To investigate the complex pathways of association by which VC in green spaces is linked with cardiovascular health, involving the intermediate roles of air quality and various PHRs, we employed four latent variables: VC, APM, PHR, and CHOs. Each latent variable encompassed multiple observed variables (Table 1). The bivariate correlations between all observed variables are presented in Supplementary Figure S2 in the Supplementary Material. For a comprehensive assessment of VC around patients’ residences, six vegetation indices were used as observed variables for VC. Given the known effects of VC on air quality, particularly APM (Diener and Mudu, 2021), PM2.5 and PM10 were selected as observed variables for APM. Based on PHRs leading to adverse cardiovascular outcomes (Hu, 2023), and to avoid potential temporal confounding from acute-phase reactants, we excluded inflammatory markers. Consequently, three second-level latent variables were chosen for PHR: blood lipid risk (BLR), blood glucose risk (BGR), and blood pressure risk (BPR). Corresponding biomarkers were selected as observed variables for these risks (Table 1). For CHOs, hospitalization utilization (HU) and cardiovascular disease diagnosis (CVDD) were chosen as latent variables. Data for these variables were collected from various sources, including satellite imagery for vegetation indices, air quality monitoring networks for APM, and electronic medical records for PHR and CHOs. For each patient, environmental exposure data (VC and PM) were calculated as annual averages for the calendar year preceding their hospital admission date. Appendix A provides the rationale for indicator selection, detailed data collection procedures, and data processing methods.

To explore the complex pathways of association linking VC to CHOs, we employed a pre-specified, four-stage analytical strategy using a series of path models. This hierarchical approach was designed to move progressively from broad, composite constructs to more specific, granular indicators, allowing for a systematic exploration of the associative pathways. VC was treated as the exogenous variable, CHOs as the final endogenous variables, with APM and PHR positioned as intermediate variables in the pathway. The observed variables for each latent variable are presented in Table 1. Each of the four model groups (Model Groups 1-4) was analyzed using partial least squares structural equation models (PLS-SEMs) (Figure 1).

The model evaluation results (Supplementary Table S1) indicate that the composite constructs for VC and APM have high reliability and validity. However, the overall PHR construct, now excluding inflammatory markers, still shows CR, AVE, and Cronbach’s α values below recommended thresholds, suggesting heterogeneity among the remaining physiological risks. The adjusted R² values for the final outcomes were 0.120 for HU (Model 1a) and 0.080 for CVDD (Model 1b), indicating modest explanatory power at this broad level.

The path analysis results for Model Group 1 (Figure 2) show that all path coefficients are statistically significant. VC was significantly and negatively associated with APM (β = -0.222, p< 0.001). In turn, APM was positively associated with the composite PHR (β = 0.132 in Model 1a; β = 0.131 in Model 1b, p< 0.001). Finally, PHR was significantly associated with both HU (β = 0.315, p< 0.001) and CVDD (β = 0.245, p< 0.001).

The analysis of indirect associations (Table 2) reveals that all pathways are statistically significant. For HU (Model 1a), the primary indirect pathway is VC→APM→HU (Specific Indirect Effect = -0.05352, p< 0.001). For CVDD (Model 1b), a notable inconsistent association was observed: while the total association was negative (Total Effect = -0.01866, p< 0.001), the specific indirect path through APM alone (VC→APM→CVDD) was positive (Specific Indirect Effect = 0.00333, p< 0.001). This statistical phenomenon, potentially indicative of a suppressor effect, suggests complex underlying relationships that warrant further disaggregation in subsequent models.

In Model Group 2, the PHR construct was disaggregated into BLR, BGR, and BPR. Model evaluation (Supplementary Table S2) showed that while VC and APM remained robust, the BLR construct had a low Cronbach’s α (0.463). This low internal consistency indicates that TC and HDL-C do not form a reliable single construct, methodologically necessitating the disaggregation of BLR into its constituent biomarkers in the subsequent analytical stage.

