Longitudinal Trajectories of Body Roundness Index and Risk of Incident Hypertension in Non-Diabetic Elderly Adults: A Prospective Nationwide Cohort Study Involving Two Cohorts

What is new?

This study demonstrates that long-term BRI trajectories are independently associated with hypertension risk in non-diabetic adults across two large cohorts. Persistently high BRI trajectories confer the greatest risk. Incorporating BRI trajectories into machine learning models significantly enhances predictive performance and risk stratification. This study provides the first large-scale evidence for integrating longitudinal BRI trajectories with explainable machine learning for hypertension prediction.

What are the clinical implications?

Longitudinal BRI monitoring may enable earlier and more precise identification of individuals at high risk of hypertension, supporting personalized prevention before overt metabolic disease develops.

Study design and population

This study used data from CHARLS (14), a nationally representative, longitudinal survey of community-dwelling Chinese adults aged ≥45 years. CHARLS collects comprehensive information on sociodemographic characteristics, socioeconomic status, health behaviors, and biological markers. The baseline survey (Wave 1, 2011–2012) recruited 13,523 participants from 10,257 households across 28 provinces. Follow-up assessments were conducted biennially using face-to-face computer-assisted personal interviews, with blood samples collected in alternate waves. For the present analysis, data were obtained from Wave 1 and four subsequent follow-ups: Wave 2 (2013–2014), Wave 3 (2015–2016), and Wave 4 (2018–2020). Participant selection is shown in Figure 1. Participants were excluded if they had any of the following: (1) incomplete baseline nurse visit data required for body roundness index (BRI) assessment; (2) fewer than two visits with available BRI measurements; (3) age <45 years or a diagnosis of diabetes before follow-up; (4) prevalent hypertension before follow-up; or (5) incomplete follow-up information for incident hypertension. Participants were required to have valid BRI measurements at a minimum of two of the three trajectory-assessment waves. BRI measurements from Wave 1 (2011–2012), Wave 2 (2013–2014), and Wave 3 (2015–2016) were used to construct BRI trajectories, and incident hypertension was subsequently assessed at Wave 4 (2018–2020). A total of 4,522 participants was ultimately included in the analysis examining the association between BRI trajectories and incident hypertension. Ethical approval for CHARLS was obtained from the Biomedical Ethics Review Committee of Peking University (IRB 00001052-11015).

Figure 1

To externally validate the findings, data from ELSA (15) were used. ELSA is a nationally representative, prospective cohort study of community-dwelling adults aged ≥50 years in England. BRI data were available in Wave 2 (2004–2005), Wave 4 (2008–2009), Wave 6 (2012–2013), and Wave 8 (2016–2017). Consistent with the requirements of trajectory modeling in the CHARLS cohort, BRI measurements from Wave 2, Wave 4, and Wave 6 were used to construct BRI trajectories, with follow-up for incident hypertension until Wave 8. Except for the age criterion (≥50 years), the inclusion and exclusion criteria were consistent with those applied in the CHARLS cohort. A total of 1,105 participants was included in the validation analysis. Ethical approval for the ELSA study was granted by the London Multi-Centre Research Ethics Committee (MREC/01/2/91). All participants in both cohorts provided written informed consent. This study followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guidelines (16).

Assessment of body roundness index and trajectory modeling

The body roundness index was calculated using the formula proposed by Thomas et al. (8). $\text{BRI}=364.2-365.5\times \sqrt{1-{\left(\frac{WC}{2\pi }\right)}^2/{\left(0.5\times \mathit{\text{Height}}\right)}^2}$ where WC (waist circumference) and height are measured in meters. To characterize the dynamic patterns of BRI over time, group-based trajectory modeling (GBTM) (17) was applied using the PROC TRAJ procedure in SAS software. In the CHARLS cohort, BRI measurements from Waves 1–3 (2011–2016) were used to construct longitudinal trajectories, and incident hypertension was subsequently assessed at Wave 4 (2018–2020). In the ELSA cohort, BRI measurements from Waves 2, 4, and 6 (2004–2013) were used for trajectory construction, and incident hypertension was assessed at Wave 8 (2016–2017). We adhered to principles of model parsimony and clinical interpretability. Models with two to four groups were fitted, testing linear, quadratic, and cubic functional forms for each group.

