How does high-density built-environment affect the incidence of gestational diabetes mellitus

Summary

As one of the prevalent complications during pregnancy, this study discussed the association between gestational diabetes mellitus (GDM) and multi-source built-environment from the perspective of different spatial scales in mega-cities. This study centered on 4,355 women at a Shanghai hospital. Thirteen variables were selected, then multi-scale geographically weighted regression (MGWR) was used to clarity how these variables relate to GDM. The results showed that: MGWR effectively revealed the relationship between the built environment and GDM at different scales. At specific spatial scales, building-density (BD) and sky view factor (SVF) exhibit pronounced spatial heterogeneity. Based on the mean coefficients in MGWR, it can be inferred that for every 0.1 increase in SVF and green view index (GVI), the probability of GDM decreases by 3% and 1%. These findings delve into the association between the built environment and lifestyle-related diseases. They underscore the significance of developing place-specific policies in health interventions and urban planning.

Graphical abstract

Highlights

• •The relationship between multisource variables and GDM is discussed
• •Multiscale geographically weighted regression can better explain GDM in Shanghai
• •Green view index and sky view factor are influential factors
• •Some hierarchical and site-specific health interventions are proposed

Public health; Human metabolism; Urban planning

Introduction

In China, due to lifestyle changes caused by urbanization, the prevalence rate of Gestational Diabetes Mellitus (GDM) has reached 14.8%, which poses a serious threat to the health of pregnant and lying-in women.^1,2,3 As a major public health issue,^1,4 it is crucial to clarify the influencing factors of GDM to protect maternal and child health.^3,5

The potential links between built-environment variables and human behavior and disease have been extensively studied. As a place for human settlement, the built-environment is a synthesis of physical space,^6,7,8 social culture,^9,10,11 and economic environment.^12,13 It will affect residents' eating habits, outdoor communication, and frequency of physical activity, and thus affect their physical health.^14,15 Some researchers have explored the relationship between the built environment and diabetes.^{15,16,17,18,19} For example, living in a higher level of green space reduces maternal blood glucose levels and the risk of GDM.²⁰ On the contrary, in some regions, it has been shown that there is insufficient evidence for a possible association between residential greening and GDM risk.⁵ Research shows that land use patterns can significantly affect health. Specifically, living near a major road is associated with a higher probability of GDM.²¹ Urban form indicators such as building density (BD), building height (BH), floor area ratio (FAR), and sky view factor (SVF) influence the risk of GDM by modulating urban microclimate,^22,23 affecting residents' physical activity levels and psychological stress states, and thereby linking to metabolic health outcomes. Meanwhile, the spatial distribution characteristics of transportation infrastructure, including road network density (RND), accessibility, and the distribution of bus and subway stations, may alter residential exposure to air pollutants while also shaping daily physical activity patterns.¹⁹ Together, these factors constitute key built-environment variables that influence the occurrence of GDM. Previous studies on the relationship between the built environment and GDM have mainly focused on a single factor,²⁴ but few studies have taken the built environment as a complex integrated system. In particular, there are few studies on the relationship between GDM and complex built-environment variables in special groups such as pregnant women.

The proliferation of high-density cities is a remarkable phenomenon in the global urbanization process,²⁵ which brings complex challenges to the construction of livable and healthy cities.²⁶ In general, cities in the United States and Europe are less dense than densely populated cities in Asia, including Hong Kong, Singapore, Tokyo, and Shanghai.²⁷ The existing literature mainly comes from studies in the context of the United States and Europe, with limited attention to China’s rapidly developing mega-cities.²⁸ Shanghai is China’s typical mega-city, characterized by high population density, high building mass, limited green space, high housing costs, popular private car travel,^29,30 and a widespread diet high in sugar and salt. Therefore, how the unique built-environment of these cities affects the incidence of GDM in pregnant women still needs to take targeted investigated to provide a scientific basis for the formulation of public policies on urban governance.