The path results (Figure 3) confirm a significant negative association between VC and APM (β ≈ -0.222, p< 0.001). APM was significantly and positively associated with both BLR (β ≈ 0.115, p< 0.001) and BGR (β ≈ 0.100, p< 0.001), but its association with BPR was not significant. In turn, BLR and BGR were significantly associated with all four CHO variables (TH, DH, CIHD, HF). BPR was only significantly associated with HF (β = -0.005, p< 0.01).

The analysis of indirect associations (Table 3) indicates that the most substantial pathways from VC to CHOs operate through APM to BLR and BGR. For hospitalization outcomes, the indirect pathways involving BLR and BGR were strongest, with BGR being slightly more prominent for DH (Specific Indirect Effect = -0.00266) than BLR (Specific Indirect Effect = -0.00260). For disease diagnoses, the pathways via BLR and BGR were also the most prominent, with BLR being stronger for CIHD (Specific Indirect Effect = -0.00150) and BGR being stronger for HF (Specific Indirect Effect = -0.00141). The role of BPR as an intermediate variable was negligible in most models, showing only a minor significant indirect association for HF.

Based on the findings from Model Group 2, the third stage of our analysis simultaneously disaggregated APM into PM2.5 and PM10, and the key PHR constructs into their constituent biomarkers (TC, HDL-C, and GLU). Model evaluation metrics remained robust (Supplementary Table S3).

The path analysis results (Figure 4) reveal a critical finding: while VC was significantly and negatively associated with both PM2.5 (β ≈ -0.214, p< 0.001) and PM10 (β ≈ -0.228, p< 0.001), only PM10 showed significant subsequent associations with the key biomarkers. Specifically, PM10 was positively associated with TC (β ≈ 0.064, p< 0.001) and GLU (β ≈ 0.100, p< 0.001), and negatively associated with HDL-C (β ≈ -0.119, p< 0.001). In contrast, the paths from PM2.5 to these biomarkers were not statistically significant.

Consequently, the analysis of indirect associations (Table 4) demonstrates that the significant sequential pathways from VC to all CHOs operate exclusively through PM10. For TH, the strongest indirect pathway involved PM10 and TC (Specific Indirect Effect = -0.00282, p< 0.01). For DH, CIHD, and HF, the pathways involving PM10 and GLU were consistently the most prominent (e.g., for DH: Specific Indirect Effect = -0.00273, p< 0.01). These results identify PM10 as the primary PM variable in this chain of association, linking VC to CHOs via specific metabolic biomarkers.

In the final stage, we compared the relative strength of six individual VC indicators within the most salient pathway identified in Model Group 3. Based on those results, PM10 was used as the APM indicator, while TC was selected as the key biomarker for the TH outcome, and GLU was selected for the DH, CIHD, and HF outcomes.

The comprehensive-effect model was found to be invalid due to severe multicollinearity (VIF > 5), as anticipated in our pre-specified analytical plan. Consequently, we proceeded with single-factor models. The results (Supplementary Figure S3 and Table 5) show that all six VC indicators were significantly and negatively associated with PM10. In the subsequent pathways to CHOs, LAI consistently emerged as the indicator with the strongest indirect association for disease diagnoses, particularly for CIHD (Specific Indirect Effect = -0.00164) and HF (Specific Indirect Effect = -0.00164). For hospitalization outcomes (TH and DH), SAVI and LAI showed the most prominent associations. Notably, FVC also demonstrated a strong association for TH (Specific Indirect Effect = -0.00172) and DH (Specific Indirect Effect = -0.00259), although slightly weaker than LAI. This suggests that vegetation quality and structure, as captured by LAI and FVC, are particularly important in this environmental health pathway.

The results from our four-stage analytical strategy consistently support an association between VC and CHOs, which is statistically explained by indirect pathways involving APM and PHR. This highlights the complex web of associations through which green spaces are linked to cardiovascular health.

A central finding, established in Model Group 3, is the dominant role of PM10. While VC was associated with reductions in both PM2.5 and PM10 concentrations, only the pathways involving PM10 were significantly linked to subsequent changes in metabolic biomarkers and CHOs. This differential impact could be attributed to PM10 being more easily captured by vegetation. Additionally, PM10’s deposition in the upper respiratory tract may more readily trigger systemic responses affecting lipid and glucose metabolism.

