Effect of Low-Carbohydrate and Low-Fat Diets on Metabolomic Indices and Coronary Heart Disease in US Individuals

Abstract

Background

Low-carbohydrate diet (LCD) and low-fat diet (LFD) patterns are practiced by many in the U.S., although their health effects, as well as the role of diet quality in the effects, are not fully understood.

Objectives

This study aimed to prospectively examine the associations of these diets, which emphasize different quantities and qualities of macronutrients, as well as their objective metabolomic indices, with coronary heart disease (CHD) risk in U.S. individuals.

Methods

We followed 42,720 men in the Health Professionals Follow-Up Study (HPFS, 1986–2016), 64,164 women in the Nurses’ Health Study (NHS, 1986–2018), and 91,589 women in the NHSII (1991–2019) for CHD incidence. Five LCD and five LFD indices were derived based on food frequency questionnaire (FFQ) assessments, each emphasizing different sources and qualities of macronutrients (animal products vs. plant-based foods, whole grains vs. refined carbohydrates, etc.). Multi-metabolite scores of LCD and LFD indices assessed using FFQ assessments were developed through elastic net regressions among 1,146 healthy participants in the Lifestyle Validation Studies (LVS), sub-studies embedded in the NHS/NHSII/HPFS.

Results

During 5,248,916 person-years of follow-up, we documented 20,033 CHD cases. When comparing individuals with the highest LCD scores (emphasizing lower carbohydrate contents) to those with the lowest, the pooled multivariable-adjusted hazard ratios (HRs, 95% CIs) for CHD were 1.05 (1.01, 1.10) for overall LCD, 1.07 (1.02, 1.12) for animal LCD, 0.94 (0.90, 0.99) for vegetable LCD, 1.14 (1.09, 1.20) for unhealthy LCD, and 0.85 (0.82, 0.89) for healthy LCD. These estimates were 0.93 (0.89, 0.98) for overall LFD, 0.94 (0.90, 0.98) for animal LFD, 0.87 (0.83, 0.91) for vegetable LFD, 1.12 (1.07, 1.17) for unhealthy LFD, and 0.87 (0.83, 0.91) for healthy LFD. The healthy versions of the LCD and LFD patterns were also linked to lower triglycerides, higher HDL-C, and lower hs-CRP levels, as well as favorable metabolomic profiles, including increased 3-indolepropionic acid and decreased valine. Unhealthy patterns showed opposite associations. Multi-metabolite scores of LCD and LFD indices were developed in the LVS (Spearman r = 0.57–0.68), and replicated in the NHS, NHSII, and HPFS (r = 0.21–0.38). They showed associations with CHD risk highly consistent with those based on FFQ assessments.

Conclusions

These findings highlight the critical role of diet quality in determining health effects of low-carbohydrate and low-fat diets on CHD risk. The healthy versions of these diets may exert their health benefits through some common pathways that together entail favorable cardiovascular risk profile and lower CHD risk.

Condensed Abstract:

We investigated associations of low-carbohydrate (LCD) and low-fat diet (LFD) patterns, along with metabolomic indices, with coronary heart disease (CHD) in 198,473 U.S. adults. Over 5.2 million person-years, 20,033 CHD cases occurred. Healthy LCD and LFD patterns emphasizing plant-based macronutrients were consistently linked to lower CHD risk, improved lipid and inflammatory biomarkers, and favorable metabolomic profiles, whereas unhealthy versions, high in refined carbohydrates and animal products, were associated with higher risk and adverse profiles. These findings highlight that macronutrient quality, rather than carbohydrate or fat quantity alone, is crucial for CHD prevention.

Graphical Abstract

Central Illustration. Effects of Low-Carbohydrate Diet and Low-Fat Diet on Host Metabolome, Cardiovascular Risk Markers, and Coronary Heart Disease.

Abbreviations: ALCD, animal low-carbohydrate diet; ALFD, animal low-fat diet; HLCD, healthy low-carbohydrate diet; HLFD, healthy low-fat diet; TLCD, total low-carbohydrate diet; TLFD, total low-fat diet; ULCD, unhealthy low-carbohydrate diet; ULFD, unhealthy low-fat diet; VLCD, vegetable low-carbohydrate diet; VLFD, vegetable low-fat diet; LPC, lysophosphatidylcholine; PC, phosphatidylcholine; PE, phosphatidylethanolamine; TG, triacylglycerol.

Analyses were adjusted for age (continuous), white ancestry (yes or no), physical activity (<3.0, 3.0–8.9, 9.0–17.9, 18.0–26.9, ≥27.0 metabolic equivalents of task-hours/week), smoking status (never, former, current), alcohol consumption (0, 0.1–4.9, 5.0–9.9, 10.0–14.9, 15.0–29.9, ≥30.0 g/day), multivitamin use (yes, no), aspirin use (yes, no), post-menopausal hormone use (never, former or current hormone use; Nurses’ Health Study only), oral contraceptive use (never, former or current hormone use; Nurses’ Health Study only), family history of myocardial infarction (yes or no), intake of total energy (quintile), BMI status at baseline (<24.9, 25.0–29.9, ≥30.0 kg/m²), baseline hypertension (yes or no), and baseline hypercholesterolemia (yes or no).

(A) Regression coefficients for top metabolites associated with low-carbohydrate diet and low-fat diet. (B) Top metabolites responsive to low-carbohydrate diet and low-fat diet indices. (C) Effect of low-carbohydrate and low-fat diet associated metabolites on biomarkers. (D) Hazard ratios of CHD for low-carbohydrate and low-fat diet associated metabolites.

Diet quality is essential in the prevention of coronary heart disease (CHD), as demonstrated by landmark interventional and observational studies.^1–4 However, whether diet quality outweighs macronutrient composition in relation to CHD risk is understudied. In the past two decades, diets high in fat and low in carbohydrates and the vice versa have been promoted for their potential benefits, including weight management, glycemic control, and improved lipid profiles.^5–10 However, the efficacy of low-carbohydrate diets (LCD) and low-fat diets (LFD) in reducing CHD risk remains a topic of ongoing debate, with studies reporting mixed findings.^{1, 3, 10–17} These inconsistencies might stem from the fact that foods with contrasting quality and health effects can similarly be used to construct the same dietary pattern that solely emphasizes macronutrient composition.^{18, 19} For example, an LFD emphasizing high carbohydrate intake can consist primarily of refined carbohydrates or whole grain foods, fruits, and vegetables, which may lead to distinct cardiovascular outcomes.²⁰ Similarly, replacing foods or cooking fats high in saturated or trans fats with those rich in cis unsaturated fats has demonstrated protective effects against cardiovascular disease.⁴ Previous studies have consistently shown that LCD and LFD patterns emphasizing less healthy sources were associated with higher mortality, while healthy versions of these diets were linked to lower mortality across different populations.^{19, 21–24} Meanwhile, the long-term effects of LCD and LFD patterns that emphasize various food sources and macronutrient quality on CHD risk are largely underexplored. Additionally, the biological mechanisms underlying these dietary patterns, particularly their impact on the plasma metabolome, are not well understood. Lastly, the lack of objective assessments of these dietary patterns may further hinder causal inference regarding the health effects before clinical trials can be conducted to establish their effects on CHD incidence.