Due to the diversity and extensiveness of pregnant women’s activities during pregnancy, the selection of study areas became a key factor affecting the accuracy of the assessment results. In previous studies, the relationship between the built-environment and disease was assessed by delineating a fixed range (300m or 500m buffer zone) centered on the residential address,^21,31 and the analysis process was simplified to a certain extent, and assess the health such as BMI and physical activity and diseases such as cardiovascular disease and respiratory disease associated with urban climate,^32,33 but considering the cross-regional activity patterns of pregnant women who need to work, leisure, and recreation during pregnancy, the built-environment exposed to them is far more complex and diverse than that of a single residential buffer zone. Therefore, the relationship between the built environment and disease based on the residential address method alone may have certain limitations and cannot represent the true level of exposure to or near the natural space area, so this study expands the single residential address to a broader and more flexible community-scale built-environment variables, and deeply explores the relationship between built-environment variables and GDM.

High-density built environments exhibit significant spatial heterogeneity, posing challenges for traditional regression models to accurately describe their impact on health outcomes across different scales. Some studies have used traditional methods, such as generalized estimating equations (GEEs) and logistic regression, to uncover the overall relationship between the built environment and GDM. For example, some used GEEs to demonstrate that increased exposure to green spaces can reduce the risk of GDM, emphasizing the relevance of built-environment characteristics in influencing GDM incidence.³⁴ Similarly, some applied logistic regression to reveal a significant association between extreme temperatures and increased GDM risk, highlighting the importance of considering extreme temperatures in GDM research.³⁵

However, while these traditional regression models can explain the overall relationships between variables, they fall short in elucidating associations across different spatial scales. In contrast, the geographically weighted regression (GWR) model,^36,37 as a spatial analysis method, can identify the risk factors of diseases and their geographic variations within urban built environments. GWR enhances the understanding of the complex relationships between environmental variables from a spatial perspective, which is difficult to achieve with traditional regression models.³⁸

For instance, a previous study utilized the GWR model to explore the impact of the built environment on chronic obstructive pulmonary disease (COPD) mortality rates and their geographic differences. The study found that the functional effects of built-environment variables varied by location, with the impact of main road density on COPD decreasing from west to east.³⁹ Additionally, a study on the relationship between the urban environment and residents' cardiovascular health in Madrid (Spain) showed significant differences in cardiovascular health between residents of the urban core and those in less populated surrounding areas.⁴⁰ This method can similarly be applied to explain the spatial dynamics of GDM. In the context of GDM, using GWR can help identify specific areas in cities where built-environment variables have a more pronounced impact on GDM incidence. Multi-scale geographically weighted regression (MGWR) model⁴¹ further improves GWR by adjusting the bandwidth (BW) to more precisely describe the relationships between variables across multiple spatial scales.⁴² MGWR holds significant potential for applications in urban health research, as more precise spatial analyses can help identify high-risk areas and facilitate effective interventions.⁴³ This refinement is particularly valuable in urban health studies because factors influencing health outcomes often operate at different spatial scales. For example, some demonstrated that the spatial heterogeneity in NO₂ distribution between central and suburban areas of Wuhan leads to varying degrees of respiratory disease hospitalization rates.⁴⁴ Similarly, some showed that MGWR effectively explains the spatial scales at which different built-environment variables influence cardiovascular diseases.⁴⁵

However, despite MGWR’s excellent performance in various health studies, no research has yet used MGWR to study the association between the built environment and GDM. By capturing these multi-scale effects, MGWR can conduct more detailed analyses to understand how specific built-environment characteristics at different spatial scales contribute to GDM risk. Therefore, integrating GWR and MGWR into studies on the impact of high-density built environment on GDM can significantly enhance our understanding of the spatial dimension of this relationship, ultimately providing insights for better urban planning and public health interventions aimed at reducing GDM incidence.⁴⁶

In summary, in view of the complexity of the built environment, it is necessary to establish a comprehensive variable evaluation system from the three levels of material space, social culture, and economic environment, so as to scientifically and accurately reveal its impact on GDM. This study takes the ultra-high density city of Shanghai as an example, obtains the cohort data of 4533 pregnant women residing in the central urban area of a hospital, establishes the index system of built-environment variables through multi-source data, and applies the geographical spatial regression model to analyze the relationship between built-environment and GDM in pregnant women. The overall effect of the built-environment on GDM and the spatial scale effect of different variables were revealed, and the key factors affecting the density of GDM in different regions are identified. These findings offer fresh perspectives and methodologies for utilizing non-pharmacological interventions (NPIs) to mitigate the risk of GDM in the future, while simultaneously providing a scientific basis for the formulation of public health policies in high-density cities, thereby contributing to sustainable urban development.