The analysis further specified the biological pathways involved. Model Group 3 demonstrated that lower PM10 concentrations were associated with reduced TC and GLU levels, and increased HDL-C levels. These biomarkers, in turn, were linked to better health outcomes. Notably, the indirect pathways showed specificity: the VC→PM10→TC pathway was most prominent for TH, whereas the VC→PM10→GLU pathway was strongest for DH, CIHD, and HF. This suggests that PM10 exposure may influence hospitalization frequency and disease severity through distinct metabolic routes. The significant role of HDL-C in the pathways for DH, CIHD, and HF, but not for TH, further underscores this complexity.

Finally, the single-factor models in Model Group 4 confirmed that while all vegetation indicators were associated with reduced PM10, LAI and FVC showed the most significant indirect associations with CHOs. This indicates that vegetation structure and density are key characteristics. LAI, reflecting the total leaf surface area, is directly related to the capacity for air purification and particle deposition. FVC represents the density of VC, which also contributes to overall air quality improvement.

These findings collectively underscore the multifaceted nature of urban greening’s health benefits. They suggest that urban planning strategies should prioritize vegetation types that effectively reduce PM10 concentrations. Based on this study’s results, increasing LAI and FVC could be effective strategies for improving cardiovascular health. It is also important to note the presence of inconsistent associations in some pathways, such as the VC→APM→CVDD path in Model 1b. This phenomenon, potentially a suppressor effect, suggests the existence of complex underlying relationships not fully captured by the current model and highlights a valuable direction for future research.

Our general finding that green space is associated with better cardiovascular health via pathways involving PM and metabolic biomarkers is consistent with a growing body of literature. For example, studies have similarly identified PM as a key intermediate variable in the link between greenness and atherosclerotic CVD risk (Shen and Lung, 2016; Dong et al., 2021), and have associated green space with improved lipid profiles (R. Lei et al., 2024; Xu et al., 2022, better glucose metabolism (Yang et al., 2019; Li et al., 2021), and reduced inflammation (Egorov et al., 2024; Lai et al., 2024). Our study builds upon this foundation by using a path analysis framework to explore these sequential relationships, aligning with findings that cardiometabolic disorders mediate the relationship between green space and CVD (Yang et al., 2020).

A notable finding of our study, confirmed robustly in our competitive model (Model Group 3), is the dominant role of PM10 over PM2.5 in the associative pathways. This diverges from several studies that have specifically implicated PM2.5 as the primary mediator (Li et al., 2022; Lei et al., 2024). Several factors could explain this discrepancy. First, urban vegetation may be more effective at capturing larger PM10 particles through deposition and impaction (Diener and Mudu, 2021). Second, our patient cohort might be particularly susceptible to the inflammatory and metabolic effects induced by PM10. While PM2.5 is known for its ability to penetrate deeper into the bloodstream, the metabolic responses observed in our cross-sectional data may be more strongly linked to PM10 exposure.

Furthermore, our granular analysis pinpointed TC and GLU as the most significant biomarkers. While previous research has broadly linked greenness to cardiometabolic disorders (Yang et al., 2020), our study identifies TC and GLU as particularly sensitive indicators in this environmental health context. The strong association with GLU aligns with studies suggesting that air pollution can induce insulin resistance and impair glucose metabolism (Yang et al., 2018; Burkart et al., 2022). The specific prominence of TC in the pathway to TH, and the significant roles of both TC and HDL-C in other outcomes, suggests that PM10 exposure in this population may have a pronounced effect on overall cholesterol regulation, a finding that complements broader evidence linking air pollution to adverse lipid profiles (Cesaroni et al., 2014). This specificity provides a more focused target for future mechanistic research, suggesting that TC and GLU could serve as key biomarkers for assessing the cardiovascular benefits of urban greening initiatives.

