Machine learning identification of influencing factors of global Nation-Level hypertension prevalence

Abstract

Hypertension remains a critical global public health challenge, with its complex etiology poorly captured by traditional linear models, especially regarding macro-level structural and gender-specific drivers. To address this, we employed an interpretable machine learning framework, combining XGBoost with SHAP and bootstrap resampling. This approach analyzed a global, nation-level panel dataset from 190 countries (1990–2019) across four macro dimensions: natural geography, behavior, socioeconomic status, and healthcare. The XGBoost model achieved excellent predictive performance. SHAP analysis identified mean annual precipitation (PRC), hospital beds (HOS), obesity prevalence (OB), and access to safely managed drinking water as the dominant macro-determinants of global hypertension prevalence. Critically, factor influence showed profound gender heterogeneity: HOS was the most impactful predictor for males, while OB and specific socioeconomic indicators were key drivers for females. Our findings offer a transparent and actionable perspective for policymakers, underscoring the necessity of integrating macro-environmental and gender-stratified insights to formulate precise and equitable public health interventions and resource allocation strategies globally.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12889-025-25335-y.

Introduction

Hypertension, a central risk factor for noncommunicable diseases (NCDs) globally, remains largely uncontrolled worldwide and continues to be the leading cause of NCD-related deaths, posing a significant public health threat [1]. At present, its effect extends to over 1 billion people worldwide [2, 3]. In China, hypertension prevalence has steadily increased over the past two to three decades due to extended life expectancy and lifestyle changes [4]. In the United States, approximately 116 million adults are affected, accounting for about 45% of the adult population [5, 6]. Notably, the burden of hypertension is more severe in low- and middle-income countries (LMICs). An estimated 31.1% of adults globally (1.39 billion) had hypertension in 2010, with a significantly higher prevalence in LMICs (31.5%, 1.04 billion people) compared with high-income countries (28.5%, 0.349 billion people) [7]. More than 10 million deaths annually worldwide are attributed to hypertension, with LMICs bearing the greatest burden, characterized by poorly controlled blood pressure and increased cardiovascular disease (CVD) mortality [8]. For instance, systolic blood pressure exceeding 115 mmHg was estimated to have contributed to 8.5 million deaths in 2015, with 88% of these fatalities occurring in LMICs [9]. Notwithstanding the advancements in high-income nations, inequalities continue to prevail, particularly in the United States, where CVD mortality rates among Black individuals are four times higher than among White individuals [8]. Hypertension is not only a leading cause of CVDs (e.g., myocardial infarction, heart failure, coronary heart disease, and stroke) and premature death globally [2, 6, 10, 11], but it also contributes to chronic kidney disease and cognitive decline; it even increases the risk of vascular cognitive impairment and Alzheimer’s disease [12–14], highlighting its multifaceted negative effects on health.

The etiology of hypertension has been extensively discussed in the academic literature, encompassing several factors, such as deficiency in or excessive intake of trace elements [3, 15, 16], obesity [7, 17, 18], alcohol consumption [3, 5, 18], smoking [19–21], lack of physical activity [22–24], living environment [8, 25], inflammation [26–28], healthcare conditions [9, 29, 30], environmental pollution [25, 31, 32], and patient adherence to treatment [1, 33, 34]. Although these studies are crucial for understanding the causes of hypertension, they frequently overlook the influences of macro-level factors, such as natural geography and socioeconomic development, leading to an incomplete analysis of influencing factors. Furthermore, existing research often lacks a comparative analysis of gender differences. Gender also affects hypertension prevalence. Although hypertension is generally more common in males, female hypertension prevalence accelerates with age. Moreover, Females exhibit a higher predisposition to cardiovascular incidents at comparatively lower blood pressure levels, potentially due to the persistent effects of hypertensive disorders during pregnancy and interactions between sex hormones and other factors [35]. Hypertension prevalence is significantly influenced by the interactive effects of multidimensional and complex factors. Traditional statistical models (e.g., linear regression) struggle to effectively capture these nonlinear relationships.

Given the complexity of the global prevalence of hypertension and the urgency of intervention, this study aims to establish a more precise and interpretable analytical framework. Traditional research typically over-relies on linear regression models or narrowly focuses on individual clinical and behavioral risk factors within limited geographical regions. In sharp contrast, this study adopts an innovative strategy: it utilizes a nonlinear XGBoost regression model to process macro panel data from 190 countries between 1990 and 2019, combined with the SHAP interpretability framework, thereby establishing a more precise and transparent analytical paradigm.

