ORIGINAL RESEARCH article

Front. Nutr., 04 November 2026

Sec. Nutrition Methodology

Volume 13 - 2026 | https://doi.org/10.3389/fnut.2026.1924089

Macronutrient composition and incident hypertension during China's nutrition transition: a longitudinal cohort analysis of the China Health and Nutrition Survey

  • 1. Research Center of Machine Learning and Public Health, China Three Gorges University, Yichang, China

  • 2. School of Construction Engineering, Shenzhen Polytechnic University, Shenzhen, China

Abstract

Background:

China's nutrition transition has shifted dietary macronutrient composition while hypertension prevalence has risen. Whether the routinely available macronutrient composition adequately captures the risk of incident hypertension remains unclear.

Methods:

We followed 14,384 baseline normotensive adults in the China Health and Nutrition Survey (baseline 1991–2011; blood pressure follow-up through 2015). Baseline diet was assessed using three consecutive 24-h recalls combined with household food inventories. Primary composition analyses retained records whose calculated protein, fat, and carbohydrate shares summed to 95%–105%. These shares were closed to 100% for substitution models; acceptable macronutrient distribution ranges (AMDR) from the Chinese Dietary Reference Intakes (DRI) were instead applied to the original percentages of total reported energy. Cox models used Efron handling of tied event times and community-clustered robust standard errors. Sensitivities included alternative composition-sum windows, event-time models, categorical and nonlinear exposure forms, cumulative-average repeated diet, measured-blood-pressure-only outcomes, and censoring-weighted 6-year temporal prediction.

Results:

Hypertension was detected in 5,234 adults (36.4%). In primary Model 3 (n = 13, 276; 4,745 events), closed protein replacing carbohydrate [hazard ratio (HR) per five composition points 0.986, 95% confidence interval (CI) 0.920–1.056; p = 0.679], closed fat replacing carbohydrate (HR 0.986, 95% CI 0.968–1.004; p = 0.120), and AMDR count (HR per component 1.007, 95% CI 0.974–1.041; p = 0.679) crossed the null. Categorical AMDR and alternative event-time models were concordant. From 1991 to 2011, the standardized AMDR count increased from 1.37 to 1.53 and standardized hypertension prevalence from 17.4% to 22.6%. Among 9,284 verified 6-year endpoints, dietary features did not materially improve temporal prediction.

Conclusion:

A single baseline measure of gross macronutrient composition had limited association and predictive value for incident hypertension. Hypertension-related surveillance still requires sodium, potassium, dietary fiber, fat quality, and food-group information. Repeated-diet findings were exploratory and require dedicated longitudinal confirmation.

1 Introduction

Hypertension remains a leading modifiable risk factor for cardiovascular disease and premature mortality worldwide, and its burden has grown rapidly in China (1–4). In six national surveys of Chinese adults aged 18–69 years, crude hypertension prevalence was 24.9% in 2004 and 38.1% in 2018, with persistently low rates of awareness and control (5); other large surveys likewise document a substantial burden (6–8). Over the same broad period, China has undergone one of the most rapid nutrition transitions ever documented: a shift from traditional cereal-based diets toward higher dietary fat and more animal-source foods, accompanied by urbanization and broad socioeconomic change (9–11). These parallel trends pose a measurement question: does the dietary information routinely captured in large cohorts and surveillance systems help explain who develops hypertension?

Blood-pressure-relevant dietary factors are well-established. Controlled-feeding trials show that blood pressure responds to dietary patterns rich in fruits, vegetables, legumes, low-fat dairy, and whole grains, and to sodium reduction (12–14). High sodium and low potassium intakes are well-documented in Chinese samples (10, 15), and a large pragmatic trial showed that potassium-enriched salt substitution reduced the risk of stroke, major cardiovascular events, and death (16). Diet-quality indices that capture pattern- and food-quality features are associated with cardiovascular outcomes (17), and food-group-level dietary patterns in the China Health and Nutrition Survey (CHNS) have been linked to hypertension risk (18). A blood pressure signal from diet is therefore more likely to emerge when measurement instruments capture these dimensions.

In practice, a dietary layer routinely available in many longitudinal datasets is macronutrient composition: energy and the energy shares of protein, fat, and carbohydrate. These variables are computed centrally, are comparable across survey waves, and depend far less on complete food composition mapping than micronutrient or food group estimates. Whether this macronutrient layer is sufficient to capture incident hypertension risk during a nutrition transition remains unclear. This analysis does not test whether diet affects blood pressure (an exposure space that omits sodium, potassium, and food groups cannot address that question) but whether this accessible dietary information is fit for this particular purpose.