Due to the large public interest in these diets in the U.S.,^{5, 17} and to address the multiple knowledge gaps, we investigated the association between LCD and LFD patterns, with a specific focus on the macronutrients sources and quality, and the risk of CHD in three U.S. cohort studies with more than 30 years of follow-up. We further examined the impact of these diets on plasma metabolome and developed objective metabolomic indices responsive to the adherence to the diets. Lastly, we examined associations of the plasma metabolomic indices with CHD risk to corroborate findings based on self-reported diets.

Methods

Study Population

The current analysis was comprised of three inter-connected stages of analyses that involved different study populations (Figure 1).

Figure 1. Schematic of the Three Interconnected Analysis Stages.

Abbreviations: NHS, Nurses’ Health Study; NHSII, Nurses’ Health Study II; HPFS, Health Professionals Follow-up Study; FFQ, food frequency questionnaire.

Stage 1 strategy

For Stage 1 analysis that examined the dietary indices in relation to CHD incidence, we included data from 198,473 participants, contributing 5,248,916 person-years of follow-up, across three large-scale cohort studies: the Nurses’ Health Study (NHS), NHSII, and Health Professionals Follow-up Study (HPFS). The NHS was initiated in 1976 and enrolled 121,700 registered female nurses aged 30–55 years. The NHSII was established in 1989 and recruited 116,429 younger female nurses aged 25–42 years.²⁵ The HPFS was launched in 1986 and enrolled 51,529 male health professionals aged 40–75 years.²⁶ Since then, participants in each cohort were followed every 2–4 years via mailed questionnaires inquiring about diet, lifestyle factors, medication use, and diagnoses of chronic diseases, with follow-up rates exceeding 90% in each 2-year cycle. We collected blood samples from 32,826 participants in the NHS in 1989–1990, 29,611 NHSII in 1996–1999, and 18,225 HPFS participants in 1993–1995.²⁷ Participants arranged to have samples drawn and shipped by overnight courier to the laboratory (95% were received within 24 h after blood draw).²⁸ After sample processing, plasma aliquots were archived in liquid nitrogen freezers (≤−130 °C). The study protocol for each cohort was approved by the institutional review boards of the Brigham and Women’s Hospital and Harvard T.H. Chan School of Public Health.

In the present analysis, we excluded participants who self-reported having a history of diabetes, cardiovascular disease, or cancer at baseline (1986 for NHS, 1991 for NHSII, and 1986 for HPFS). We also excluded participants with missing dietary data at baseline or those who reported unusual total energy intakes (<500 or >3,500 kcal/day for women and <800 or >4,200 kcal/day for men). After these exclusions, 64,164 participants from NHS, 91,589 from NHSII, and 42,720 from HPFS who had valid food frequency questionnaire (FFQ) data remained in our main analyses.

Stage 2 strategy

For Stage 2 analysis that aimed to cross-sectionally examine the dietary indices in relation to cardiovascular risk markers and plasma metabolome and to develop objective metabolomic scores of the indices, we included data from 1,146 participants who had valid metabolomics data in the Lifestyle Validation Studies (LVS). The LVS consists of two sub-studies among men and women, respectively, with the same goal of examining the validity of instruments of self-reports using objective methods. The Men’s LVS was conducted from 2011 to 2013 and consisted of 914 men aged 46–82 years participating in the HPFS and the Harvard Pilgrim Health Care cohort (HPHC).²⁹ The Women’s LVS was conducted in 2010–2012 and included 796 women aged 45–80 years from the NHS and NHSII.³⁰ Women and men who had no history of coronary heart disease, stroke, cancer, or major neurological disease were invited to participate.

Stage 3 strategy

For the Stage 3 analysis that prospectively examined the metabolomics scores in relation to CHD incidence, we included 11,220 participants, contributing 249,694 person-years of follow-up, with existing metabolomics data exclusively generated at the Broad Institute, which are from 14 individual nested case-control or other sub-studies within the three cohorts.³¹ Associations between metabolomics indices and CHD risk were examined in this population (henceforth referred to as the “blood cohort”).

Dietary Assessments and Construction of LCD and LFD Patterns

In three large-scale cohorts (Stage 1), diet was assessed every 4 years using a validated semiquantitative FFQ. In the LVS (Stage 2), two validated FFQs, 12 months apart, were administered to assess habitual diet. Daily nutrient and energy intakes were calculated by multiplying the frequency of intake by the nutrient and energy content in specific food items with pre-specified portion sizes and then summing the products across all food items. The validity and reproducibility of measurements on macronutrient intake levels have been reported in detail previously.^32–34

We derived five LCD indices which emphasize different compositions and quality of macronutrients based on FFQ assessments (Table S1). Based on the percentage contributions to total energy, the intakes of fat and protein were ranked into 11 equal-sized categories in ascending order, while carbohydrate intake level was ranked in the same way but in descending order.^{3, 23, 35} The carbohydrate categories were scored from 10 (lowest intake) to 0 (highest intake), whereas protein and fat categories were scored from 0 (lowest intake) to 10 (highest intake). We then summed the rank from three macronutrients to create the total LCD (TLCD). Animal LCD (ALCD) and vegetable LCD (VLCD) were calculated in the same manner but using animal fat/protein and vegetable fat/protein instead of total fat/protein. Two additional LCD indices were created to further account for the quality of carbohydrates.²³ The unhealthy LCD (ULCD) was derived by summing the rank of animal protein, the rank of animal fat, and the reverse rank of healthy carbohydrates [sum of carbohydrate from non-starchy vegetables (excluding potatoes and French fries), fruits (excluding fruit juice), legumes, and whole grains)]. The healthy LCD (HLCD) was calculated as the sum of rank of vegetable protein, rank of vegetable fat, and reverse rank of low-quality carbohydrates (sum of carbohydrates from potato, added sugar, and refined grain). All LCD indices had a range of 0 (highest carbohydrates intake) to 30 (lowest carbohydrates intake).

Similarly, we derived five LFD indices. The fat categories were scored from 10 (lowest intake) to 0 (highest intake), whereas protein and carbohydrate categories were scored from 0 (lowest intake) to 10 (highest intake). We then summed the rank from three macronutrients to create the total LFD (TLFD). Animal LFD (ALFD) and vegetable LFD (VLFD) were calculated in the same manner but using animal protein and vegetable protein instead of total protein, respectively, and total carbohydrate. The unhealthy LFD (ULFD) was derived by summing the rank of animal protein, the rank of low-quality carbohydrates, and the reverse rank of vegetable fat. The healthy LFD (HLFD) was calculated as the sum of rank of vegetable protein, the rank of healthy carbohydrates, and reverse rank of animal fat. All LFD indices had a range of 0 (highest fat intake) to 30 (lowest fat intake).

Ascertainment of CHD

The primary outcome was incident CHD, including self-reported cases of coronary revascularization, in addition to nonfatal myocardial infarction (MI) and fatal CHD death for the current analysis.³⁶ In all three cohorts, permission was sought to access medical records of participant who reported having a nonfatal MI on a follow-up questionnaire. Study physicians who were blinded to exposure status reviewed the medical records and confirmed/refuted reported MI cases according to the WHO criteria, which require the presence of symptoms, and either typical electrocardiographic changes or elevated cardiac enzyme levels.^{37, 38} Deaths were identified through reports from the next of kin, the postal authorities, or by searching the National Death Index (NDI).³⁹ Fatal CHD was confirmed by a review of hospital records or autopsy reports if CHD was listed as the underlying cause of death and if evidence of previous CHD was available from medical records.