Specifically, this research comprehensively analyzes the influencing factors of hypertension prevalence across four dimensions: natural geographical factors, personal behavioral habits, socio-economic conditions, and health care conditions. The core innovation lies in highlighting structural and environmental drivers at the macro level, particularly the global role of natural geographical and socio-economic factors. By integrating the SHAP framework, we aim to deliver transparent and highly policy-relevant insights, revealing nonlinear marginal effects and complex interactions that are unattainable with traditional models. Crucially, this study also clearly analyzes the differentiated impact of these factors across genders, exposing the unique sensitivities of men and women in risk exposure. Ultimately, the results of this study will provide new and precise evidence to inform global hypertension intervention strategies, leveraging the full interpretability of nonlinear modeling.

Data and methods

This study constructs a machine learning-driven framework for global hypertension risk factor analysis (Fig. 1). The framework consists of the following: (1) integration and preprocessing of multisource data from the World Health Organization (WHO) and the World Bank; (2) spatial analysis of global distribution, temporal trends, and regional heterogeneity of hypertension prevalence; (3) gender-stratified modeling using XGBoost (combined with Bayesian optimization, fivefold cross-validation); (4) SHAP-based interpretation of feature importance and nonlinear relationships; and (5) bootstrap resampling to validate result robustness.

Fig. 1.

Research Flowchart of Global Hypertension Risk Factor Analysis

Data collection and preprocessing

Hypertension prevalence data in this study were sourced from the age-standardized hypertension prevalence among adults aged 30–79 years, published by WHO. A total of 18 key independent variables were included, as summarized in Table 1, encompassing 4 major domains: natural geographic elements, personal behavioral patterns, socioeconomic status, and healthcare conditions, providing a multidimensional perspective for a comprehensive assessment of hypertension prevalence. All data sources are authoritative. WHO, as a specialized agency of the United Nations, collects disease and health service data through standardized survey methods (e.g., stepwise approach to surveillance) via its Global Health Observatory database (https://data.who.int/indicators). Indicators, such as the Universal Health Coverage index (UHC), obesity prevalence (OB), and physician density, in the current study are derived from this source. The Human Development Index (HDI), published by the United Nations Development Programme (UNDP) through its Human Development Report, is available on the Our World in Data platform (https://ourworldindata.org). This index reflects national development levels by integrating dimensions, such as life expectancy, education, and quality of life. The World Bank, an international economic organization, provides public datasets (https://data.worldbank.org) that cover global environmental, economic, and social indicators for countries worldwide. A total of 14 indicators, including fine particulate matter (PM_2.5) and urbanization level (URB), used in the current study are sourced from the World Bank, with data validated collaboratively by national statistical bureaus and international agencies. Furthermore, indicators with gender disparities were classified and statistically analyzed by total, male, and female categories to ensure the completeness of the gender dimension in the analysis.

Table 1.

Summary of data sources utilized in the study

Primary Category	Subcategory	Detail	Source
Hypertension Prevalence	Rates	Hypertension in adults (30–79)	https://data.who.int/indicators
Natural Geographic Elements	PM2.5	PM2.5 air pollution, mean annual exposure (micrograms per cubic meter)	https://data.worldbank.org
	VEG	Forest area (% of land area)	https://data.worldbank.org
	PRC	Average precipitation depth (mm/year)	https://data.worldbank.org
Personal Behavioral Patterns	ALC	Total alcohol consumption per capita (15+)	https://data.worldbank.org
	MAL	Prevalence of malnutrition (% of population)	https://data.worldbank.org
	OB	Obesity in adults (18+)	https://data.who.int/indicators
	SMK	Prevalence of current tobacco use (% of adults)	https://data.worldbank.org
Socioeconomic Status	PGDP	GDP per capita (current USD)	https://data.worldbank.org
	SLM	Population living in slums (% of urban population)	https://data.worldbank.org
	GIN	GINI Coefficient	https://data.worldbank.org
	URB	Urban Population Percent	https://data.worldbank.org
	HDI	Human Development Index	https://ourworldindata.org
Healthcare Conditions	PHY	Density of doctors	https://data.who.int/indicators
	UHC	UHC Service coverage index	https://data.who.int/indicators
	SAN	People using safely managed sanitation services (% of population)	https://data.worldbank.org
	WAT	People using safely managed drinking water services (% of population)	https://data.worldbank.org
	HOS	Hospital beds (per 1,000 population)	https://data.worldbank.org
	GHE	Domestic general government health expenditure per capita (current USD)	https://data.worldbank.org