We addressed this question in the CHNS, a longitudinal cohort spanning more than two decades of China's nutrition transition (19, 20). We scoped the dietary exposure to the macronutrient layer that can be reliably reconstructed from the official CHNS nutrition files. We (i) estimated the association of macronutrient energy percentages and an acceptable macronutrient distribution range (AMDR) concordance count with incident hypertension using sequentially adjusted survival models; (ii) examined whether a flexible machine learning approach concurred with the survival model findings; and (iii) compared standardized secular trends in macronutrient composition and hypertension prevalence.

2 Materials and methods

We report the cohort component according to the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement and its nutritional epidemiology extension (STROBE-nut) (21, 22). The completed STROBE and STROBE-nut reporting checklist is included in the Supplementary material. The machine learning component is explicitly exploratory rather than a deployable clinical prediction model. Supplementary Code 1 supplies the complete R4 workflow: named CHNS-module acquisition, construction and audit of the harmonized person-wave input, cohort reconstruction from the official integrated physical examination file, pinned dependencies, parameter and variable specifications, executable analyses, and non-identifying aggregate outputs.

2.1 Study design, data source, and study population

CHNS is an ongoing longitudinal cohort initiated in 1989 that monitors how socioeconomic transformation affects health and nutrition across multiple Chinese provinces through repeated individual-, household-, and community-level measurements (19, 20). CHNS household and individual files were obtained through the official CHNS/UNC data portal under the applicable data-use conditions and harmonized across survey years. Individual dietary intake was assessed on three consecutive days using 24-h recalls, while household food inventories and condiments were weighed over the same period; CHNS investigators constructed average daily energy and macronutrient intakes using the Chinese Food Composition Tables (20, 23). We used the official constructed daily energy and gram intakes and calculated protein, fat, and carbohydrate energy percentages using 4, 9, and 4 kcal/g, respectively. We separately attempted to reconstruct micronutrient and food group intakes from local CHNS food-detail records, but incomplete food-code mapping produced implausible values; these reconstructions were excluded from all primary analyses (Supplementary Table S12).

The cleaned adult dataset comprised 110,325 individual-wave records from 30,666 adults aged ≥18 years across waves 1989–2015; constructed macronutrient variables were available for 1991–2011. We used the official integrated physical examination master, which includes blood pressure measurements through 2015. At the first wave meeting dietary, energy, and outcome criteria and not reporting current pregnancy or breastfeeding, participants had to be normotensive, have all three macronutrient percentages, and have plausible energy intake (men 800–5,000 kcal/day; women 500–4,000 kcal/day). Missing pregnancy/breastfeeding flags were retained because these items are not applicable to men and were not administered uniformly; only reported current pregnancy or breastfeeding triggered exclusion. A participant with hypertension at or before this baseline was excluded. Incident hypertension was then ascertained in each subsequent wave with a valid outcome, regardless of whether follow-up dietary data were available. Participants without an event were censored at their last valid outcome wave. Primary exposures used the three-day average at baseline; repeated post-baseline diet was used only in an additional exploratory sensitivity analysis and never in prediction.

Among 24,114 adults with at least one diet-, energy-, and outcome-valid candidate wave, 175 had no candidate without reported pregnancy/breastfeeding and 411 had baseline deferred to a later eligible wave. We further excluded 4,755 adults with hypertension at or before baseline and 4,800 without a later valid outcome, leaving 14,384 participants. Study size was determined by including all adults who met these criteria in the available CHNS data; no a priori sample-size calculation was performed. Missingness was 1.6% for education, 0.7% for income, 1.3% for smoking, 0.6% for alcohol, and 0.8% for body-mass index (BMI); systolic blood pressure (SBP) and diastolic blood pressure (DBP) were complete by the strict baseline-outcome definition. Physical activity was unavailable for 80.3% and was examined only in a restricted sensitivity analysis. Cox models used model-specific complete cases and report their sample and event counts.

2.2 Dietary exposure assessment

The dietary exposures began with percentages of total reported energy from protein (%E), fat (%E), and carbohydrate (%E), computed from the official CHNS constructed-nutrient file using standard Atwater conversion factors (protein and carbohydrate, 4 kcal/g; fat, 9 kcal/g). Dietary supplements were not part of the macronutrient exposure definition; the analysis concerns food-derived energy as recorded by CHNS. Primary substitution analyses required the three calculated shares to sum to 95%–105%, a pragmatic quality-control window intended to tolerate small discrepancies from rounding and non-macronutrient energy while excluding materially inconsistent records, and then divided each share by their sum so that the closed protein, fat, and carbohydrate shares totalled exactly 100%. Because this closure rescales to macronutrient-derived energy alone, a closed share is expressed as a percentage of macronutrient-derived energy rather than as a percentage of total reported energy. Closed protein and fat shares were entered simultaneously with total reported energy in a relative-mixture parameterization analogous to a nutrient-density model (24); each coefficient therefore represents a five-composition-point increase at the expense of carbohydrate, conditional on total reported energy. The corresponding unclosed, composition-valid total-energy shares were analyzed as a conventional nutrient-density sensitivity check. To assess sensitivity to the pragmatic composition screen, Model 3 was repeated using 97.5%–102.5% and 90%–110% windows and with no sum-window restriction.