Assessments of Blood Metabolomics and Other Biomarkers

In the MLVS and WLVS, two blood samples were collected about 6 months apart. Blood lipids [total cholesterol, high-density lipoprotein (HDL) cholesterol, and triglycerides] and high-sensitivity C-reactive protein (hs-CRP) were measured in the MLVS (680 and 662 in bloods 1 and 2, respectively) and WLVS (769 and 759 in bloods 1 and 2, respectively). Non-HDL cholesterol was calculated by subtracting HDL cholesterol from total cholesterol. Low-density lipoprotein (LDL) cholesterol was calculated using the Martin–Hopkins equation.⁴⁰ We averaged the available measurements to maximize the sample size and obtain a more stable estimate that better reflects long-term biomarker levels.

Metabolomics profiles were measured among participants in the LVS using plasma samples from the second blood collection (Stage 2), as well as the 11,220 participants in the blood cohort (Stage 3). Liquid chromatography-tandem mass spectrometry (LC-MS) was used to measure endogenous, polar metabolites and lipids [Broad Institute of MIT and Harvard University (Cambridge, MA)]. The reproducibility and stability of metabolomic profiles have been validated as described previously.^{41, 42} The detailed procedures for metabolite profiling and data processing are provided in the Supplementary Methods. After excluding the metabolites that were unstable to our blood collection conditions,^{41, 43} failing quality control (CV>0.3), or missing rate >25%,⁴¹ there were 299 unique named metabolites included for analyses in the LVS and 261 metabolites in the blood cohort.

Assessment of Covariates

The information on demographic, lifestyle factors, disease status and medication use, including age, ancestry, smoking status, physical activity, vitamin supplement use, aspirin use, family history of MI, alcohol consumption, body weight, oral contraceptives and postmenopausal hormone use (NHS and NHSII only), physician diagnosed hypertension and hypercholesterolemia, and other variables was updated through a biennial follow-up questionnaire. Physical activities were repeatedly assessed, and the metabolic equivalent tasks in hours per week (METs-h/wk) were calculated to represent the total physical activity level.⁴⁴ Fasting status was determined based on self-reported time since the last meal through a questionnaire administered before the blood sample was taken, with fasting defined as ≥8 hours before blood draw. For NHS and NHSII, we further calculated z scores of standardized neighborhood socioeconomic status using census tract variables from the Neighborhood Change Database.⁴⁵

Statistical analysis

The LCD and LFD indices and total energy intake were cumulatively averaged to reflect long-term usual intake. Person-years of follow-up for each participant were calculated from the return of the baseline questionnaires to the CHD diagnosis, deaths, or the end of follow-up (30 June 2018 in NHS, 30 June 2019 in NHSII, and 31 January 2016 in HPFS), whichever occurred first. To minimize potential reverse causality if participants with existing diseases change their usual diet intake, we stopped updating diet once participants developed diabetes, stroke, or cancer during follow-up in the main analysis.⁴⁶ We replaced missing values with valid values in the immediately-preceding questionnaire cycle.

For the Stage 1 analysis, data from three large cohorts were combined and analyzed to maximize the statistical power. Age-standardized characteristics at baseline were presented according to NHS, NHSII, and HPFS cohorts, as well as quintiles of LCD and LFD indices. Characteristics were expressed as mean (standard deviation) for continuous variables, and percentages for categorical variables. We estimated hazard ratios (HRs) and 95% CIs using age (years), calendar year, and cohort–stratified Cox proportional hazards models for the associations between diet indices and CHD risk. Our models were further adjusted for White ancestry, physical activity (<3.0, 3.0–8.9, 9.0–17.9, 18.0–26.9, ≥27.0 METs-h/wk), smoking status (never, former, current), alcohol consumption (0, 0.1–4.9, 5.0–9.9, 10.0–14.9, 15.0–29.9, ≥30.0 g/day), multivitamin use (yes, no), aspirin use (yes, no), post-menopausal hormone use (never, former or current hormone use; women only), oral contraceptive use (never, former or current use; women only), family history of MI (yes or no), intake of total energy (quintile), baseline BMI status (<24.9, 25.0–29.9, ≥30.0 kg/m²), baseline hypertension (yes, no), and baseline hypercholesterolemia (yes, no). We also conducted cubic spline regressions with the same covariates adjusted in the primary analysis for each LCD and LFD indices to characterize potential nonlinear associations with CHD incidence. The proportional hazards assumption was evaluated by including an interaction term between each categorical LCD and LFD indices and the duration of follow-up, and we did not detect violations in the main analyses.

We conducted a stratified analysis by age, sex, obesity, hypertension, and hypercholesterolemia. P values for interaction were calculated using the likelihood ratio tests. We conducted several sensitivity analyses to assess the robustness of the primary findings. First, we repeated the analyses using simply-updated LCD and LFD indices. Second, CHD incidence within one year of baseline was excluded to account for potential reverse causality. Third, we further adjusting for neighborhood socioeconomic status in two female cohorts. Fourth, we conducted a competing risk survival analysis to account for non-CHD death.⁴⁷ Lastly, to examine whether the effects of LCD patterns on CHD risk were independent of LFD and vice versa, we mutually adjusted for LCD indices and the corresponding LFD indices in the same model, e.g., mutually adjusting for HLCD and HLFD indices.

At Stage 2 analysis, to examine the impact of the dietary scores on metabolome, we first used multivariate linear regression models to analyze the associations of LCD and LFD patterns with individual metabolites. Metabolite data were log-transformed and converted to z-scores. We used Bonferroni correction to account for multiple comparisons. To summarize the global metabolomic signatures of LCD and LFD patterns, we used linear regression models with elastic net penalty to derive multi-metabolite scores for LCD and LFD indices (Supplement Methods).

For the Stage 3 analysis, to apply the multi-metabolite scores to participants in NHS, NHSII, and HPFS, we first identified the subset of metabolites measured in both LVS and the blood cohort. Then, a ‘reduced’ score in LVS was built by fitting the multi-metabolite score as a function of overlapping metabolites using linear regression; the coefficients from this ‘reduced’ score were applied to the metabolites to allow the examination of comparable metabolomic indices in the blood cohort. The correlation between LCD and LFD indices and their multi-metabolite scores were assessed using Pearson correlation coefficients. We further evaluated the prospective associations between the multi-metabolite scores of LCD and LFD indices with CHD risk in the blood cohort using Cox proportional hazards models, adjusting for the same covariates described above.

Data were analyzed using SAS version 9.4 (SAS Institute) and R software (version 4.2.0), at a two-tailed α level of 0.05.

Results

Stage 1 Analyses

Characteristics of the NHS, NHSII, HPFS Participants

The study population consisted of 198,473 individuals, giving rise to 20,033 confirmed incident CHD cases. Table 1 shows the age-standardized characteristics of study participants at baseline by cohorts. The distribution of the LCD and LFD indices was similar across NHS, NHSII and HPFS. Table S2 compares the age-standardized characteristics of study participants between the lowest and the highest quintiles of LCD and LFD indices. On average, participants in the highest level of TLCD tended to have lower AHEI-2010 compared with those in the lowest (44 vs. 46), while participants with highest level of TLFD tended to have higher AHEI-2010 (52 vs. 41).

Table 1.