Model selection was guided by the following criteria: (1) minimization of the Bayesian Information Criterion (BIC); (2) adequate sample size per group (not less than 5% of the total population); (3) average posterior probability (AvePP) of assignment >0.70; (4) odds of correct classification (OCC) > 5; and (5) distinct visual separation of trajectories conforming to clinical plausibility.

Outcome definition and covariates

The outcome was incident hypertension, assessed at Wave 4 (2018–2020) in CHARLS and at Wave 8 (2016–2017) in ELSA among participants without hypertension during the corresponding trajectory-assessment period. Hypertension was defined as a systolic blood pressure (SBP) ≥ 140 mmHg and/or a diastolic blood pressure (DBP) ≥ 90 mmHg, current use of antihypertensive medication, or a self-reported physician diagnosis of hypertension (based on the question, ‘Have you been diagnosed with hypertension by a doctor?’). The diagnostic criteria for hypertension in the ELSA cohort were identical to those applied in the CHARLS cohort.

Covariates assessed from Waves 1–4 in the CHARLS cohort included the following: (1) sociodemographic characteristics (age, sex, education level, and residence); (2) lifestyle behaviors (smoking status and alcohol consumption); (3) health status (history of cardiovascular and metabolic diseases); and (4) biomarkers (C-reactive protein, triglycerides, and remnant cholesterol). Continuous covariates were summarized using their mean values across the exposure period, where appropriate. Education level was categorized as below high school or high school and above. Residence was classified as rural or urban. Marital status was categorized as married or living with a stable partner versus other statuses (divorced, separated, widowed, or single). Alcohol consumption was defined as drinking alcohol at least once per week. Dyslipidemia, lung disease, kidney disease, and heart disease were identified based on self-reported physician diagnoses. We imputed the missing value for further exploration using the K Nearest Neighbors imputation method.

Statistical analysis

Baseline characteristics were compared across BRI trajectory groups using analysis of variance (ANOVA) or Kruskal–Wallis tests for continuous variables, and χ² tests for categorical variables. Multivariable logistic regression models were employed to estimate odds ratios (ORs) and 95% confidence intervals (CIs) for the association between BRI trajectories and incident hypertension. As a separate but complementary predictive analysis, a structured ML framework was implemented to assess whether BRI trajectory information could be integrated with clinical variables for multivariable hypertension risk prediction. Prior to model development, correlation analyses were conducted among all clinical variables. For pairs of variables with a correlation coefficient greater than 0.6, only one variable was retained to reduce multicollinearity among the independent variables in the models. Subsequently, feature selection was performed using a Light Gradient Boosting Machine (Light-GBM (18, 19)) to identify the most informative predictors based on feature importance rankings. Specifically, feature importance was first quantified using the Light-GBM model, and variables were ranked accordingly. Sequential forward feature selection was then performed by iteratively adding variables in order of importance until the inclusion of additional features resulted in an area under the receiver operating characteristic curve (AUC) increase of less than 0.1. The selected feature set was used for final model construction (20). Model performance was evaluated using 10-fold cross-validation for each candidate feature set, with the mean receiver operating characteristic (ROC) AUC calculated across folds. The 95% confidence intervals were estimated based on 10 fold-specific results and visualized to illustrate model performance uncertainty. To balance the positive and negative samples, the training set underwent preprocessing with random under-sampling and synthetic minority over-sampling technique (SMOTE) over-sampling. Subsequently, several ML algorithms were developed (21), including logistic regression (LR (22)), decision tree (DT (23)), random forest (RF (24)), support vector machine (SVM (25)), and gradient boosting decision tree (GBDT (26)). In addition, a stacking ensemble model (27) was constructed by combining SVM, GBDT, RF, and DT as base learners, with logistic regression serving as the meta-learner to integrate predictions from individual models (28, 29, 30). The stacking ensemble model was used to improve predictive performance by combining multiple heterogeneous base learners, thereby reducing model-specific bias and variance and enhancing generalization ability. The dataset was randomly split into a training set (70%) and a testing set (30%). Model training and hyperparameter tuning were conducted within the training set using 10-fold cross-validation. Model performance was evaluated in the testing set using multiple metrics, including accuracy, sensitivity, specificity, F1-score, area under the receiver operating characteristic curve, positive predictive value (PPV), and negative predictive value (NPV). To enhance model interpretability, Shapley Additive Explanations (SHAP) (31) were applied, particularly for tree-based models (29, 32). Mean absolute SHAP values were used to quantify global feature importance, while SHAP summary plots illustrated the direction and magnitude of each predictor’s contribution to hypertension risk. Furthermore, SHAP dependence plots with locally weighted regression (LOWESS) smoothing were used to explore potential non-linear relationships between key predictors and SHAP value.