As a secondary summary measure, we constructed an AMDR concordance count anchored to the Chinese Dietary Reference Intakes (DRI; 2023 edition) (25). Among composition-valid records, each macronutrient contributed one when its original percentage of total reported energy was in range: protein 10%E-20%E at ages 18–64 years and 15%E-20%E at ages ≥65 years, fat 20%E-30%E, and carbohydrate 50%E-65%E; the indicators were summed (0–3; Supplementary Table S1). No partial credit, weighting, or distance penalty was applied. This count is not a validated diet-quality index and omits sodium, potassium, fiber, micronutrients, fat quality, and food groups. We examined an AMDR count calculated after closure, a macronutrient-risk count, a binary indicator for at least two risk features, and the fully specified original four-component proxy, only as sensitivities (Supplementary Table S1).

2.3 Outcome definition and covariates

Blood pressure (BP) was summarized as the mean of up to three readings after excluding SBP outside 70–260 mmHg and DBP outside 40–160 mmHg. Measured hypertension was present if mean SBP was ≥140 mmHg or mean DBP was ≥90 mmHg; a negative outcome required both SBP and DBP to be observed and below threshold. A self-reported physician diagnosis or antihypertensive medication use was sufficient for a positive composite outcome even if measured BP was unavailable, but a negative self-report alone was not treated as normotension. Incident hypertension was the first later positive wave. Because ascertainment occurred at discrete visits, the true onset is interval censored. Primary Cox analyses assigned events to the detection wave; discrete-time and midpoint alternatives are described below. A measured-BP only outcome was examined in sensitivity analysis.

Covariates were age, sex, province, baseline year, urban/rural residence, the CHNS urbanization index, education, per capita household income (standardized to 2015 values), smoking, alcohol use, body-mass index (BMI), baseline SBP, baseline DBP, and total energy intake. Education used the categorical highest-attainment item: no formal education, primary school, lower middle school, upper middle school, technical/vocational training, and college/university or higher (the last category combined the two highest codes); code 9 was treated as missing. CHNS does not provide valid cross-sectional or longitudinal sampling weights, as documented in the official methodology (19); analyses were therefore unweighted.

2.4 Statistical analysis

Associations with incident hypertension were estimated using sequentially adjusted Cox proportional hazards models (26). Unpenalized partial likelihood used the Efron method for tied event times; sandwich standard errors were clustered by the CHNS community identifier (27). Model 1 adjusted for age, sex, province (as fixed effects), baseline year, and total energy; Model 2 added urban/rural residence, the urbanization index, education, and income; Model 3 further included smoking and alcohol use and served as the primary confounder-adjusted model; Model 4 additionally included BMI, baseline SBP, and baseline DBP and was treated as a secondary mediator-adjusted model. Closed protein and fat composition shares were entered simultaneously with total reported energy, so that each coefficient represents a relative closed composition increase at the expense of carbohydrate, conditional on total reported energy. The AMDR concordance count and other sensitivity exposures were entered as single continuous variables.

The proportional hazards assumption was assessed with variable-specific Schoenfeld-residual tests from the corresponding unclustered Model 3 fit (28). The exposure tests did not reject proportionality, but age, baseline year, and sex did in at least one primary model. We therefore fitted a sensitivity model stratified by seven baseline-age groups, baseline year, and sex while retaining age within each stratum as a continuous covariate. Records censored before the first event in their joint stratum, which do not enter any risk set or contribute to the partial likelihood, were excluded from this stratified sensitivity analysis only. Further analyses included a discrete-time complementary log–log model with interval-specific intercepts and a log-interval-length offset; a Cox analysis assigning events to interval midpoints; measured-BP only hypertension; age strata with robust interaction tests; restricted physical-activity analyses; and the original proxy. AMDR was additionally fitted as a four-level categorical variable. Five-knot restricted cubic splines tested departures from linear protein and fat effects using community-clustered robust Wald tests; Benjamini–Hochberg false-discovery-rate correction covered the three functional-form tests (29). Finally, a counting-process Cox sensitivity split follow-up at every valid dietary assessment and outcome visit. At each interval start, exposure was the cumulative mean of all composition-valid dietary waves observed up to that time; a diet measured at an event-detection visit was not used retrospectively for the interval ending with that event. Variance was clustered at the community level. This repeated-diet model is exploratory because the interval-censored onset, time-varying confounding, and diet changes in response to health remain possible. Confidence intervals, rather than post-hoc observed power, were used to characterize precision (30).