The study data cover 192 countries worldwide from 1990 to 2019, with an initial sample size that exceeds 17,000 records. In the data preprocessing stage, records with completely missing key indicators (Nauru and Slovakia) were first removed, followed by the exclusion of individual samples with more than 30% missing values. For the remaining missing values, median imputation and XGBoost prediction strategies were employed and compared. The resulting R² values obtained after applying each imputation method showed a difference of less than 0.01, indicating a highly consistent model fit regardless of the missing value imputation approach used. Although the feature importance ranking of some variables exhibited slight variations, the overall trend remained stable, suggesting strong robustness of the study results. Logarithmic transformations were applied to mitigate the effect of skewed distributions, such as PM_2.5 and gross domestic product per capita (PGDP). Finally, all continuous features were dimensionless-processed using Z-score standardization to eliminate dimensional differences. After the described data processing, 190 countries were ultimately included in the analysis, excluding Palestine, Vatican City, San Marino, Monaco, Liechtenstein, Slovakia, and Nauru.

The dataset of this study is clearly organized into a country-year panel, with each row of records representing the data of a specific country in a certain year. This panel contains 18 explanatory variables, covering four major aspects: physical geography, behavioral patterns, socioeconomic status, and healthcare conditions. To achieve gender-stratified analysis, the prevalence data of hypertension were obtained respectively for the total population, male and female, thus forming three independent datasets. The valid sample size for each dataset is 2,964 records, and the total valid sample size in total is 8,892 records. These three datasets were subsequently analyzed independently using the same methodology, ensuring a fair and rigorous comparison of the predictors in the overall, male, and female populations.

Research methods

Linear correlation analysis

Given the non-normal distribution and discrete nature of different data, the current study employed Spearman correlation coefficients to analyze variable associations. To enhance the explanatory power of the XGBoost model, a collinearity test of independent variables was conducted using the variance inflation factor (VIF). When VIF < 10 and tolerance > 0.1, variables are considered to exhibit no significant collinearity, and this step aims to eliminate the interference of multicollinearity on the model’s explanatory power [36].

XGBoost-based model for analyzing the influencing factors of hypertension prevalence

XGBoost is an ensemble learning model based on gradient boosting. It is widely applied in regression and classification tasks due to its efficient computational performance and powerful predictive capabilities. Compared with other traditional methods, XGBoost offers significant advantages in training speed for large-scale data or parallel computing scenarios. It also exhibits built-in feature selection capabilities, robustness to nonlinear relationships, and good interpretability [37].

In this study, we utilize the XGBoost model to analyze the effects of environmental, behavioral, socioeconomic, and other factors on hypertension prevalence in different gender populations. During model training, data were divided into three gender-stratified groups: total, female, and male. Separate models were built for each group to capture gender heterogeneity, e.g., a factor that exerted a stronger influence on male hypertension prevalence but a weaker influence on female prevalence. Each group’s data were divided into training and testing sets at a ratio of 8:2. To prevent model overfitting and improve generalization ability, fivefold cross-validation, combined with Bayesian optimization, was used to tune model hyperparameters, including max_depth, learning_rate, subsample, min_child_weight, and regularization coefficients.

Its simplified mathematical principle is as follows:

Assuming a dataset that contains n samples and m features, the XGBoost model iteratively builds k regression trees through an additive model, wherein each iteration adds a new tree to minimize the objective function, as follows [37]:

1

2

3

where T is the number of leaf nodes, W_j​ is the score of the j-th leaf node, and and are regularization hyperparameters. Through this optimization framework, XGBoost can effectively identify the major drivers of hypertension prevalence while balancing accuracy and interpretability.

SHAP interpretability framework for hypertension influencing factors

SHAP is a game theory-based model interpretation method. Its core idea is to decompose the contribution of each feature to the model’s prediction into additive values, quantifying the degree of influence of each feature on the prediction result. This study adopts the TreeSHAP algorithm, which efficiently calculates Shapley values based on tree path decomposition, combining the marginal contribution principle of game theory with the structural characteristics of tree models [38].