Sensitivity analyses were conducted to examine the robustness of the main findings in this prospective cohort study. First, analyses were restricted to participants with complete data for all covariates. Second, the association between baseline BRI and the incidence of hypertension during follow-up was re-evaluated. Third, participants who were using lipid-lowering medications at baseline were excluded. Fourth, participants with a history of cancer at baseline were excluded. In addition, inverse probability weighting (IPW) logistic regression was applied to further reduce potential confounding bias. Furthermore, sensitivity analyses were conducted after excluding potential metabolic and clinical intermediates to re-evaluate the associations between BRI trajectory groups and incident hypertension. Finally, Cox regression was performed as a sensitivity analysis, with follow-up starting at Wave 3 after completion of BRI trajectory assessment. Participants with hypertension before or at Wave 3 were excluded, and follow-up continued until incident hypertension. All statistical analyses were performed using R software (version 4.5.2). Two-sided P values <0.05 were considered statistically significant.

Baseline characteristics

A total of 4,522 participants from the CHARLS cohort were included in the final group-based trajectory modeling analysis. Linear, quadratic, and cubic trajectory models with two, three, and four groups were fitted, respectively. Model selection was based on a comprehensive evaluation of fit indices, including the Bayesian Information Criterion (BIC), Akaike Information Criterion (AIC), entropy, and average posterior probabilities, as well as the substantive interpretability of the trajectory shapes (Tables S1–S2). On the basis of these criteria, a three-group model, comprising two quadratic trajectories and one linear trajectory, was identified as the optimal model (Table 1). Three distinct BRI trajectories were identified: low-level inverted U-shaped (trajectory 1; n = 2,183), moderate-level inverted U-shaped (trajectory 2; n = 1,971), and high-level increasing (trajectory 3; n = 368) patterns (Figure 2A). The incidence of hypertension was 12%, 17%, and 23% in trajectories 1, 2, and 3, respectively (Figure 2B). Baseline clinical characteristics according to BRI trajectory groups are summarized in Table 2. The mean age of the study population was 60.29 ± 8.51 years. Significant differences in baseline characteristics were observed across BRI trajectory groups. Compared with participants in the low-level inverted U-shaped BRI trajectory, those in the high-level increasing trajectory had markedly higher body mass index and were more likely to be female and reside in urban areas. Participants in the high-level increasing BRI group also had lower proportions of current smoking and alcohol consumption. In addition, the baseline prevalence of dyslipidemia and heart disease was higher in the high-level increasing trajectory group. No significant differences were observed across trajectory groups with respect to marital status, education level, stroke, lung disease, cancer, or kidney disease.

Figure 2

In the ELSA validation cohort, three BRI trajectories comparable to those identified in CHARLS were observed: low-level inverted U-shaped (trajectory 1; n = 486), moderate-level inverted U-shaped (trajectory 2; n = 513), and high-level increasing (trajectory 3; n = 106) patterns, with satisfactory model fit (Figure S1A). The corresponding incidences of hypertension were 17%, 19%, and 27%, respectively (Figure S1B), supporting the general consistency of the findings across cohorts.