To examine secular patterns, we computed directly standardized annual means across diet-observed, sex-specific energy-plausible, composition-valid person-waves within the seven-province panel present in every diet wave (provinces 32, 37, 41, 42, 43, 45, and 52). Macronutrient composition shares were closed to 100%; AMDR classifications used the original total-energy shares. The standard population was the pooled covariate distribution across the age-group, sex, province, and urban/rural cells represented for every reported outcome in every diet wave. We estimated macronutrient shares, the age-specific AMDR count, and hypertension prevalence under composite and measured-BP-only definitions, with 95% confidence intervals from 500 community-level exponential-multiplier bootstrap replicates. A sensitivity analysis included all provinces available in each wave, standardized to common age-group, sex, and urban/rural cells while allowing the province mix to vary (Supplementary Figure S3). These are repeated cross-sectional, ecological descriptions and are neither nationally representative prevalence estimates nor evidence of individual-level causation.

Machine learning was used solely as an exploratory concordance check (31). The fixed target was incident hypertension within 6 years among composition-valid baselines through 2009. A case required detection by 6 years; a non-case required a documented normotensive assessment at or after 6 years. When the last normal visit preceded 6 years, and the next positive visit or censoring occurred after the horizon, 6-year status was classified as unknown rather than a non-case. These 2,871 records contributed to the censoring model but not the binary endpoint. The binary endpoint was verified for 9,284 participants, who were evaluated using inverse-probability-of-censoring weighting (IPCW). An untuned L2-penalized logistic model (C = 1, lbfgs solver) and an untuned random forest (400 trees, maximum depth 8, minimum leaf size 12) were evaluated with nested feature sets in temporal validation (1991/1993 training, n = 5, 265; 1997–2009 testing, n = 4, 019). Censoring weights were estimated separately within the early training and later test periods. The baseline year was excluded from the predictors. The weighted evaluation metrics were area under the receiver operating characteristic curve (AUC), area under the precision-recall curve (AUPRC), Brier score, calibration intercept, and calibration slope. The diet-inclusive random forest was designated for held-out SHapley Additive exPlanations (SHAP) before temporal performance was evaluated (32, 33); neither model selection nor SHAP refitting used test outcomes. SHAP is model-internal and associational, and unavailable features such as sodium and food groups cannot be ranked (34). Post-baseline dietary trajectories were excluded from prediction to prevent future-information leakage.

3 Results

3.1 Cohort characteristics and the macronutrient transition

The source data comprised 110,325 adult person-waves from 30,666 adults (Figure 1). Of 24,114 adults with at least one diet-, energy-, and outcome-valid candidate wave, 175 had no wave without reported pregnancy/breastfeeding, 4,755 had hypertension at or before baseline, and 4,800 lacked a later valid outcome. The final cohort included 14,384 adults; 5,234 (36.4%) developed hypertension during a median of 7.0 years (IQR 4–15) of follow-up. Adults who developed hypertension were older (43.7 vs. 35.5 years), had higher BMI (22.5 vs. 21.8 kg/m2), SBP (115.1 vs. 110.6 mmHg), and DBP (74.8 vs. 72.2 mmHg), and more often smoked or used alcohol (Table 1).

Figure 1

Table 1

CharacteristicOverallNo incident HTNIncident HTN
(n = 14, 384)(n = 9, 150)(n = 5, 234)
Age, years38.5 (14.4)35.5 (13.7)43.7 (14.2)
Male, n (%)6,721 (46.7)4,028 (44.0)2,693 (51.5)
Urban residence, n (%)5,300 (36.8)3,610 (39.5)1,690 (32.3)
Urbanization index56.7 (21.6)59.5 (21.5)51.8 (20.9)
Per-capita income (2015 CNY)6,975.0 (9,918.1)7,833.8 (10,867.6)5,475.1 (7,770.4)
Current smoker, n (%)4,268 (30.1)2,423 (26.8)1,845 (35.8)
Alcohol use, n (%)5,041 (35.3)2,940 (32.3)2,101 (40.4)
Body-mass index, kg/m222.0 (3.1)21.8 (3.0)22.5 (3.1)
Systolic BP, mmHg112.2 (11.7)110.6 (11.5)115.1 (11.7)
Diastolic BP, mmHg73.1 (8.2)72.2 (8.2)74.8 (7.9)
Total energy, kcal/day2,362.2 (710.4)2,294.8 (700.1)2,480.1 (713.0)
Protein share of macro-derived energy, %12.5 (2.8)12.7 (3.0)12.1 (2.6)
Fat share of macro-derived energy, %26.0 (12.2)27.2 (12.2)24.0 (11.9)
Carbohydrate share of macro-derived energy, %61.5 (12.8)60.2 (12.8)63.9 (12.4)
AMDR concordance count (0–3)1.5 (0.9)1.5 (0.9)1.4 (0.9)
Education, n(%)
 No formal education3,038 (21.5)1,346 (14.9)1,692 (32.9)
 Primary school2,722 (19.2)1,618 (17.9)1,104 (21.5)
 Lower middle school4,666 (33.0)3,300 (36.6)1,366 (26.6)
 Upper middle school1,987 (14.0)1,422 (15.8)565 (11.0)
 Technical/vocational education892 (6.3)648 (7.2)244 (4.7)
 College/university or higher853 (6.0)682 (7.6)171 (3.3)

Baseline characteristics by incident-hypertension status.