The specific formula is as follows:

4

represents the SHAP value of the -th feature on sample , and this value is used to quantify the contribution of this feature to the model’s prediction result. is the feature index, is the input sample, and is the value of the -th feature in the sample. and represent the leaf node index and total number of leaf nodes in the tree model, respectively. is the number of samples falling into the -th leaf node, and ​ is the set of samples in that leaf node. denotes the operation of replacing feature values. determines whether the feature participates in splitting through the tree’s path conditions. The entire formula efficiently calculates the marginal effect of features in the tree model by averaging the contributions of samples within all leaf nodes, combining tree structure decomposition with the Shapley value idea from game theory. This formula is capable of capturing nonlinear interactions among features and quantifying their actual contribution to predictions, making it suitable for tree models, such as XGBoost.

Bootstrap resampling for the stability testing of hypertension influencing factors

To rigorously validate model robustness, the current study introduces bootstrap resampling for analysis. As a classic self-sampling statistical inference method, its core idea is to generate multiple bootstrap samples by sampling with the replacement from the original sample, and thus, estimating the distribution, bias, and variance of statistics [39]. The present study performs sampling with replacement on the original sample to generate bootstrap samples with the same size as the original. For each bootstrap sample, an XGBoost model is trained, and feature importance is recorded. This process is repeated 1000 times to accumulate the importance distribution. Based on this condition, 95% confidence intervals are calculated (via the 2.5% and 97.5% quantiles), and the significance of features is determined by whether the interval includes zero [40].

Results

Global hypertension prevalence patterns and evolution

Figure 2 illustrates the global distribution of hypertension prevalence in 1990 and 2019. In 1990, Russia and Eastern European and Central Asian countries were high-prevalence areas, while Western Europe and North America mostly had medium-to-low prevalence. By 2019, prevalence decreased in North America and parts of Eastern Europe, while it remained high in South America and parts of Africa, indicating a shift in the high-prevalence areas. Over the past 30 years, global hypertension prevalence has exhibited a spatial heterogeneity of “low in Europe and America, high in parts of Asia, Africa, and Latin America” and an overall increasing trend.

Fig. 2.

Comparisons between the 1990 and 2019 Global Hypertension Prevalence Distributions

Figure 3 illustrates the spatial distribution pattern of global hypertension prevalence growth rates from 1990 to 2019, revealing significant spatial heterogeneity. In Asia (China), Southeast Asia (e.g., Thailand, Myanmar, and Indonesia), parts of Africa, and some South American countries, hypertension prevalence growth rates are in the high range, reflecting a significant upward trend in prevalence over the past 30 years in these regions. By contrast, in North America and parts of Europe, the growth rates are negative or in a low positive growth range, indicating a declining or slowly increasing trend in prevalence.

Fig. 3.

Spatial Variation of Global Hypertension Prevalence Growth Rates

Spearman correlation coefficient analysis and XGBoost performance

Table 2 provides the Spearman correlation coefficients between each feature variable and hypertension prevalence, and the VIF and tolerance results from the multicollinearity test. The tolerance values for all the variables were greater than 0.1, and all VIF values were less than 10, indicating no severe multicollinearity among the variables and making them suitable for subsequent modeling analysis. In terms of correlation, PM_2.5, total alcohol consumption per capita (ALC), prevalence of current tobacco use (SMK), OB, URB, HDI, and hospital beds (HOS) exhibited significant positive correlations with prevalence (p < 0.01). Mean annual precipitation (PRC), prevalence of malnutrition (MAL), GINI Coefficient (GIN), Population living in slums (SLM), and vegetation coverage (VEG) presented significant negative correlations (VEG with p < 0.05, others with p < 0.01). Furthermore, the correlations for PGDP, government health expenditure per capita (GHE), safely managed drinking water (WAT), people using safely managed sanitation services (SAN), and UHC were not statistically significant, indicating no clear linear association with hypertension prevalence.

Table 2.