Results

Distribution of the pregnant women in this study

ELP participants came from all over the country, and were distributed in 32 provincial-level administrative regions across the country except Tibet and Taiwan Province according to Figure 1A, so the data were extensive, and Xinhua Hospital has established cooperative relations with a number of medical institutions, moreover, about 85.7% of the participants of the ELP were located within the Shanghai inner city, according to Figure 1B, this data can reflect the relationship between the built-environment and GDM in the Shanghai inner city (Figure 1C) to a certain extent. Within the study area, these eleven administrative regions involve 117 community units (N = 117) at the street level and below. Figures 1D and 1E were the distribution of the women and pregnant women diagnosed with GDM. The number of GDM pregnant women was summarized at the community level, and then the density of GDM was calculated by measuring the number of GDM diagnoses in the unit of land area of each community resident (Figure 1F).

Figure 1.

Distribution and density of diagnosis of GDM

(A–F) showed the distribution of pregnant women and those diagnosed with GDM across different geographical scopes.

Built-environment variables’ selection

BD was calculated using building footprint data from Bigemap, while the FAR was computed by combining this data with height information retrieved from Zenodo.⁴⁷ The road network data obtained by OpenStreetMap (OSM) calculated the RND and accessibility of the road network (ARN). ARN refers to the average time for residents to drive between road network nodes, with a shorter time indicating higher accessibility. RND was calculated by dividing the total length of the road network by the unit area. The land use diversity (LUD) was calculated using the Shannon Diversity Index based on the land use type data provided by Zenodo.⁴⁸ The data of points of interest (POIs) in Shanghai and the distribution of bus and subway stations in each community were provided by Bigemap. Kernel distribution density of the restaurant facilities (DKDRF) reflected the spatial distribution and aggregation degree of POIs in milk tea, fast food, and other restaurant facilities. Subway station density (SSD) and bus stop density (BSD) reflected the number of subway stations and bus stops per unit community area. Google Earth Engine obtains and calculates the normalized difference vegetation index (NDVI) (annual average data from 2016 to 2024) data, ultimately yielding the annual average NDVI values during the pregnant women’s gestation period. Green view index⁴⁹ (GVI) and SVF data were calculated using street view images obtained from Baidu Maps. We obtained the house price (HP) data at the community level in the Shanghai inner city from Beike, which was used as a socio-economic variable corresponding to the building value.

According to the multi-collinearity analysis, with the correlation of each index being less than 0.8 (Figure 2), variance inflation factor (VIF) < 5, 12 variables were finally selected as the variable system to explore the relationship between the built-environment and the density of GDM. According to the correlation analysis, there is a significant positive correlation between BH and FAR (0.681), RND (0.579), BSD (0.586), SSD (0.549), DKDRF (0.645), and HP (0.580). Meanwhile, BD has a significant positive correlation with BSD (0.664) and DKDRF (0.680), and a significant negative correlation with LUD (−0.581) and SVF (−0.476). Additionally, LUD shows a relatively significant negative correlation with BSD (−0.566), SSD (−0.356), DKDRF (−0.396), and HP (−0.447). NDVI has a relatively significant negative correlation with BH (−0.341), BD (−0.465), RND (−0.410), BSD (−0.365), SSD (−0.303), DKDRF (0.284), and HP (0.400). GVI exhibits a significant negative correlation with BH (−0.294), BD (−0.402), RND (−0.328), BSD (−0.336), SSD (−0.367), DKDRF (−0.265), and HP (−0.488). It is found that BD, FAR, SVF, ARN, and DKDRF had a significant impact on GDM (p < 0.05 or p < 0.01), other variables have a lower correlation. Among them, BD, FAR, BSD, SSD, and DKDRF are positively correlated with GDM, while SVF, NDVI, and HP are negatively correlated with GDM.