Values are mean (SD) or n (%); denominators vary with missing data. Income is standardized to 2015 CNY. The table is descriptive; no baseline significance tests were performed.

Across the eight diet waves from 1991 to 2011 in the qualifying seven-province trend sample, the crude mean closed fat share rose from 21.8% to 32.5% of macronutrient-derived energy, carbohydrate fell from 66.4% to 54.3%, and protein rose from 11.9% to 13.1% (Figure 2). These changes reproduce the documented Chinese nutrition transition (9, 10).

Figure 2

3.2 Macronutrient composition and incident hypertension

In primary Model 3 (13,276 complete cases; 4,745 events), closed protein replacing carbohydrate was not associated with incident hypertension [hazard ratio (HR) per five composition points 0.986, 95% confidence interval (CI) 0.920–1.056; p = 0.679], and closed fat replacing carbohydrate crossed the null (HR 0.986, 95% CI 0.968–1.004; p = 0.120; Table 2, Figure 3). The total-energy AMDR count was also null in the same analysis sample (HR per additional in-range component 1.007, 95% CI 0.974–1.041; p = 0.679). Unclosed composition-valid substitution estimates were nearly identical. The closed-share AMDR sensitivity (HR 1.008, 95% CI 0.975–1.042), macronutrient-risk count (HR 1.023, 95% CI 0.979–1.069), binary multi-risk indicator (HR 1.013, 95% CI 0.940–1.092), and original proxy (HR 0.997, 95% CI 0.985–1.009) crossed the null. The closed-fat estimate was borderline in Model 1 (p = 0.051) and crossed the null in all four models.

Table 2

ExposureModel 1Model 2Model 3Model 4
Substitution n/events13,790/4,94413,484/4,82913,276/4,74513,168/4,697
Closed protein share (per +5 points)0.9790.9840.9860.985
(0.914–1.048)(0.917–1.055)(0.920–1.056)(0.920–1.055)
Closed fat share (per +5 points)0.9830.9840.9860.985
(0.967–1.000)(0.967–1.001)(0.968–1.004)(0.967–1.003)
AMDR n / events13,790/4,94413,484/4,82913,276/4,74513,168/4,697
AMDR count (per +1)1.0021.0051.0070.988
(0.969–1.036)(0.972–1.039)(0.974–1.041)(0.955–1.023)

Sequentially adjusted Cox proportional-hazards associations with incident hypertension.

Hazard ratios (95% CI) use Efron handling of tied event times, community-clustered robust standard errors, and model-specific complete cases. Model 3 is the primary confounder-adjusted estimate; Model 4 additionally conditions on potential mediators. Protein and fat coefficients represent relative closed composition increases at the expense of carbohydrate, conditional on total reported energy.

Closed-share points are percentage points of macronutrient-derived energy; carbohydrate is replaced.

M1: age, sex, province, baseline year, total energy. M2: +urban residence, urbanization, education, income.

M3: +smoking and alcohol. M4: +BMI, baseline SBP, and baseline DBP.

All models use community-clustered robust standard errors.

Figure 3

Schoenfeld residual tests did not reject proportionality for protein (p = 0.302), fat (p = 0.558), or AMDR count (p = 0.502); age, baseline year, and sex showed violations (Supplementary Table S4). The stratified analysis remained null (13,273 persons after excluding three records that entered no risk set; protein HR per five composition points 0.977, 95% CI 0.912–1.047; fat HR 0.985, 95% CI 0.967–1.003; AMDR HR 1.005, 95% CI 0.971–1.039). The complementary log-log analysis yielded protein HR 0.983 (95% CI 0.917–1.055), fat HR 0.986 (95% CI 0.968–1.005), and AMDR HR 1.008 (95% CI 0.975–1.043). Midpoint Cox estimates were directionally concordant. Using macronutrient-sum windows of 97.5%–102.5% or 90%–110%, or removing the sum-window restriction, produced protein HRs of 0.980–0.997, fat HRs of 0.984–0.987, and AMDR-count HRs of 1.001–1.009; every confidence interval crossed the null (Supplementary Tables S3, S4, S13).

3.3 Standardized secular trends

After direct standardization among qualifying seven-province person-waves, the mean age-specific AMDR count increased from 1.367 (95% CI 1.325–1.418) in 1991 to 1.525 (95% CI 1.488–1.567) in 2011 (Figure 4). Over the same period, composite hypertension prevalence increased from 17.4% (95% CI 16.1–18.5) to 22.6% (95% CI 21.5–23.7), and measured-BP only prevalence increased from 16.4% (95% CI 15.1–17.5) to 18.6% (95% CI 17.6–19.6). The all-province sensitivity analysis was qualitatively similar but numerically distinct (Supplementary Figure S3). These estimates describe the selected diet-valid CHNS panel, not national prevalence. Their ecological co-movement neither supports a causal link nor establishes that AMDR concordance prevents hypertension.

