcome of interest, death from any cause, or 31 August 2018−whichever came first. The model was adjusted for baseline age, sex, race, total energy intake, maximal educational attainment, parental history of CVD, smoking pack-years, physical activity level, use of lipid-lowering medications, and BMI; time-varying covariates were education, total energy intake, smoking, physical activity, use of lipid-lowering medications, and BMI. In separate models, the diet scores were adjusted for each other and lipid variables. Parallel analyses were performed using % energy from total fat and carbohydrate intake to represent other popular diet messages. Statistical significance of multiplicative interactions between each diet and lipid-lowering medications, race, sex, and BMI in association with CVD outcomes were assessed using Wald test. Statistical analyses were performed using SAS version 9.4 (SAS Institute Inc., Cary, NC, USA), with two-tailed significance set to 0.05.

Study Population
At Y0, the mean ± SD was 62.7 ± 13.0 for APDQS and 48.0 ± 10.5 for Keys Score. Participants with higher APDQS tended to be older, female, White race, more active, more highly educated, and less likely to smoke than participants with lower APDQS (Table 1). They had lower BMI, lower LDL-C and non-HDL-C, and higher HDL-C compared with those in the lowest quintile. Participants with higher Keys Score had characteristics similar to those with lower APDQS, except for no physical activity difference. While extreme quintiles of APDQS were associated with somewhat lower range of saturated fat than were extreme quintiles of Keys Score, APDQS had the larger range of micronutrients and salting behavior.

Characteristics of Diet Indices
Detailed analysis of servings/day of the 46 food groups illuminates how high and low APDQS and Keys Scores were actuated ( Table 2). With about 45 servings/day eaten, those in the highest vs. lowest APDQS quintiles ate 21 vs. 9 servings/day of beneficially rated foods. In contrast, those in the lowest vs. highest Keys Score quintiles ate 16 vs. 13 servings/day of beneficially rated foods. The pattern for adversely rated foods also favored APDQS (24.92 in the lowest vs. 13.42 in the highest), while it was unfavored for Keys score (14.49 in the lowest vs. 23.19 in the highest). Despite the fact that the Keys score was driven by dietary lipids and cholesterol, the data showed that high Keys score diet was generally characterized by a diet low in nutritionally rich plant foods but high in unhealthy plant foods and red and processed meats. Overall, the data indicate that high APDQS was largely driven by nutritionally rich plant foods, and nutritionally rich plant foods make a large majority of daily servings of those who had a high APDQS.
Each 1 SD (11 points) decrease in Keys Score over 20 years was associated with reduced LDL-C (−0.05 ± 0.01, p < 0.001) and reduced non-HDL-C (−0.06 ± 0.01, p < 0.001), and similarly for 7-and 13-year changes. These associations were slightly attenuated after excluding any lipid-lowering medication users during the period evaluated (data not shown).