Outcome of the multicollinearity evaluation and correlation coefficients of the determinant factors

Features	Correlation coefficient		VIF	Tolerance
PM2.5	0.037**	2.251		0.444
PRC	−0.254**	2.191		0.456
VEG	−0.027*	1.941		0.515
ALC	0.073**	2.182		0.458
SMK	0.038**	1.094		0.914
MAL	−0.066**	2.228		0.449
OB	0.253**	1.65		0.606
GIN	−0.060**	1.462		0.684
URB	0.066**	2.785		0.359
HDI	0.033**	7.059		0.142
SLM	−0.111**	1.287		0.777
PGDP	−0.017	8.787		0.114
SAN	−0.02	2.845		0.351
GHE	−0.003	3.072		0.326
WAT	0.003	2.21		0.453
HOS	0.254**	1.987		0.503
UHC	0.018	1.384		0.722
PHY	0.126**	1.815		0.551
Note: The significance levels for correlation coefficients are denoted by * and **, representing p-values of 0.05 and 0.01, respectively.

To explore the gender heterogeneity of hypertension prevalence, the current study constructed XGBoost models for different groups and performed Bayesian optimization. All three models demonstrated high goodness-of-fit and stability, verifying the effectiveness of XGBoost in capturing the complex mechanisms of hypertension. The model parameters are provided in Table 3.

Table 3.

XGBoost model parameters for different groups

Hyperparameter	Search Range	Total	Female	Male
Learning Rate	0.01–0.2	0.142	0.2	0.2
Max Depth	3–6	6	5	6
Number of Trees	50–300	182	300	300
Subsample	0.7–0.9	0.7	0.9	0.9
Colsample Bytree	0.7–0.9	0.7	0.9	0.9
Min Child Weight	3–10	3	10	10
Gamma	0.1–2.1	0.1	0.1	0.1

The model evaluation results are presented in Table 4. All models demonstrated outstanding fitting capabilities on the training set (with R² values all above 0.99, with the highest reaching 0.9988), which confirmed the powerful ability of XGBoost in capturing the complex nonlinear relationships of macro factors. The R² values of the test sets for all groups remained at a high level of 0.94 to 0.96, and the difference (generalization loss) between the R² values of the training and test sets was only controlled within 0.03 to 0.05. This extremely small difference strongly proved that the models successfully avoided severe overfitting. In terms of prediction accuracy, all models performed excellently, with the root mean square error (RMSE) of the test sets ranging from 1.340 to 1.484 and the mean absolute error (MAE) ranging from 0.855 to 0.870, confirming the high-accuracy predictive ability of the models for the prevalence of hypertension.

Table 4.

XGBoost model parameters for different groups

Group	Test-R²	Train-R²	Test-RMSE	Train-RMSE	Test-MAE	Train-MAE
Total	0.9478	0.9963	1.484	0.383	0.855	0.261
Female	0.9627	0.9962	1.340	0.413	0.870	0.309
Male	0.9634	0.9988	1.389	0.246	0.866	0.176

Feature importance and distribution visualization of hypertension influencing factors

The analysis of feature importance based on SHAP values (Fig. 4) reveals significant differences across groups. Overall, PRC (1.38) is the core influencing feature, with HOS (1.05), OB (0.98), and others acting as synergistic drivers. In the female group, OB (1.38) dominates the model’s influence, with PRC (1.08), PGDP (1.02), and GHE (0.81) playing prominent roles. In the male group, HOS (1.73) exerts the greatest influence, followed by significant contributions from PRC (1.48), OB (0.98), VEG (0.8), and ALC (0.7).

Fig. 4.

SHAP Feature Importance Bar Charts for the Total, Female, and Male Groups

Figure 5 illustrates the importance and directional impact of each variable on hypertension prevalence across distinct population groups. The horizontal axis denotes SHAP values, quantifying whether a variable exerts a positive or negative influence on the model output; color gradients from blue to red correspond to low-to-high feature values. The findings reveal that obesity prevalence (OB) and annual mean precipitation (PRC) are the dominant drivers across all three groups. Specifically, higher OB values markedly elevate predicted hypertension risk, with the female cohort exhibiting an even stronger positive effect than the others. Conversely, rising PRC values are associated with a pronounced negative influence on hypertension prevalence. Additionally, PM_2.5, ALC, VEG, and WAT display heterogeneous effects across groups, underscoring that hypertension is shaped by a complex interplay of socio-environmental and health behavioral factors.

Fig. 5.