Age-Standardized Characteristics of Participants at Baseline.

	Cohort
Characteristic	NHS (n=64 164)	NHSII (n=91 589)	HPFS (n=42 720)
Age, y	52 (7)	36 (5)	53 (10)
White, %	98.1	96.5	91.3
Smoking status, %
never	46.1	66.5	45.4
past	32.9	21.3	46.0
current	21.0	12.1	8.6
BMI, kg/m2	24.9 (4.7)	24.5 (5.1)	25.4 (3)
BMI category, %
<24.9	62.5	67.4	45.8
25.0–29.9	25.0	19.7	47.5
≥30.0	12.5	12.9	6.6
Physical activity, METs-h/wk	14.3 (20.4)	21.3 (27.1)	19.8 (23)
Alcohol intake, g/d	5.8 (10.2)	3.1 (5.7)	11.9 (15.6)
Family history of CHD, %	25.2	43.9	31.5
Hypertension, %	17.1	5.8	27.5
Hypercholesterolemia, %	8.3	14.2	11.8
Use of multivitamins, %	42.1	44.6	44.6
Use of aspirin, %	68.3	11.0	31.4
Total energy intake, kcal/d	1784 (531)	1794 (531)	1974 (601)
Alternative Healthy Eating Index–2010	46 (11)	43 (10)	48 (10)

Abbreviations: BMI, body mass index (calculated as weight in kilograms divided by height in meters squared); HPFS, Health Professionals Follow-up Study; NHS, Nurses’ Health Study; NHSII, Nurses’ Health Study II.

Correlations within and between LCDs and LFDs and with Nutrient Profiles

Figure S1 presents the inter-correlation of diet patterns. The Spearman correlation coefficients of LCD and LFD indices with AHEI-2010 ranged from −0.45 to 0.64. Intercorrelations among LCD patterns ranged from −0.31 to 0.92, and among LFD patterns from −0.29 to 0.99. Notably, ALCD and TLCD were highly correlated (Spearman coefficient: 0.92), and ALFD closely resembled TLFD (Spearman coefficient: 0.99), due to the fact that these indices emphasize the same component, i.e., total carbohydrates, without differentiating on sources. Figure S2 illustrates correlations of LCD and LFD indices with major food groups and dietary fats among LVS participants. TLCD, ALCD, and ULCD were all positively correlated with saturated fat (Spearman coefficient: 0.42–0.43), trans fat (0.23–0.38), red meat (0.34–0.42), and processed meat (0.22–0.30), whereas VLCD and HLCD showed inverse correlations with these components. In contrast, all LFDs, except ULFD, were inversely correlated with dietary fats and meats (−0.42 to −0.18) and positively correlated with fruits, vegetables, and whole grains (0.09–0.41).

Dietary Patterns and CHD Risk

Pooled multivariable-adjusted HRs (95% CIs) of CHD for LCD and LFD indices were shown in Table 2. Comparing the highest and lowest quintiles, the HRs (95% CIs) were 1.05 (1.01, 1.10) for TLCD (P trend = 0.01), 1.07 (1.02, 1.12) for ALCD (P trend = 0.001), 0.94 (0.90, 0.99) for VLCD (P trend = 0.01), 1.14 (1.09, 1.20) for ULCD (P trend < 0.001), and 0.85 (0.82, 0.89) for HLCD (P trend < 0.001). Spline regression revealed linear dose-response associations between the diet scores and CHD, without evidence of a plateauing effect, except for the VLCD (P = 0.01 for nonlinearity) (Figure 2).

Table 2.

Pooled HRs (95% CIs) of CHD for Low-Carbohydrate and Low-Fat Diets Among 198,473 Participants in NHS, NHSII, and HPFS.

	Quintiles of Low-Carbohydrate/Low-Fat Diets Score					Per 1-SD Increase	P for Trend b
	Q1	Q2	Q3	Q4	Q5
Low-Carbohydrate Diet
TLCD score
Median	7	12	15	19	23
Cases/person-years	4,002/1,042,972	3,900/1,037,726	3,955/1,069,865	4,048/1,056,759	4,128/1,041,594
Age-adjusted model	1	1.03 (0.99, 1.08)	1.04 (0.99, 1.09)	1.11 (1.06, 1.16)	1.19 (1.14, 1.24)	1.06 (1.05, 1.08)	<.001
Multivariable-adjusted model a	1	1.01 (0.97, 1.06)	1.00 (0.96, 1.05)	1.05 (1.00, 1.10)	1.05 (1.01, 1.10)	1.02 (1.00, 1.03)	.01
ALCD score
Median	6	11	15	19	24
Cases/person-years	3,946/1,048,319	3,940/1,052,196	4,000/1,052,680	3,974/1,045,367	4,173/1,050,353
Age-adjusted model	1	1.04 (0.99, 1.08)	1.06 (1.02, 1.11)	1.12 (1.07, 1.17)	1.21 (1.16, 1.26)	1.08 (1.06, 1.09)	<.001
Multivariable-adjusted model a	1	1.01 (0.97, 1.06)	1.02 (0.97, 1.06)	1.05 (1.00, 1.09)	1.07 (1.02, 1.12)	1.03 (1.01, 1.04)	.001
VLCD score
Median	9	13	15	18	21
Cases/person-years	4,277/1,039,464	3,999/1,032,628	4,092/1,073,002	3,774/1,064,407	3,891/1,039,417
Age-adjusted model	1	0.93 (0.89, 0.97)	0.95 (0.91, 1.00)	0.92 (0.88, 0.96)	0.93 (0.89, 0.97)	0.97 (0.96, 0.99)	<.001
Multivariable-adjusted model a	1	0.95 (0.91, 0.99)	0.97 (0.93, 1.01)	0.93 (0.89, 0.98)	0.94 (0.90, 0.99)	0.98 (0.97, 0.99)	.01
ULCD score
Median	7	12	15	19	23
Cases/person-years	3,837/1,037,454	4,096/1,053,698	4,010/1,075,220	3,908/1,044,353	4,182/1,038,194
Age-adjusted model	1	1.12 (1.07, 1.17)	1.13 (1.08, 1.18)	1.23 (1.18, 1.29)	1.37 (1.31, 1.43)	1.12 (1.10, 1.13)	<.001
Multivariable-adjusted model a	1	1.07 (1.03, 1.12)	1.04 (1.00, 1.09)	1.10 (1.05, 1.15)	1.14 (1.09, 1.20)	1.05 (1.03, 1.06)	<.001
HLCD score
Median	9	13	15	18	21
Cases/person-years	4,272/1,044,259	3,897/1,041,552	4,139/1,055,607	3,805/1,054,213	3,920/1,053,286
Age-adjusted model	1	0.89 (0.85, 0.92)	0.87 (0.84, 0.91)	0.82 (0.78, 0.85)	0.77 (0.74, 0.81)	0.91 (0.90, 0.92)	<.001
Multivariable-adjusted model a	1	0.92 (0.88, 0.96)	0.92 (0.88, 0.96)	0.87 (0.84, 0.91)	0.85 (0.82, 0.89)	0.95 (0.93, 0.96)	<.001
Low-Fat Diet
TLFD score
Median	8	12	15	18	22
Cases/person-years	4,163/1,046,713	3,940/1,057,565	4,044/1,057,574	3,767/1,047,301	4,119/1,039,766
Age-adjusted model	1	0.96 (0.92, 1.01)	0.95 (0.91, 1.00)	0.89 (0.85, 0.93)	0.90 (0.86, 0.94)	0.96 (0.95, 0.97)	<.001
Multivariable-adjusted model a	1	0.97 (0.93, 1.02)	0.97 (0.93, 1.02)	0.92 (0.88, 0.96)	0.93 (0.89, 0.98)	0.97 (0.96, 0.99)	<.001
ALFD score
Median	9	12	15	18	21
Cases/person-years	4,126/1,028,663	4,021/1,090,650	3,877/1,011,185	3,835/1,092,918	4,174/1,025,502
Age-adjusted model	1	0.97 (0.92, 1.01)	0.95 (0.91, 0.99)	0.91 (0.87, 0.95)	0.92 (0.88, 0.96)	0.97 (0.96, 0.98)	<.001
Multivariable-adjusted model a	1	0.98 (0.93, 1.02)	0.97 (0.92, 1.01)	0.94 (0.90, 0.98)	0.94 (0.90, 0.98)	0.98 (0.96, 0.99)	.002
VLFD score
Median	6	11	15	19	24
Cases/person-years	4,213/1,032,948	3,974/1,061,572	3,872/1,060,757	4,127/1,049,129	3,847/1,044,511
Age-adjusted model	1	0.90 (0.87, 0.94)	0.87 (0.83, 0.91)	0.90 (0.86, 0.94)	0.79 (0.76, 0.83)	0.93 (0.92, 0.94)	<.001
Multivariable-adjusted model a	1	0.94 (0.90, 0.98)	0.93 (0.89, 0.97)	0.96 (0.92, 1.00)	0.87 (0.83, 0.91)	0.96 (0.95, 0.98)	<.001
ULFD score
Median	10	13	15	17	20
Cases/person-years	3,935/1,039,641	3,902/1,037,733	3,964/1,081,858	3,972/1,048,515	4,260/1,041,170
Age-adjusted model	1	1.08 (1.04, 1.13)	1.11 (1.06, 1.16)	1.16 (1.11, 1.22)	1.28 (1.22, 1.33)	1.09 (1.07, 1.10)	<.001
Multivariable-adjusted model a	1	1.03 (0.99, 1.08)	1.04 (0.99, 1.09)	1.06 (1.02, 1.11)	1.12 (1.07, 1.17)	1.04 (1.02, 1.05)	<.001
HLFD score
Median	6	11	15	19	24
Cases/person-years	4,193/1,042,261	4,026/1,055,319	3,863/1,049,938	4,067/1,054,514	3,884/1,046,885
Age-adjusted model	1	0.92 (0.88, 0.96)	0.84 (0.80, 0.88)	0.82 (0.79, 0.86)	0.75 (0.72, 0.79)	0.90 (0.89, 0.91)	<.001
Multivariable-adjusted model a	1	0.97 (0.93, 1.01)	0.92 (0.88, 0.96)	0.91 (0.88, 0.96)	0.87 (0.83, 0.91)	0.94 (0.93, 0.96)	<.001