Figure 4

3.4 Machine-learning concordance and sensitivity analyses

Among 9,284 verified 6-year endpoints, traditional-predictor logistic regression had the numerically highest temporal AUC (weighted AUC 0.785, AUPRC 0.459, Brier 0.118; Table 3); performance differences were not formally tested. Adding socioeconomic/lifestyle and dietary variables did not improve logistic discrimination (diet-inclusive AUC 0.777). Random forest AUCs were 0.775 with traditional predictors, 0.777 after adding socioeconomic/lifestyle variables, and 0.778 after diet; the diet increment over the socioeconomic/lifestyle set was 0.001. The designated diet-inclusive random forest remained miscalibrated (intercept 0.275; slope 1.373; Supplementary Figure S2). In held-out SHAP analysis, age, baseline SBP, baseline DBP, and BMI dominated; mean absolute values were 0.0056 for protein and 0.0029 for fat (Figure 5). Baseline year was not a predictor. These rankings are model-internal and cannot quantify the importance of unmeasured dietary dimensions.

Table 3

Model and feature setAUCAUPRCBrierΔAUC
Logistic, traditional0.7850.4590.118–
Logistic, +SES/lifestyle0.7810.4500.118−0.004
Logistic, +diet0.7770.4430.119−0.007
Random forest, traditional0.7750.4400.120–
Random forest, +SES/lifestyle0.7770.4350.119+0.002
Random forest, +diet0.7780.4350.119+0.003

IPCW temporal validation of verified 6-year incident-hypertension endpoints (train: 1991/1993 baselines, n = 5, 265; test: 1997–2009 baselines, n = 4, 019).

A non-case required documented normotension at or after 6 years; 2,871 status-unknown intervals were excluded from binary evaluation but retained in the censoring model. Baseline year was excluded from predictors. Delta AUC is relative to the traditional feature set for the same algorithm. Performance differences were not formally tested. SES, socioeconomic status.

Figure 5

Age interaction tests were significant jointly for protein and fat (p = 0.016) and for fat specifically (p = 0.007), but not for protein (p = 0.243) or the AMDR count (p = 0.582; Supplementary Table S5 and Supplementary Figure S1). Fat replacing carbohydrate was nominally inverse only among adults aged < 40 years (HR per five composition points 0.9766, 95% CI 0.9541–0.9996; p = 0.046); all other within-age-group estimates crossed the null. AMDR categories were jointly null (p = 0.201). Protein nonlinearity was nominal before multiple-testing correction (p = 0.041) but not after false-discovery-rate correction (p = 0.122); fat and joint nonlinearity tests were null after correction. In the cumulative-average repeated-diet model, whose 37,643 risk intervals were updated at every valid dietary wave (13,276 participants; 4,745 events), fat replacing carbohydrate was inversely associated with incidence (HR per five composition points 0.954, 95% CI 0.930–0.978; false-discovery-rate-adjusted p < 0.001), whereas protein and AMDR crossed the null. This post-review exploratory contrast may reflect reduced exposure error, dietary change, interval-censored onset, or time-varying confounding and is not a basis for dietary advice. Physical-activity adjustment in the available subset produced negligible changes (n = 2, 521; 853 events). Measured-BP only results crossed the null (fat HR per five composition points 0.987, 95% CI 0.969–1.006; AMDR HR 1.008, 95% CI 0.972–1.045; Supplementary Tables S6–S9).

4 Discussion

In this longitudinal cohort spanning more than two decades of China's nutrition transition, single-baseline closed protein and fat shares replacing carbohydrate and the age-specific AMDR count crossed the null after confounder adjustment. Dietary variables produced no material improvement in 6-year temporal prediction, and both the standardized AMDR count and hypertension prevalence increased from 1991 to 2011. Taken together, the primary results indicate that a single gross macronutrient measurement contributed little information about individual risk; they do not test the effects of sodium, potassium, food groups, or fat and carbohydrate quality. The repeated-diet analysis was post-review and exploratory and is interpreted separately below.

4.1 Interpretation and biological plausibility

The primary AMDR confidence interval (0.974–1.041 per additional in-range component) excludes large associations on the fitted scale but does not prove equivalence. Protein and fat substitution intervals likewise include the null and remain compatible with small effects. Three recall days may not represent usual long-term intake, and measurement error can attenuate associations. The cumulative-average analysis partly addresses this limitation but introduces susceptibility to diet changes after emerging health concerns and to time-varying confounding; its approximately 5% lower hazard per five composition points higher fat therefore requires replication with explicit longitudinal confounding control. Exact event onset is unknown between survey waves, although discrete-time and midpoint analyses were concordant. Neither the baseline nor repeated-diet estimates distinguish unsaturated from saturated fat or whole from refined carbohydrate and should not support recommendations about macronutrient quantity or quality.