SHAP Feature Distribution Violin Plots across the Total, Female, and Male Groups

To analyze the patterns of feature influence on model output and group differences, the current study generated SHAP dependence plots (Fig. 6) for the top four ranked features, illustrating the SHAP dependence relationships of key variables on the hypertension prevalence prediction model. In these plots, the gray histogram provides the distribution of feature values, and the colored scatter points reflect the marginal contribution to the model output and the interactive effects of variables. The horizontal axis of each subplot represents the original feature value, the vertical axis represents the SHAP value of that feature, and the color of the points represents the value of another variable that interacts with the feature. The total group (Fig. 6A) focuses on PRC, HOS, OB, and WAT, demonstrating the synergistic effect of multiple factors, including climate, medical care, and lifestyle, with the interactive threat of water quality and air quality being particularly prominent. The female group (Fig. 6B) exhibits a significant buffering effect of economic level and government investment on the risk of hypertension in women. The male group (Fig. 6C) reflects a strong correlation between “medical resources and environmental exposure,” and the interactive influence of vegetation cover and pollution is more specific.

Fig. 6.

SHAP Dependency Plots for Total(A), Female(B), and Male Groups (C)

Based on the XGBoost model combined with bootstrap resampling, we evaluated the influence of various features on the total population and different gender groups, and quantified feature importance with 95% confidence intervals (Fig. 7). The results further corroborate the previous SHAP-based findings. PRC, HOS, and WAT presented the highest feature importance in the total population, and their confidence intervals did not include 0, indicating a significant and robust influence. In particular, PRC’s importance was particularly prominent in the female group, with a high mean importance and a relatively wide confidence interval, suggesting potentially large heterogeneity. Meanwhile, HOS showed overwhelming importance in the male group, with its mean importance significantly higher than those of other features and a narrower confidence interval, indicating its highly stable core role in male hypertension prevalence. Furthermore, PGDP and GHE exhibited higher importance in the female group, while OB and ALC showed relatively increased importance in the male group.

Fig. 7.

Feature Importance with 95% CI for Total, Female, and Male Groups

Discussion

The spatial heterogeneity of global hypertension prevalence, characterized by “low in Europe and America, high in parts of Asia, Africa, and Latin America,” along with an overall increasing trend, arises from significant differences in population structure and socioeconomic development. From a demographic perspective, global population growth coupled with aging directly expands the base number of people with hypertension. At the socioeconomic level, low- and middle-income regions (such as parts of Asia, Africa, and Latin America) are constrained by the lack of medical resources and weak health awareness, leading to persistently low rates of hypertension awareness and treatment. Furthermore, the prevalence of unhealthy lifestyles, such as high-salt and high-fat diets and reduced physical activity, with accelerating urbanization further drives the increase in prevalence. By contrast, high-income regions in Europe and America, which benefit from well-established hypertension prevention and control systems, have effectively curbed or even reversed the growth trend of prevalence through widespread health screenings, standardized treatment protocols, and continuous health lifestyle interventions (e.g., salt control and exercise advocacy). This disparity in health management, which is influenced by regional development imbalances, collectively shapes the current spatially heterogeneous distribution pattern of global hypertension prevalence.

The current study integrates the XGBoost model, SHAP interpretability tool, and bootstrap resampling techniques to conduct a large-scale systematic analysis of nearly 30 years of data from 190 countries, spanning 4 dimensions: natural geography, personal behavior, socioeconomic status, and healthcare conditions. The results show that PRC, HOS, and WAT are three universally important factors that influence global hypertension prevalence. The importance of PRC may reflect the effects of climate and related agricultural activities on people’s lifestyles and health. The importance of HOS highlights the critical role of medical service accessibility and quality in chronic disease management; sufficient hospital beds likely signify good conditions for diagnosis, treatment, and management [41–43]. WAT emphasizes the importance of infrastructure development and public health services in disease prevention. The widespread availability of clean drinking water helps reduce waterborne diseases and improve overall health, indirectly lowering hypertension risk. These findings are consistent with the existing literature on the effects of environmental and socioeconomic factors on health [23, 44, 45]; however, the current study provides precise insights by quantifying their contributions.