Values are hazard ratio (95% confidence interval) unless otherwise indicated.

Abbreviations: ALCD, animal low-carbohydrate diet; ALFD, animal low-fat diet; BMI, body mass index (calculated as weight in kilograms divided by height in meters squared); CHD, coronary heart disease; HLCD, healthy low-carbohydrate diet; HLFD, healthy low-fat diet; HPFS, Health Professionals Follow-up Study; NHS, Nurses’ Health Study; NHSII, Nurses’ Health Study II; TLCD, total low-carbohydrate diet; TLFD, total low-fat diet; ULCD, unhealthy low-carbohydrate diet; ULFD, unhealthy low-fat diet; VLCD, vegetable low-carbohydrate diet; VLFD, vegetable low-fat diet.

^aAdjusted for age (continuous), white ancestry (yes or no), physical activity (<3.0, 3.0–8.9, 9.0–17.9, 18.0–26.9, ≥27.0 METs-h/week), smoking status (never, former, current), alcohol consumption (0, 0.1–4.9, 5.0–9.9, 10.0–14.9, 15.0–29.9, ≥30.0 g/day), multivitamin use (yes, no), aspirin use (yes, no), post-menopausal hormone use (never, former or current hormone use; Nurses’ Health Study only), oral contraceptive use (never, former or current hormone use; Nurses’ Health Study only), family history of myocardial infarction (yes or no), intake of total energy (quintile), baseline BMI status (<24.9, 25.0–29.9, ≥30.0 kg/m²), baseline hypertension (yes or no), and baseline hypercholesterolemia (yes or no).

^bMedian value in each quintile category was used to calculate P value for trend.

Figure 2. Restricted Cubic Spline Analysis of Associations Between Low-Carbohydrate Diet, Low-Fat Diet Indices and CHD Risk Among Participants of NHS, NHSII and HPFS.

Abbreviations: ALCD, animal low-carbohydrate diet; ALFD, animal low-fat diet; HLCD, healthy low-carbohydrate diet; HLFD, healthy low-fat diet; TLCD, total low-carbohydrate diet; TLFD, total low-fat diet; ULCD, unhealthy low-carbohydrate diet; ULFD, unhealthy low-fat diet; VLCD, vegetable low-carbohydrate diet; VLFD, vegetable low-fat diet.

Adjusted for age (continuous), white ancestry (yes or no), physical activity (<3.0, 3.0–8.9, 9.0–17.9, 18.0–26.9, ≥27.0 metabolic equivalents of task-hours/week), smoking status (never, former, current), alcohol consumption (0, 0.1–4.9, 5.0–9.9, 10.0–14.9, 15.0–29.9, ≥30.0 g/day), multivitamin use (yes, no), aspirin use (yes, no), post-menopausal hormone use (never, former or current hormone use; Nurses’ Health Study only), oral contraceptive use (never, former or current hormone use; Nurses’ Health Study only), family history of myocardial infarction (yes or no), intake of total energy (quintile), baseline BMI status (<24.9, 25.0–29.9, ≥30.0 kg/m²), baseline hypertension (yes or no), and baseline hypercholesterolemia (yes or no).

Median of each score was used as the reference value for the restricted cubic spline analysis.

Comparing extreme quintiles (Table 2), the HRs (95% CIs) of incident CHD were 0.93 (0.89, 0.98) for TLFD (P trend < 0.001), 0.94 (0.90, 0.98) for ALFD (P trend = 0.002), and 0.87 (0.83, 0.91) for VLFD (P trend < 0.001), 1.12 (1.07, 1.17) for ULFD (P trend < 0.001), and 0.87 (0.83, 0.91) for HLFD (P trend < 0.001). We did not detect any nonlinear relationships between LFD indices and CHD in the spline regressions (all P values for linearity < 0.001) (Figure 2).

In a secondary analysis assessing potential effect modification by age, sex, and health status, the associations between LCD and LFD indices with CHD appeared to be more pronounced among younger, non-obese, non-hypertensive, and non-hypercholesterolemic participants, although most P values for interactions were not statistically significant (Figure S3, Figure S4). Results remained consistent when performing the analyses within NHS, NHSII and HPFS participants, separately (Table S3, Table S4). In sensitivity analyses using LCD and LFD indices based on each time-varying individual FFQ assessment, excluding participants who reported CHD within one year of baseline, additionally adjusting for socioeconomic status score, or using the competing risk survival model, the results were largely unchanged (Table S5, Table S6, Table S7, Table S8). Mutually adjusting for LCD and LFD indices slightly attenuated the estimates, whereas the results remained statistically significant, except for TLCD which changed to null (P trend = 0.45) (Table S9).