Figure 2.

Results of the correlation analysis

∗∗p < 0.01; ∗p < 0.05.

The spatial auto-correlation analysis of built-environment variables

The global Moran’s I and the local indicators of spatial association (LISA) were used to analyze the spatial distribution characteristics and spatial autocorrelation of each variable, and measure whether the spatial distribution of each built-environment variable is different and clustered.

From Table 1, the global Moran’s I statistical value (I) ranging from 0.3664 to 0.7456, was in the range of 0 < I ≤ 1 which means it shows a positive spatial autocorrelation pattern. The p-value is statistically significant, and the Z score is positive (Z score>2.58), indicating that the variable is spatially clustered more than expected. HP showed the highest I value (I = 0.7456), indicating that the HP of urban communities was basically consistent with those in neighboring areas and showed obvious spatial agglomeration. It was also consistent with the SVF of the adjacent region overall, though it showed the lowest I value (I = 0.3664).

Table 1.

The results of the global Moran’s I

Variable	I	Z score	p-value
BH	0.5973	10.2009	< 0.0000
BD	0.4164	7.1623	< 0.0000
SVF	0.3664	6.3126	< 0.0000
LUD	0.5501	9.4359	< 0.0000
RND	0.5137	8.8359	< 0.0000
ARN	0.5578	9.5747	< 0.0000
BSD	0.4801	13.5206	< 0.0000
SSD	0.3974	6.9487	< 0.0000
NDVI	0.4771	8.2254	< 0.0000
GVI	0.4437	7.6041	< 0.0000
DKDRF	0.5734	9.8439	< 0.0000
HP	0.7456	12.7559	< 0.0000

The Figure 3 reflected that the built-environment variables of Shanghai exhibit four patterns of spatial agglomeration: high-high cluster, high-low cluster, low-high cluster, and low-low cluster. In the center of the study area, LUD, ARN, and NDVI showed a low-low cluster, while BH, BD, GVI, DKDRF, and HP showed a high-high cluster. It shows that in districts such as Hongkou District and Jing’an District, the building functions are relatively simple, the green spaces are scarce and concentrated in distribution, the road network is densely covered, the buildings are tall and closely arranged, and the catering facilities are highly concentrated in distribution, which is consistent with the real situation in Shanghai. BD in some communities in Pudong presents a high-low cluster situation, where, despite having a relatively high BD, the surrounding areas have lower BD, indicating potential future development into a high-high cluster spatial characteristic. In addition, the regions with no significant LISA are mostly distributed on the edge of Shanghai's inner city, suggesting that the spatial correlation of the marginal communities was low, and there were multiple spatial correlations or heterogeneity in the whole study area.

Figure 3.

LISA cluster patterns

(A–L) showed the cluster pattern of every variable.

Model evaluation

In order to compare the performance of different models in terms of fitting effects and results, the same set of variables was used to perform ordinary least squares (OLS), GWR, and MGWR. The OLS model displays a low adjusted R² value of 0.1531, indicating that 84.7% of the variability in GDM across Shanghai remains unaccounted for and likely influenced by other unknown factors. Notably, the limitation of the OLS model is that it cannot account for the variability of spatial local regression in the relationship between the dependent and independent variables, whereas the GWR model is used to deal with spatial data regression, allowing the coefficients to change spatially. To investigate the spatial variation in the relationships between the built-environment and GDM, both GWR and MGWR are utilized with the same set of variables used in the OLS model. The GWR model outperforms the OLS model, as demonstrated by its higher explanatory power of 0.1757 of the variance in GDM and a lower Akaike’s information criterion (AICc) value of 692.562. Moreover, the MGWR model exhibits the most favorable performance among all models, with the highest adjusted R² value of 0.6430 and the lowest AICc value of 262.0740 (Table 2).

Table 2.