Notably, the present study delves into and reveals significant gender heterogeneity in influencing factors. In the female group, OB emerged as the most significant influencing factor, followed by PRC, PGDP, and GHE. This result suggests that intervention strategies for female hypertension should focus on obesity management, improving socioeconomic development levels, and government investment in health. A clear physiological link exists between obesity and hypertension [46, 47], and women’s unique physiological cycles and hormonal changes may make them more susceptible to obesity-related metabolic disorders [48, 49]. The importance of PGDP and GHE may reflect the profound effects of economic development level and healthcare investment on women’s health and well-being [50–52], particularly in improving medical accessibility and quality of life.

By contrast, the influence of HOS considerably outweighs those of other factors in the male group, becoming the decisive driving factor, followed by PRC, OB, and VEG. This finding may imply that men are more dependent on medical services for hypertension management, or that their health behavior patterns (e.g., smoking, alcohol consumption) have more complex interactions with medical resource accessibility. The importance of VEG may be related to the effects of outdoor activities, mental health, and environmental quality on male health [53–55]. Furthermore, the importance of ALC is relatively high in the male group, and this result is consistent with the general understanding of alcohol intake as a risk factor for hypertension [56–58]. These findings reveal significant gender-based differences in hypertension etiology at a macro scale, providing cross-scale evidence to support the formulation of refined intervention strategies.

This study utilizes SHAP values to quantify the importance of each feature variable and reveals the direction and extent of the impact of different feature values on hypertension prevalence through violin plots and partial dependence plots. For instance, an increase in OB significantly elevates the risk of hypertension in women, whereas a rise in PRC reduces the risk of hypertension across all groups. Through these visualizations, we can gain a deep understanding of the impact patterns, nonlinear relationships, and interactive effects of various influencing factors on hypertension prevalence in different groups. These findings provide data support for future precise interventions and emphasize the necessity of considering gender differences when formulating public health policies. Specifically, priority should be given to investing in basic environmental sanitation facilities such as safe drinking water, optimizing the accessibility of medical services for men, and designing targeted metabolic health and socio-economic support plans for women, so as to achieve more cost-effective and equitable global hypertension control targets.

However, although this study has yielded the aforementioned findings, certain limitations persist. This study focuses on macro-level and annual data, without incorporating microscopic clinical data, individual behavioral details, genetic backgrounds, or potential dynamic interactions, all of which play important roles in the development of hypertension. Future research may consider integrating detailed individual-level data and adopting time-series analysis methods to uncover the dynamic changes and deeper causal mechanisms of hypertension-influencing factors.

Conclusion

This study utilized XGBoost, SHAP, and bootstrap resampling techniques to comprehensively analyze the influencing factors of global hypertension prevalence. For the first time, it quantified the differential effects of these factors across the total population, female, and male groups. The study found that PRC and WAT are universally important factors that influence global hypertension prevalence, indicating that natural geographical background conditions and socioeconomic factors are crucial perspectives for explaining differences in hypertension incidence among countries. OB, PGDP, and GHE exerted more prominent effects on female hypertension, while HOS, VEG, and ALC exhibited stronger associations in the male group. These findings not only provide new insights for understanding the complex etiology of hypertension but also emphasize the necessity of fully considering gender-specific strategies in public health interventions and policy formulation to achieve precise and effective global hypertension control goals.

Authors’ contributions

HL. Zheng：Writing – original draft，Writing – review and editing，Visualization，Software，Data curation，MethodologyHS. Yu* ：Writing – original draft，Writing – review and editing，Data curation，Formal analysis，ConceptualizationM.W. Rosenberg：Writing – review and editing，Formal analysis，MethodologyYM. Wu：Writing – review and editing，Project administration，Funding acquisitionY.L：Writing – review and editing，Project administrationJ.C：Writing – review and editing，Software，ValidationM.Y：Visualization，Data collection，Writing – review and editingXl. Y：Validation，Data curation，Writing – review and editingY. W*：Conceptualization，Writing – original draft，Writing – review and editing， Supervision，Funding acquisition， Project administration.

Funding

This work was financially supported by the “Yunnan Revitalization Talent Support Program” of Yunnan Province (Grant No. XDYC-QNRC-2022-0740; XDYC-WHMJ-2022-0016) and the Yunnan Province Innovation Team Project (Grant No. 202305AS350003).

Data availability

The relevant data of this study can be provided by the corresponding authors under reasonable circumstances.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

All authors have consented to the publication of the findings of this study.

Competing interests

The authors declare no competing interests.