Stage 2 Analyses

Dietary Patterns, Metabolites, and CVD Risk Markers

Table S10 shows the characteristics of 1,146 LVS participants included in the analysis. Figure 3 presents the top metabolites significantly associated with LCD and LFD indices and their effects on biomarkers and risk of CHD (Central Illustration). Metabolites positively associated with unhealthy LCD and LFD patterns and inversely associated with their healthy counterparts were linked to adverse biomarkers and higher CHD risk, including triacylglycerols with fewer double bonds. Figure S5 and Supplement Data 1 show the detailed information for the metabolites associated with LCD and LFD indices. Most metabolites responsive to these diet indices were lipids, such as triacylglycerol (TG), phosphatidylcholine (PC), and phosphatidylethanolamine (PE), amino acids and derivatives. The metabolites linked to healthy and unhealthy dietary patterns were distinct, despite both emphasizing a reduced intake of carbohydrates or fats. For example, gentisic acid, hippuric acid, trigonelline, N-acetylornithine, 3-indolepropionic acid (IPA), and 2-aminoheptanoic acid were positively associated HLCD and HFLD and inversely associated with ULCD and ULFD; C7 carnitine, creatine, and valine were positively associated ULCD and ULFD and inversely associated with HLCD and HLFD.

Figure 3. Individual Metabolites Associated with Low-Carbohydrate Diet and Low-Fat Diet Indices in the LVS participants.

Analyses were adjusted for age (continuous), white ancestry (yes or no), physical activity (<3.0, 3.0–8.9, 9.0–17.9, 18.0–26.9, ≥27.0 metabolic equivalents of task-hours/week), smoking status (never, former, current), alcohol consumption (0, 0.1–4.9, 5.0–9.9, 10.0–14.9, 15.0–29.9, ≥30.0 g/day), multivitamin use (yes, no), aspirin use (yes, no), post-menopausal hormone use (never, former or current hormone use; Nurses’ Health Study only), oral contraceptive use (never, former or current use; Nurses’ Health Study only), family history of myocardial infarction (yes or no), intake of total energy, BMI status at baseline (<24.9, 25.0–29.9, ≥30.0 kg/m2), baseline hypertension (yes or no), and baseline hypercholesterolemia (yes or no).

Multi-marker Index Development

There were 76–140 metabolites selected to construct the multi-metabolite scores for LCD and LFD indices in LVS through the elastic net regression, while 45–88 metabolites were available to calculate the “reduced” multi-metabolite scores in the blood cohort. Spearman coefficients between diet indices and derived multi-metabolite scores ranged from 0.57 to 0.68 in LVS, and from 0.21 to 0.38 in three cohorts, respectively (Table S11). Figure 4 shows the associations of per-SD increase of LCD and LFD indices and their multi-metabolite scores with blood lipids and hs-CRP. In general, HLCD and HLFD were associated with lower triglyceride and hs-CRP levels, along with higher HDL cholesterol levels. Conversely, ULCD and ULFD were linked to higher triglyceride and hs-CRP levels and lower HDL cholesterol levels. However, these dietary patterns were not significantly associated with LDL cholesterol or non-HDL cholesterol levels.

Figure 4. Biomarkers Associated with Low-Carbohydrate Diet and Low-Fat Diet Indices in the LVS participants.

Abbreviations: ALCD, animal low-carbohydrate diet; ALFD, animal low-fat diet; HLCD, healthy low-carbohydrate diet; HLFD, healthy low-fat diet; TLCD, total low-carbohydrate diet; TLFD, total low-fat diet; ULCD, unhealthy low-carbohydrate diet; ULFD, unhealthy low-fat diet; VLCD, vegetable low-carbohydrate diet; VLFD, vegetable low-fat diet; TG, triglyceride; TC, total cholesterol; HDL-C, high-density lipoprotein cholesterol; LDL-C, low-density lipoprotein cholesterol (calculated using Martin-Hopkins equation).

Analyses were adjusted for age (continuous), white ancestry (yes or no), physical activity (<3.0, 3.0–8.9, 9.0–17.9, 18.0–26.9, ≥27.0 metabolic equivalents of task-hours/week), smoking status (never, former, current), alcohol consumption (0, 0.1–4.9, 5.0–9.9, 10.0–14.9, 15.0–29.9, ≥30.0 g/day), multivitamin use (yes, no), aspirin use (yes, no), post-menopausal hormone use (never, former or current hormone use; Nurses’ Health Study only), oral contraceptive use (never, former or current use; Nurses’ Health Study only), family history of myocardial infarction (yes or no), intake of total energy, BMI status at baseline (<24.9, 25.0–29.9, ≥30.0 kg/m²), baseline hypertension (yes or no), and baseline hypercholesterolemia (yes or no).

Stage 3 Analyses

Multi-metabolite Scores of Dietary Patterns and CHD

We included 11,220 participants with available metabolomics data, among whom 1,111 incident CHD cases were confirmed. The multivariable-adjusted hazard ratios (HRs, 95% CIs) of CHD per 1-SD increment of the multi-metabolite scores were 1.11 (1.04–1.18) for TLCD, 1.10 (1.04–1.17) for ALCD, 1.06 (0.99–1.13) for VLCD, 1.12 (1.06–1.20) for ULCD, 0.96 (0.90–1.02) for HLCD, 0.89 (0.84–0.95) for TLFD, 0.89 (0.84–0.95) for ALFD, 0.88 (0.82–0.93) for VLFD, 1.06 (0.99–1.13) for ULFD, and 0.89 (0.83–0.95) for HLFD, respectively (Figure 5, Table S12). The associations of multi-metabolite scores with CHD risk were only slightly attenuated after further adjustment for the corresponding diet indices. Figure S6 illustrates the correlations between individual metabolites that were used to build the multi-metabolite scores for LCD and LFD indices in blood cohort, and the risk of CHD. Most metabolites linked to the dietary patterns showed associations with CHD risk that aligned with the corresponding dietary pattern (Figure S7). For instance, decreased glycine was observed in response to ULCD and ULFD patterns and was significantly associated with a lower risk of CHD. Conversely, decreased valine was linked to HLCD and HLFD patterns and was significantly associated with a higher risk of CHD.

Figure 5. Associations of Metabolomic Scores of Low-Carbohydrate and Low-Fat Diets with CHD Among 11,220 Participants with Metabolomic Data in NHS, NHSII, and HPFS.

Abbreviations: ALCD, animal low-carbohydrate diet; ALFD, animal low-fat diet; HLCD, healthy low-carbohydrate diet; HLFD, healthy low-fat diet; TLCD, total low-carbohydrate diet; TLFD, total low-fat diet; ULCD, unhealthy low-carbohydrate diet; ULFD, unhealthy low-fat diet; VLCD, vegetable low-carbohydrate diet; VLFD, vegetable low-fat diet.