These near-null and small estimates are biologically plausible because macronutrient energy percentages are only weak proxies for the physiological mechanisms governing long-term arterial pressure regulation. Blood pressure is strongly influenced by sodium-volume homeostasis and potassium balance; evidence from INTERSALT and potassium-intervention studies underscores the importance of sodium and potassium for blood pressure (35, 36). These mineral quantities are not determined by macronutrient distribution: a diet with a given fat or carbohydrate share can carry widely different sodium loads depending on discretionary salt, sauces, and processed foods. The original DASH trial established the effect of a fruit-, vegetable-, and low-fat-dairy-rich pattern, DASH-Sodium separately demonstrated additional benefit from sodium reduction, and OmniHeart compared macronutrient substitutions within a DASH-like pattern (12, 13, 37). An identical fat or carbohydrate share can be achieved through opposite food choices (unsaturated vs. saturated fat, whole vs. refined carbohydrate) whose vascular implications a share-based metric averages away (38, 39). Potassium-enriched salt substitution also reduced cardiovascular events in a Chinese high-risk rural population, rather than in the CHNS cohort itself (16). Consistent with this distinction, a CHNS study found a hypertension signal when food-group-level patterns were measured (18).

4.2 Descriptive secular trends and cross-method concordance

The standardized AMDR count increased modestly from 1991 to 2011 while hypertension prevalence also increased. Parallel movement does not establish concordance, because this three-component range count cannot capture minerals, food group quality, adiposity, or other non-dietary determinants. Direct standardization addresses measured differences in age, sex, province, and residence, but not concurrent changes in diagnostic practice, medication use, adiposity, sodium intake, or the built environment. The trend comparison is therefore descriptive; the individual-level evidence comes from the survival models.

The machine learning analysis adds a prediction perspective rather than an independent causal test. Traditional-predictor logistic regression had the numerically highest temporal AUC, adding diet reduced its discrimination, and random forest performance changed little; these differences were not formally tested. Held-out SHAP confirmed that age and baseline blood pressure dominated 6-year prediction. The corrected endpoint deliberately excluded status-unknown intervals rather than calling them non-cases, but IPCW still depends on an adequate censoring model. Machine learning does not resolve the nominal protein non-linearity or the cumulative-average fat association: both approaches share the same cohort, recall data, and incomplete feature space (31).

4.3 Implications for dietary surveillance

These findings have practical implications for dietary monitoring systems. Macronutrient surveillance retains value for tracking energy and protein adequacy, but the present results do not support using a single gross-composition measurement as a standalone hypertension-risk instrument. Together with trial evidence (13, 16) and Chinese hypertension management guidelines (40), the findings support the collection of sodium, potassium, dietary fiber, fat quality, and food groups, ideally calibrated against urinary biomarkers where feasible. Repeated dietary measurement remains methodologically important, although the exploratory difference between baseline and cumulative-average estimates cannot establish a longitudinal substitution effect. A version-controlled food-code-to-composition-table mapping should be maintained across waves, and dietary monitoring should be linked with standardized repeated blood pressure measurement.

4.4 Strengths and limitations

The study's strengths include 14,384 adults and 5,234 incident detections over two decades, centrally constructed macronutrient data, explicit composition closure, age-specific Chinese DRI ranges, the exclusion of visits reporting current pregnancy or breastfeeding from baseline selection, a strict measured-BP requirement for negative outcomes, sequential adjustment, community-clustered robust standard errors, alternative event-time and functional-form models, repeated-diet and measured-BP-only analyses, interval-censoring-safe temporal validation, and an end-to-end reproduction package.

This study has several limitations. Three recall days may not capture usual intake, and no formal measurement-error correction was possible. The 95%–105% composition-sum screen is a pragmatic quality-control rule rather than a validated biological cut-off; alternative-window analyses reduce, but do not eliminate, concern about selection from this rule. Model-specific complete-case analysis may also have introduced selection bias. In addition, 4,800 otherwise eligible baseline-normotensive adults had no later valid outcome and could not enter the survival risk sets. Compared with retained participants, they entered later (standardized mean difference [SMD] 0.606), were more urbanized (SMD 0.377), younger (SMD 0.332), and reported lower energy intake (SMD 0.349; Supplementary Table S13). Informative retention and censoring therefore remain possible because the primary Cox models were not selection-weighted. The cumulative-average model uses repeated observations but cannot eliminate reverse causation or time-varying confounding. The composite, interval-censored outcome combines incident with newly detected hypertension; alternative time assignments do not recover the exact onset date. Physical activity was unavailable for 80.3% of the cohort and could be examined only in a smaller subset. BMI and baseline blood pressure may act as both confounders and mediators; they were therefore reserved for Model 4. Competing risks were not modeled, and prediction-performance differences were not formally tested. IPCW depends on correct modeling of censoring, and the prediction analysis was not externally validated. CHNS samples selected provinces and is not nationally representative (19). The nominal age and protein functional-form findings arose among multiple exploratory tests. Robustness across specifications reduces, but cannot remove, these uncertainties.