Association of Diet Score and Plasma Lipids with CVD Outcomes
During the median 32-year follow-up, 116 incident CHD, and 80 stroke events occurred. For CHD, the HRs (95% CIs) for every 1 SD higher were as follows: 0.73 (0.55-0.96) for APDQS, 0.86 (0.68-1.08) for Keys Score, 1.70 (1.42-2.03) for LDL-C, and 1.74 (1.46-2.08) for non-HDL-C ( Table 4). The association between APDQS and CHD persisted after adjustment for LDL-C, non-HDL-C or Keys Score (Supplementary Materials, Table S1). APDQS was associated with risk of stroke 0.70 (0.50-0.99; per 1 SD), but Keys Score, LDL-C, and non-HDL-C were not. Adjustment for LDL-C, non-HDL-C, or Keys Score did not substantially alter this finding. The associations of APDQS (or Keys Score) with CHD and stroke risk did not vary by lipid-lowering medication use, race, or BMI (P-interaction was >0.05 for each).
In analyses of other simple messages that are currently popular, neither total fat nor total carbohydrate intake were associated with risk of CHD or stroke (Table 5), although total fat was minimally correlated with LDL-C (Supplementary Materials, Table S2).  APDQS, A Priori Diet Quality Score; LDL-C, low-density lipoprotein cholesterol; non-HDL-C, non-high-density lipoprotein cholesterol. a Each row is a separate linear regression. Model was adjusted for baseline LDL-C (or non-HDL-C), baseline age, sex, race (White or Black), total energy intake (baseline and change), maximal educational attainment, parental history of CVD (yes or no), pack-years of smoking (baseline and change), physical activity level (baseline and change), use of lipid-lowering medications (yes or no), and BMI (baseline and change). b 1 SD changes were 13 for APDQS, and −11 for Keys Score.  APDQS, A Priori Diet Quality Score; CHD, coronary heart disease; CI, confidence interval; HR, hazard ratio; LDL-C, low-density lipoprotein cholesterol; non-HDL, non-high-density lipoprotein cholesterol; SD, standard deviation. a Each row is a separate proportional hazards regression. Model was adjusted for Y0 age, sex, race (White or Black), total energy intake (time-varying average), maximal educational attainment, parental history of CVD (yes or no), pack-years of smoking (time-varying), physical activity level (time-varying average), use of lipid-lowering medications (yes or no), and BMI (time-varying average). b Time-varying variables that were cumulatively averaged over follow-up at Y0, Y7, and Y20. Y0 predicted events from Y0 to Y7, average of Y0 and Y7 APDQS (average of Y0, Y2, Y5, and Y7 LDL-C) predicted events from after Y7 to Y20, and average of Y0, Y7, and Y20 (average of Y0, Y2, Y5, Y7, Y10, Y15, and Y20 LDL-C) predicted events from after Y20 to Y32. c 1 SD changes were +0.80 mmol/L for LDL-C, +0.87 mmol/L for Non-HDL-C, +13 for APDQS, and −11 for Keys Score. d Statistical significance was estimated by modeling APDQS as a continuous variable in the model.

Discussion
Expanding on our previous finding that APDQS predicted incident CVD outcomes [10], we conducted a head-to-head comparison of predictiveness for CHD and stroke of APDQS vs. Keys score (strongly based in theory and representing a blood cholesterol-lowering diet), total fat restriction, and total carbohydrate restriction. In so doing, we interpreted observed long-term diet features as indicative of different dietary recommendations. There are three main findings of this 32-year prospective study starting with a younger, generally healthy, community-based sample. First, LDL-C predicted incident CHD, consistent with historical assertions within the diet-heart hypothesis that is reflected in the Keys Score [3]. Second, a decrease in Keys Score, primarily driven by low saturated fat intake, was associated with concurrent reductions in both LDL-C and non-HDL-C, again in line with historical assertions. Similarly, APDQS, discouraging high saturated fat foods to some extent, but more importantly emphasizing nutritionally rich plant foods, was inversely associated with concurrent changes in LDL-C and non-HDL-C. Third, we clarified that the diet based on Keys Score has little association with risk of incident CHD and stroke. In contrast, a nutritionally rich plant-centered diet represented by high APDQS was associated with a 27% and 30% lower risk of incident CHD and stroke, respectively, per 1 SD higher of APDQS. Notably, only APDQS had an inverse association with incident stroke. Although restricting total fat and restricting total carbohydrate are dietary messages that persist in the general population, neither was related to incident CHD or stroke.
Publicly promoted messages have been oversimplified, such as "low in saturated fat", "low in total fat", and "low in carbohydrates". They do not clearly specify whether to eat foods that are low in the nutrient that is the focus of the message but may contain other constituents that should be encouraged or discouraged. Particularly, saturated fat intake has long been considered to be atherogenic, and that its reduction may decrease the risk of CHD [3]. The 2020-2025 DGA recommends that saturated fats should comprise <10% energy, and the AHA/ACC recommends 5-6% [6,23]. Yet, results from observational studies and randomized controlled trials (RCTs) are contradictory and inconclusive regarding such an association between saturated fat intake and the risk of CVD [24,25]. In addition, meta-analysis of prospective studies showed that PUFA or MUFA intake was not associated with risk of CVD, although an inverse association between PUFA and CVD risk was observed in a subgroup analysis of studies followed up for more than 10 years [25]. Results from RCTs and prospective studies generally suggested that replacing saturated fats with PUFAs, MUFAs derived from plant foods (but not from animal sources), or whole grains may reduce the risk of CHD [24,[26][27][28][29]. However, it is in practice challenging to separate specific types of fats from other constituents in food because saturated fat is always part of food.
Our data support the a priori hypothesis that LDL-C is causally related to CHD, but the LDL-C association with stroke is not as strong as with CHD [30,31]. We found that higher APDQS and lower Keys Scores were similarly associated with decreases in LDL-C and non-HDL-C. However, only high APDQS predicted CHD and stroke and this prediction may be partly through LDL-C lowering. As actually observed, diet patterns based on Keys score tended to lack numerous acaloric antioxidant nutrients and bioactive phytochemicals, which could explain its lack of association with CHD or stroke. APDQS codifies many of the guidelines in DGA, with some additional classifications and slightly different principles. In APDQS, no one food choice is very influential, as there are many other alternative and eating opportunities. Our assertion is that plant-centered diet pattern recommendations can achieve diets low in saturated fat (e.g., through recommending plant foods and lean and low-fat animal foods), added sugars, and other components that are in accordance with the current dietary guidelines without explicit mention of the nutrient. Furthermore, APDQS itself takes into account the food substitution effect to some extent. For instance, APDQS emphasizes lean and low-fat animal foods (vs. high-fat animal foods), low fat dairy (vs. whole-fat dairy), and whole grains (vs. highly refined sweetened "foods/beverages). Overall, our findings emphasize the importance of considering more than just a single or a combination of nutrients for the prevention of CVD. This thinking is closely allied with the "3V" rule of Fardet and Rock, emphasizing Végétal (plant), Vrai (real) and Varié (varied, if possible organic, local and seasonal) aspects of food choice [32].
Various mechanisms may be involved in providing the cardiovascular benefits of a plant-centered diet. By eating a variety of plant-based foods, including fruits, vegetables, whole grains, legumes, and nuts, individuals can consume an adequate set of vitamins, minerals, fibers, and phytochemicals, all of which interact with each other to reduce the risk of CVD through a number of their properties-they are anti-oxidative; anti-inflammatory; anti-hypertensive; anti-thrombotic; improving glucose control and cholesterol concentrations; as well as their functional properties of low glycemic load and energy density [33]. In contrast, animal-based foods, especially red meats may be harmful partly due to the following characteristics that have been associated with an increased risk of CVD: increased LDL-C and apolipoprotein B, independently of saturated fat [34]; high in dietary haem iron-which increases oxidative stress level [35]; and is high in dietary precursors of trimethylamine-N-oxide [36].
The strengths of this study include its prospective study design following healthy, young adults over a 32-year period, its high retention rate among survivors, and its detailed and long-term pattern measurements of overall diet quality and covariates. The present study, however, has limitations. Given its observational nature, residual and unmeasured confounding biases are possible, although dynamic changes in important risk factors during the follow-ups were controlled for in the time-varying analyses. The generalizability of findings to other populations may be limited.