Analyses were adjusted for age (continuous), white ancestry (yes or no), physical activity (<3.0, 3.0–8.9, 9.0–17.9, 18.0–26.9, ≥27.0 metabolic equivalents of task-hours/week), smoking status (never, former, current), alcohol consumption (0, 0.1–4.9, 5.0–9.9, 10.0–14.9, 15.0–29.9, ≥30.0 g/day), multivitamin use (yes, no), aspirin use (yes, no), post-menopausal hormone use (never, former or current hormone use; Nurses’ Health Study only), oral contraceptive use (never, former or current hormone use; Nurses’ Health Study only), family history of myocardial infarction (yes or no), intake of total energy (quintile), BMI status at baseline (<24.9, 25.0–29.9, ≥30.0 kg/m²), baseline hypertension (yes or no), and baseline hypercholesterolemia (yes or no).

We included metabolomic scores plus low-carbohydrate or low-fat diet indices to examine association independence.

Discussions

In the current study among U.S. men and women who were followed for over 30 years, greater adherence to low-carbohydrate or low-fat dietary patterns emphasizing high-quality macronutrients, such as plant-based proteins, healthy fats, and minimally processed carbohydrates, was associated with a significantly lowered risk of CHD. In contrast, such dietary patterns, when enriched with low-quality macronutrients characterized by refined carbohydrate, animal protein, or animal fat, were associated with a higher risk. Healthy LCD and LFD patterns were also associated with improved lipid and inflammatory biomarkers, including lower triglyceride and hs-CRP and higher HDL cholesterol levels. Additionally, plasma metabolites analysis corroborated these findings, showing distinctive metabolomic responses to healthy versus unhealthy dietary patterns. Healthy LCD and LFD patterns were linked to higher levels of trigonelline, hippuric acid, N-acetylornithine, and IPA, along with lower levels of C7 carnitine and valine. In contrast, unhealthy LCD and LFD patterns were primarily associated with these metabolites in the opposite direction. Importantly, the associations of these objective metabolomic indices with CHD risk were highly consistent with those based on FFQ assessments. These findings highlight the importance of macronutrient quality and food sources for maintaining an optimal cardiovascular health.

Macronutrient quality: the key determinant of associations of low-carbohydrate and low-fat diets

Numerous trials and cohort studies have investigated the short-term effects of low-carbohydrate and low-fat diets on cardiovascular risk factors with mixed findings,^{7, 9, 10, 18, 48, 49} but evidence of their long-term effects on cardiovascular outcomes is more scarce and inconsistent, likely due to the fact that these studies were largely focused on the overall quantity of carbohydrates and fats, without differentiating the quality and sources of these macronutrients.^{3, 50, 51} The current analysis is among the first prospective studies to explicitly examine the relationships between LCD and LFD patterns emphasizing macronutrient quality and the risk of CHD. By integrating dietary, metabolomic, clinical risk markers, and disease outcome data, our findings reconcile prior inconsistencies by demonstrating that macronutrient quality, rather than quantity alone, serves as the unifying principle underlying the health effects of low-carbohydrate and low-fat diets. Intriguingly, the protective associations of healthy LCD and LFD patterns on CHD are largely comparable. During ~30 years of follow-up, we observed a nearly 15% lower incidence of CHD when comparing extreme quintiles of healthy LCD or LFD. For the secondary prevention, the CORDIOPREV trial demonstrated that the Mediterranean diet was superior to a low-fat diet in preventing major cardiovascular events, likely because the Mediterranean pattern emphasized high-quality food sources such as olive oil, nuts, and whole grains.⁵² Similarly, the DIETFITS Randomized Clinical Trial concluded that there was no significant difference in weight change between a HLCD vs a HLFD during a 12-month weight loss diet intervention.¹⁸ A prior study also demonstrated that greater adherence to several healthy dietary patterns, including a healthful plant-based diet, was consistently associated with a lower risk of cardiovascular disease.⁵³ Such data imply that, as long as the quality is comparable, macronutrient composition variations do not seem to further predict CHD risk. The sources of macronutrients are indeed critical in determining health outcomes.^54–57 Abundant evidence has suggested that refined carbohydrates, i.e., added sugars, and processed grains, saturated fats from animal products, and trans fat lead to adverse cardiovascular outcomes,^{57, 58} whereas carbohydrates from whole grains and non-starchy vegetables and unsaturated fats from plant-based sources provide protective benefits.^{55, 59} It is notable that in our study TLFD and ALFD were both associated with lower CHD risk. This is probably due to the fact that these indices were associated with lower intakes of saturated fat, trans fat, and red and processed meat and higher consumption of low-fat dairy, whole grains, vegetables, and fruits. Nevertheless, HLFD was more strongly associated with lower CHD risk since this index is further linked to higher intake of healthy sources of all three groups of macronutrients.

Low-carbohydrate and low-fat diets in relation to CHD outcomes across health status subgroups

Interestingly, the associations between LCD and LFD patterns and CHD were stronger in individuals with lower cardiovascular risk as characterized by being free of obesity, hypertension, or hypercholesterolemia. A study using NHANES data observed similar trends when examining the associations of LCD with mortality by showing that participants without obesity had a 48% higher risk of all-cause mortality, while those with obesity had a 19% higher risk when comparing the highest to the lowest quartiles of LCD.⁶⁰ In this context, the Mediterranean diet was reported to reduce the risk of obesity, and overall and cardiovascular mortality, especially among non-obese individuals.^{61, 62} Additionally, data from NHS and HPFS found that the favorable associations between HLCD and mortality among individuals with diabetes were more pronounced among participants with more low-risk lifestyle risk factors, suggesting that people might gain additional health benefits from adopting healthy LCD and LFD patterns by maintaining a healthy metabolic status.²³ Although the exact mechanisms underlying these observations remain unclear, individuals with obesity or other metabolic disorders may already have chronic inflammation, insulin resistance, or endothelial dysfunction, which can diminish the responsiveness to dietary interventions aimed at improving cardiovascular health.^{61, 63} These associations should be investigated in future studies. Nevertheless, individuals with higher cardiovascular risk can still benefit from the healthy LCD and LFD.

Comparison with prior studies on other health outcomes

Our results are consistent with several prior studies, which emphasized that both the quantity and quality of macronutrients significantly influence health outcomes. For example, a prior NHANES study found that both healthy LCD and LFD, rather than overall LCD and LFD, were associated with lower mortality among 37,233 U.S. adults 20 years or older.¹⁹ Consistent results were observed among other populations.^{21–24, 64, 65} Similarly, an HLFD was associated with 16% lower cardiovascular mortality, comparing extreme quintiles, in the National Institutes of Health–American Association of Retired Persons (NIH–AARP) Diet and Health Study.²² All these studies have been advocating a well-balanced moderate lower-carbohydrate diet that emphasizes healthy carbohydrates and plant-based protein and fat.⁶⁶ Findings from a pooled analysis of nine prospective cohort studies indicated a positive association between LCD and cardiovascular disease mortality.⁶⁰ This may partially explain the J-shaped relationship between VLCD and CHD risk as it emphasizes reducing overall carbohydrate intake while high-quality carbohydrates are not further emphasized. Along with the existing evidence, our study highlights the need to differentiate between healthy and unhealthy sources of all macronutrients when recommending low-carbohydrate or low-fat diets.