This study presented several methodological innovations for exploring the complex relationships between green space, air quality, PHR, and CHOs. (1) While traditional structural equation modeling has been widely used in previous epidemiological studies (Hays et al., 2005; Grande et al., 2020), PLS-SEMs have gained popularity in recent years for their efficiency at handling complex structural models with multiple constructs, indicators, and model relationships. For instance, some research employed PLS-SEMs to investigate the relationships between green structures, air pollution, temperature, and CVD mortality (Shen and Lung, 2016). Our study builds upon this trend, leveraging the ability of PLS-SEMs to explore more complex mediation pathways. (2) A key innovation in our approach is the application of chain mediation analysis to examine the mechanisms linking green space to cardiovascular health. While some previous studies have considered potential mediators between environmental factors and health outcomes, they typically focused on single mediators such as air pollution or physical activity (Bauwelinck et al., 2021; Yu et al., 2023). Our study extends this by incorporating multiple potential mediators—air quality and PHR—in a chain mediation model. This approach allows for a more comprehensive understanding of the complex pathways through which green space influences cardiovascular health. (3) Our study employed a series of increasingly refined models (Model Groups 1–4) to elucidate the associations at play: Model Group 1 established the overall associative pathway, Model Group 2 deconstructed the composite PHR into its sub-risks, Model Group 3 simultaneously examined the roles of different PM sizes (PM2.5 and PM10) and specific biomarkers (TC, HDL-C, GLU) in a competitive model, and Model Group 4 explored the relative importance of various individual vegetation indicators. This progressive refinement allowed for a more nuanced and comprehensive analysis of the relationships between variables. (4) In Model Groups 2–4, we considered both the interactive effects of multiple variables in chain mediation models and the individual effects of single indicators. This dual approach, using both comprehensive-effect and single-factor models, provides a more realistic representation of complex real-world mechanisms while also allowing for the identification of key individual factors. Our study employed different models based on the specific requirements of each analysis stage. This flexible approach, using comprehensive-effect models as the primary strategy with single-factor models as a pre-specified alternative to address multicollinearity, was applied in Model Groups 2, 3, and 4. This flexible approach to model selection enhanced the scientific rigor and accuracy of our findings. In conclusion, our methodological approach—combining PLS-SEMs with chain mediation analysis and a progressive model refinement strategy—offers a novel and comprehensive framework for investigating the complex relationships between environmental factors and health outcomes. This approach allows for a more detailed and nuanced understanding of the mechanisms linking green space to cardiovascular health, and it provides valuable insights for urban planning and public health strategies.

This study has several important limitations that must be considered when interpreting the findings. First and foremost, the cross-sectional and retrospective nature of our study design is a major limitation. The environmental exposures (VC and PM) were assigned based on annual averages for the year preceding hospitalization, and physiological data were collected during the hospitalization episode. This design does not allow for the establishment of clear temporal precedence, which is a prerequisite for making causal claims. Consequently, our findings should be interpreted as evidence of statistical association rather than causal mediation. The language used throughout this paper, such as indirect association and pathway of association, reflects this non-causal, exploratory framework. Second, our analysis did not include several crucial individual-level confounders. Data on patient age, sex, socioeconomic status, lifestyle factors such as smoking and diet, and pre-existing comorbidities were not available in the electronic medical records used for this study. These factors are known to be strong predictors of both residential location choices and cardiovascular risk. Their omission means that the associations we observed could be biased due to unmeasured confounding. For example, higher socioeconomic status might be associated with both living in greener neighborhoods and better baseline health, which could partially explain the observed link between vegetation and better health outcomes. Therefore, the results should be interpreted with caution. Third, the use of annual average data for environmental exposures, while common in such studies, may mask the potentially important effects of short-term or seasonal fluctuations in PM concentrations. The health impacts of acute exposure spikes might differ from those of chronic, long-term exposure. Finally, while we used a 1000m buffer to estimate exposure, this is an approximation and may not perfectly reflect an individual’s true exposure, which is influenced by daily mobility patterns, time spent indoors versus outdoors, and occupational exposures. Future research employing longitudinal designs, with detailed individual-level data on confounders and time-activity patterns, is needed to confirm the exploratory findings of this study and to establish causal relationships.