Associations between BRI trajectories and hypertension incidence

In the CHARLS cohort, higher BRI trajectories were associated with increased odds of incident hypertension, and the associations remained after sequential multivariable adjustment (Table 3). In the fully adjusted model, compared with the low-level inverted U-shaped trajectory, the ORs were 1.40 (95% CI, 1.13–1.73) for the moderate-level inverted U-shaped trajectory and 1.76 (95% CI, 1.21–2.56) for the high-level increasing trajectory, with a significant trend across trajectory groups (P for trend <0.001).

In the ELSA cohort, the high-level increasing trajectory remained associated with increased odds of incident hypertension in the fully adjusted model (OR, 1.67; 95% CI, 1.01–2.75), whereas the association for the moderate-level inverted U-shaped trajectory was not statistically significant (Table 4). Although the trend across trajectory groups was attenuated after multivariable adjustment, the direction of the estimates was generally consistent with that observed in CHARLS. These results should be interpreted cautiously because the high-level trajectory group included only 106 participants.

Subgroup analysis

In subgroup analyses stratified by age, sex, education, smoking status, alcohol consumption, and major comorbidities, the positive association between higher BRI trajectories and incident hypertension remained generally consistent in both cohorts.

In the CHARLS cohort (Figure S2), participants in the high-level increasing BRI trajectory consistently exhibited significantly elevated risks of incident hypertension across nearly all subgroups, with odds ratios ranging from approximately 1.88 to 2.26 when compared with the low-level inverted U-shaped trajectory. A clear dose–response pattern across increasing BRI trajectories was observed in most strata (all P for trend <0.01). The magnitude of association appeared slightly stronger among older individuals (>60 years) and those without major comorbidities, although formal interaction tests did not indicate significant effect modification (all P for interaction >0.05).

In the ELSA cohort (Figure S3), the overall direction of associations was consistent with that observed in CHARLS. Participants in the high-level increasing BRI trajectory generally showed higher risks of hypertension across most subgroups; however, statistical significance was not uniformly achieved in certain strata, particularly among the oldest age group and participants with cancer, likely reflecting limited sample size and reduced statistical power.

Importantly, no statistically significant interactions were detected between BRI trajectories and any subgroup variables in either cohort, indicating that the association between long-term BRI patterns and hypertension risk was broadly stable across demographic and clinical subgroups.

Machine learning analysis

Feature screening

Feature selection was conducted using a two-step pipeline comprising multicollinearity assessment and Light Gradient Boosting Machine (Light-GBM)-based AUC-driven selection to optimize prediction of incident hypertension. First, correlation analyses identified highly collinear variable pairs (trajectory group and body mass index, smoking status and sex, remnant cholesterol and triglycerides; r > 0.64), from which one variable in each pair was retained. Consequently, 12 candidate variables, C-reactive protein (CRP), remnant cholesterol, age smoking and drinking status, trajectory group, residence, dyslipidemia, marital status, education level, kidney disease, and lung disease, were included for further analysis (Figure S4A), and their relative importance was ranked using Light-GBM (Figure S4B). Second, sequential forward selection based on incremental AUC improvement was applied to identify the most informative feature subset. A final panel of six features (CRP, remnant cholesterol (RC), age, smoke, alcohol consumption, and trajectory) was selected for model construction, with CRP, remnant cholesterol, and age emerging as the strongest predictors (AUC > 0.80). Inclusion of additional features further improved model performance, with the AUC approaching 0.88 (Figure 3A).