Conclusions
Our head-to-head comparison of different diet criteria supports the ongoing shift in diet messaging from nutrients to food-based dietary patterns. It finds that several simple nutrient-based rules for choosing what to eat are not associated with incident CHD, even when the message (i.e., low Keys score) has a strong theoretical basis. Such messages are incomplete in that they do not provide guidance about eating or not eating a wide variety of foods. In this sense, the current public health message to reduce intake of saturated fat to decrease the risk of CVD may not be helpful. Our findings provide formal support for promoting total diet quality, with nutritionally rich plant foods at its center.
Supplementary Materials: The following supporting information can be downloaded at: https: //www.mdpi.com/article/10.3390/nu14030469/s1. Supplementary Table S1: Multivariable-adjusted HRs (95% CIs) of incident CHD and stroke according to quintiles of the time-varying average APDQS, Keys Score, LDL-C, and non-HDL-C in mutually adjusted models; Supplementary Table S2: Association between change in % energy from total fat and carbohydrate and concurrent changes in LDL-C and Non-HDL-C.  . This manuscript has been reviewed by CARDIA for scientific content. The sponsor, NHLBI has a representative on the Steering Committee of CARDIA and participated in study design, data collection, and scientific review of this paper. The sponsor had no role in data analysis, data interpretation, or writing of this report.

Institutional Review Board Statement:
Institutional review boards at all CARDIA study sites approved the study, and all participants gave informed consent.
Informed Consent Statement: Written informed consent was obtained from all subjects involved in the study.

Data Availability Statement:
The data underlying this article are available in the CARDIA webpage, at https://www.cardia.dopm.uab.edu (accessed on 20 January 2022).