5 Conclusion

Among 14,384 Chinese adults, single-baseline closed protein and fat shares replacing carbohydrate and an age-specific Chinese DRI AMDR count crossed the null after confounder adjustment, and dietary features did not materially improve verified 6-year temporal prediction. These primary results do not support gross macronutrient composition as a standalone hypertension surveillance instrument. Sodium, potassium, fiber, fat quality, and food-group information remain necessary; exploratory repeated-diet findings require dedicated longitudinal confirmation.

Statements

Data availability statement

The CHNS household- and individual-level files analyzed in this study are available through the official CHNS/UNC data portal (https://chns.cpc.unc.edu/data/datasets/index/). Community-level files require an approved CHNS data-use agreement. Access and use remain subject to the applicable CHNS conditions. Supplementary Code 1 contains the complete reproducibility workflow: a manifest of every required CHNS input with acquisition identifiers where available, byte counts, and SHA-256 checksums; the raw-module-to-person-wave harmonization code and configuration; a schema audit; the cohort and analysis script; pinned dependencies; a parameter manifest; a variable dictionary; scoring and model specifications; food-code audit summaries; non-identifying aggregate outputs; and a step-by-step reproduction guide. The supplied builder was independently verified to reconstruct 110,325 person-waves from 30,666 adults with no differences in the 17 analysis-required fields relative to the development input. CHNS individual-level records and community identifiers are not redistributed; authorized users can recreate them locally from the official files.

Ethics statement

The studies involving humans were approved by Institutional Review Board of the University of North Carolina at Chapel Hill, Chapel Hill, North Carolina, USA; and the Institutional Review Board (Ethics Review Committee) of the National Institute for Nutrition and Health, Chinese Center for Disease Control and Prevention, Beijing, China. The studies were conducted in accordance with the local legislation and institutional requirements. Written informed consent for participation was not required as the present work is a secondary analysis of de-identified CHNS data obtained under the applicable data-use conditions; no new participant contact or data collection was undertaken.

Author contributions

KX: Conceptualization, Data curation, Formal analysis, Methodology, Software, Visualization, Writing – original draft. XZ: Conceptualization, Project administration, Supervision, Writing – review & editing.

Funding

The author(s) declared that financial support was not received for this work and/or its publication.

Acknowledgments

This analysis uses the China Health and Nutrition Survey (CHNS). We acknowledge the institutions and funders that made the underlying survey possible: the US National Institutes of Health, including the Eunice Kennedy Shriver National Institute of Child Health and Human Development, the National Institute on Aging, the National Institute of Diabetes and Digestive and Kidney Diseases, and Fogarty support (R01 HD30880, R01 AG065357, R01 DK104371, R01 HL108427, and D43 TW009077); the China-Japan Friendship Hospital and Ministry of Health for the 2009 CHNS; the Chinese National Human Genome Center at Shanghai from 2009; and the Beijing Municipal Center for Disease Prevention and Control from 2011. We also thank the National Institute for Nutrition and Health of the Chinese Center for Disease Control and Prevention and the other participating CHNS institutions and survey teams. These acknowledgments concern support for CHNS data collection, rather than direct funding of this secondary analysis.

Conflict of interest

The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Generative AI statement

The author(s) declared that Generative AI was used in the creation of this manuscript. The authors used Claude to polish the English language and ChatGPT to check the code. The authors take full responsibility for the content of the manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fnut.2026.1924089/full#supplementary-material

References

Summary

Keywords

diet quality, hypertension, machine learning, macronutrients, nutrition transition

Citation

Xiang K and Zhu X (2026) Macronutrient composition and incident hypertension during China's nutrition transition: a longitudinal cohort analysis of the China Health and Nutrition Survey. Front. Nutr. 13:1924089. doi: 10.3389/fnut.2026.1924089

Received

30 June 2026

Revised

15 September 2026

Accepted

25 September 2026

Published

04 November 2026

Volume

13 - 2026

Edited by

Alessandra Durazzo, Council for Agricultural Research and Economics, Italy

Reviewed by

Kun Huang, Huazhong University of Science and Technology, China

Yanzhu Yao, Chongqing Medical University, China

Updates

Copyright

*Correspondence: Xiangbo Zhu,

Disclaimer

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.

Outline

Figures

Cite article

Copy to clipboard


Export citation file


Share article

Article metrics