Figure 3

SHAP-based model interpretation and feature contribution patterns

Model interpretability was further assessed using SHAP. SHAP summary plots indicated that CRP and RC had the strongest overall influence on model predictions, with higher feature values generally contributing to an increased predicted risk of incident hypertension (Figure 4A). In addition, the proportional contribution analysis showed that CRP, RC, and age accounted for the largest shares of the total contribution among the six selected predictors, contributing 56.4%, 18.4%, and 10.7%, respectively, to the incident hypertension prediction model. In contrast, the BRI trajectory variable contributed the smallest proportion (3.84%) (Figure 4B). Stratified SHAP analyses demonstrated differential patterns of feature contributions across BRI trajectory and age subgroups. Across trajectory groups, the combined contribution of CRP and RC to the prediction model was higher in trajectory 1 (37.1%) and trajectory 2 (35.6%) than in trajectory 3 (27.3%). Across age strata, the relative contributions of CRP and RC were broadly comparable among participants aged 45–50 years (30.8%), 50–60 years (35.0%), and >60 years (34.2%). In contrast, the relative contributions of smoking and alcohol consumption increased with age, being higher in participants aged >60 years (39.7%) and 50–60 years (34.5%) compared with those aged 45–50 years (25.8%) (Figure 4C). SHAP dependence plots, complemented by LOWESS analyses, further suggested non-linear associations between key predictors and SHAP values of the corresponding predictors (Figure 4D).

Figure 4

Sensitivity analyses

A series of sensitivity analyses were conducted to assess the robustness of the main findings (Table S5–S9). First, analyses restricted to participants without missing covariate values yielded results consistent with the primary analysis (Table S5). Second, when baseline BRI was analyzed instead of BRI trajectories, higher baseline BRI remained associated with incident hypertension. In the fully adjusted model, the OR was 1.24 (95% CI, 1.14–1.35) per 1-SD increase in baseline BRI, and participants in the highest tertile had an OR of 1.69 (95% CI, 1.36–2.10) compared with those in the lowest tertile (Table S6). To minimize potential confounding by medication use, additional analyses were performed after excluding participants who used lipid-lowering drugs, with results remaining largely unchanged (Table S7). Furthermore, exclusion of participants with cancer did not materially alter the observed associations (Table S8). We also used the inverse probability weighting method to adjust for measured confounders. Standardized mean differences (SMD) for the IPW-weighted covariates were calculated, and the results showed that SMD for all covariates was less than 10% after IPW adjustment (Figure S5). Weighted logistic regression analyses accounting for the complex survey design confirmed the positive association between BRI trajectories and hypertension risk (Table S9). In sensitivity analyses excluding potential metabolic and clinical intermediates, the associations between higher BRI trajectory groups and incident hypertension remained materially unchanged (Table S10). Finally, Cox proportional hazards regression yielded results consistent with those of the logistic regression models. Compared with trajectory 1, trajectory 2 and trajectory 3 were associated with higher risks of incident hypertension, with hazard ratios of 1.46 (95% CI, 1.23–1.72) and 1.92 (95% CI, 1.48–2.50), respectively. A significant dose–response trend was observed across BRI trajectory groups (P for trend = 0.01), supporting the robustness of the primary findings (Table S11).

Strengths

Several strengths of this study merit consideration. First, the use of longitudinal BRI trajectories allowed for characterization of long-term patterns of central adiposity, providing a more comprehensive assessment of cumulative exposure than single-point anthropometric measurements. Second, the restriction to a non-diabetic population minimized potential confounding from hyperglycemia, insulin resistance, and diabetes-related treatments, thereby clarifying the independent relationship between central adiposity trajectories and incident hypertension. Third, the consistency of findings across two independent cohorts, together with extensive sensitivity and subgroup analyses, supports the robustness and external validity of the results. Finally, the integration of traditional regression models with explainable ML approaches enhanced the interpretability and credibility of the observed associations.

Limitations

Several limitations should be acknowledged. First, BRI was derived from anthropometric measurements rather than direct assessment of visceral adiposity, which may introduce measurement error. Second, residual confounding cannot be excluded, and the observational design precludes causal inference. Third, probabilistic trajectory assignment and the limited number of assessment waves may have introduced classification uncertainty and restricted the identification of more complex longitudinal patterns. Fourth, the machine learning analysis was exploratory and requires further external and prospective validation before clinical application. Finally, the findings may not be generalizable to younger populations or other ethnic groups.