Low-carbohydrate and low-fat diet on cardiovascular risk markers and metabolomic profiles

An important aspect of our study was corroborating the associations of LCD and LFD patterns using biomarkers. While prior studies have compared the effects of LCD and LFD on lipid and glucose metabolism, these studies often did not consider the quality of carbohydrates and fats.^{8, 11} Our findings showed that healthy LCD and LFD patterns were similarly associated with lower levels of triglycerides, total cholesterol, and hs-CRP, and higher HDL cholesterol. In contrast, unhealthy LCD and LFD patterns exhibited opposing associations. However, these dietary patterns did not significantly affect LDL or non-HDL cholesterol levels, possibly because the LVS cohort consisted primarily of healthy individuals without any major chronic diseases, including cardiovascular disease, resulting in limited variation in LDL cholesterol. Using metabolomic data, we further identified that lipid metabolites, such as triacylglycerols, phosphatidylcholines, and phosphatidylethanolamines, were distinctly responsive to the quality of dietary patterns, despite both emphasizing low carbohydrate or low-fat intake. For example, phosphatidylcholine metabolites (PC (34:4), PC (36:4), and PC (38:4)) and C7 carnitine were positively linked to unhealthy LCD and LFD patterns but showed inverse associations with the healthy patterns. Additionally, non-lipid metabolites such as gentisic acid, hippuric acid, trigonelline, N-acetylornithine, IPA, and 2-aminoheptanoic acid were enriched in healthy LCD and LFD patterns and reduced in unhealthy ones, whereas creatine and valine showed the opposite trend, being enriched in unhealthy patterns and depleted in healthy ones. These metabolites are largely derived from specific food sources – for instance, carnitines from red meat and hippuric acid and IPA from whole-grain carbohydrates and dietary fiber – which have been previously associated with CHD risk.^{49, 67–69} These findings provide valuable insights into the biological mechanisms through which macronutrient quality impacts health risk.³¹ Moreover, the multi-metabolite scores that consisted of these and other metabolites that together reflect the adherence to the dietary patterns showed highly consistent associations with CHD risk. The consistent pattern of associations corroborated the validity of the FFQ assessments of diet, as well as the associations based on such assessments. The performance of our metabolomics indices shall be examined further in other populations before they can be possibly applied to evaluate the adherence to these diets. Nevertheless, future studies should attempt to include a broader spectrum of metabolites to further the understanding of biological underpinning of our findings or include food biomarkers to further improve the assessments of adherence to these diets.

Study limitations

The main strengths of our study include the prospective examination of LCD and LFD patterns with CHD by emphasizing on the macronutrient quality, repeated measurements of diet using validated FFQ, large sample size, and leveraging both cohort and metabolomics data. There are also several limitations to this study. First, although our analyses extensively adjusted for lifestyle factors and health status, we still cannot exclude the role of residual or unmeasured confounding. Meanwhile, given the findings based on our objective metabolomics indices, the observed associations are unlikely to be entirely driven by uncontrolled confounding. Second, evidence from our study regarding the LCD and LFD patterns cannot be directly generalized to diets with much lower carbohydrates or fats intake, such as the ketogenic diet. The interpretation of our findings shall take into account the range of macronutrient intake. Third, the metabolites considered in the current study may still represent a small fraction of all relevant metabolites in human circulation. In addition, we did not include food biomarkers of food origin in this analysis. Fourth, the dietary assessments using the validated FFQ were still inevitably subject to measurement errors. However, since the errors are not likely to be related with outcome ascertainment, they are largely non-differential and may be more likely to bias the true associations toward the null. Finally, our findings may have limited generalizability because the study participants consisted of health professionals with higher health consciousness and better access to health care. However, the biological mechanisms are unlikely to be drastically different across populations.

Conclusions

Unhealthy LCD and LFD patterns emphasizing low-quality protein and fat from animal sources and refined carbohydrates were associated with a higher risk of CHD, whereas healthy LCD and LFD patterns emphasizing high-quality plant-based protein/fat and high-quality carbohydrates, were associated with a lower risk. These findings, based on both FFQ assessments of diet and objective metabolomics indices, suggest that the health impact of low-carbohydrate and low-fat diets may largely depend on the quality and food sources of macronutrients. Emphasizing high-quality macronutrients, such as those from plant-based foods and whole grains, is essential for promoting cardiovascular health, regardless of whether a low-carbohydrate or low-fat diet is followed.

Box 1.

Clinical and Public Health Implications

The current study findings underscore that the health effects of low-carbohydrate and low-fat diets depend primarily on the quality rather than the quantity or compositions of macronutrients. Diets emphasizing plant-based sources of fats and proteins and high-quality carbohydrates (such as whole grains, fruits, and legumes) were associated with lower CHD risk, whereas patterns rich in refined carbohydrates or animal-derived fats and proteins were linked to higher risk. These observations are corroborated using objective metabolomic indices that do not rely on self-reported diet. The findings support dietary guidelines and public health policies that emphasize on food-based, high-quality dietary patterns for maintaining cardiovascular health. In addition, promoting an overall healthy diet, regardless of the carbohydrate or fat contents, shall be among the primary prevention strategies adopted by clinicians, dietitians, and patients.

How This Study Advances Prior Evidence

Previous landmark studies have shown that diets emphasizing healthy sources of fats, proteins, and carbohydrates, particularly those rich in unsaturated fats and high-quality carbohydrates, are associated with lower risks of coronary heart disease and mortality. The current analysis extended the evidence by elucidating the role of diet quality in the settings of low-carbohydrate and low-fat diets. These diets are important, popular dietary patterns practiced by many Americans, although their health effects are subject to much debate. Using traditional nutritional epidemiology and cutting-edge metabolomic indices developed using machine-learning approaches, this work reaffirmed the notion that diet quality outweighs the particular macronutrient compositions for the primary prevention of CHD, and also provided mechanistic insights elucidating potential biological pathways.

Future Directions

In the era of precision nutrition, it is essential to identify factors that modulate the health benefits from consuming healthy diets. In this context, it is of importance to understand how these diets may be differentially associated with heart disease risk among individuals with different genetic predisposition, the gut microbiome profile, etc. The role of diet quality in the setting of these diets in the secondary prevention of cardiovascular disease, as well as the primary/secondary prevention of heart disease-related conditions, such as diabetes, excess weight gain, cognitive decline, and aging, shall be elucidated. Future research integrating multi-omics approach and randomized feeding trials in diverse populations will be critical to further elucidate the mechanistic pathways linking dietary quality to heart health and to guide precision nutrition strategies.

Abbreviations and Acronyms

CHD

coronary heart disease

LCD

low-carbohydrate diets

LFD

low-fat diets

FFQ

food frequency questionnaire

Data Availability.

Owing to participant confidentiality and privacy concerns, data cannot be shared publicly and requests to access NHS/NHSII/HPFS data must be submitted in writing. According to standard controlled access procedures, applications to use NHS/NHSII/HPFS resources will be reviewed by the External Collaborations Committee to verify that the proposed use maintains the protection of the privacy of participants and the confidentiality of the data. Investigators wishing to use NHS/NHSII/HPFS data are asked to submit a brief description of the proposed project (go to https://www.nurseshealthstudy.org/researchers (nhsaccess@channing.harvard.edu) and https://sites.sph.harvard.edu/hpfs/for-collaborators/ for details). Further information including the procedures to obtain and access data from the Nurses’ Health Studies and Health Professionals Follow-up Study is described at https://www.nurseshealthstudy.org/researchers (nhsaccess@channing.harvard.edu) and https://sites.sph.harvard.edu/hpfs/for-collaborators/.