Comparison of the goodness of fit for different models (N = 117, which means the number of community units)

Criterion	OLS	GWR	MGWR
Adjusted R2	0.1531	0.1757	0.6430
AICc	692.7785	692.5620	262.0740

Adjusted R² and AICc can be used to test and compare model performance.

OLS results

Table 3 displayed the outcomes of the OLS model for the density of the GDM. Combining positive and negative values for the coefficients illustrates that the built-environment variables have different relationships with the density of GDM in magnitude and direction. According to the coefficients of the OLS model, RND and DKDRF are positively related to the density of GDM. Additionally, the BH, LUD, and the BSD, SSD, and the NDVI are negatively related to the density of GDM. It can be seen from the results that the SVF, ARN, and HP are statistically significant, according to the probability and t-value, which are used to indicate whether there is statistical significance. The Koenker (BP) statistic is used to determine whether the explanatory variable of the model has a consistent relationship with the dependent variable in both the geographical and data space, and when it does not exhibit statistically significant, it means that the variable is suitable for GWR analysis, and it can be seen that the Koenker (BP) statistic (13.0272, degree of freedom: 0.3671) is not statistically significant, indicating that GWR analysis is performed.

Table 3.

The results of the OLS model

Variable	Coefficient	Robust Standard Error	p-value	t-value	VIF
Intercept	37.9564	18.1415	0.0232∗	2.3029	–
BH	−0.3952	0.2782	0.1023	−1.6478	3.0344
BD	0.1261	0.0865	0.2789	1.0882	3.2487
SVF	−23.2780	15.1633	0.0702	−1.8289	2.5212
LUD	−15.4824	22.8145	0.7054	−0.3789	2.1652
RND	0.1535	0.1525	0.3512	0.9362	2.1533
ARN	−1.7191	0.5574	0.0198∗	−2.3658	1.7402
BSD	−0.1111	0.1515	0.5577	−0.5881	2.9011
SSD	−0.0098	0.1635	0.9662	−0.0423	2.1541
NDVI	−4.5224	7.1572	0.7507	−0.3185	2.2674
GVI	4.4155	9.4651	0.6690	0.4286	1.8901
DKDRF	0.0062	0.0050	0.3693	0.9015	3.3761
HP	−0.0001	0.0001	0.0947	−1.6862	3.3804

∗p value < 0.05; Koenker (BP) Statistic: 13.0272, degree of freedom: 0.3671, not significant.

GWR results

GWR outputs a single optimum BW of 117 for all variables, which presumes that the associations between the built environment and GDM are at the same spatial scale and seem excessively restricted. In general, the BW can be understood as a spatial scale that describes the spatial range within which the spatial relationships of various samples remain stable. The coefficients of the variables in the GWR model have relatively large variations and are significant for a few built-environment variables (Table 4). It shows that LUD, SSD, NDVI, and GVI are positively correlated with GDM in some regions and negatively correlated in others, according to the minimum coefficient and maximum coefficient. BH, SVF, LUD, ARN, BSD, and HP are negatively correlated with GDM, while BD, RND, and DKDRF are positively correlated, as the mean coefficient of the GWR model. ARN and HP are significant in Shanghai (Figure 4).

Table 4.

Summary statistics for GWR parameter estimates (N = 117)

Variable	GWR
	Mean	STD	Min.	Max.	BW
Intercept	42.113∗	3.624	32.781	48.839	117
BH	−0.392	0.056	−0.455	−0.203	117
BD	0.131	0.023	0.090	0.184	117
SVF	−22.765	2.524	−27.065	−16.205	117
LUD	−28.345	10.518	−51.640	2.281	117
RND	0.144	0.063	0.029	0.287	117
ARN	−1.820∗	0.279	−2.527	−1.401	117
BSD	−0.118	0.032	−0.190	−0.037	117
SSD	0.016	0.032	−0.069	0.064	117
NDVI	−2.043	9.176	−13.930	21.628	117
GVI	3.888	4.309	−9.439	8.013	117
DKDRF	0.005	0.001	0.002	0.006	117
HP	−0.0001∗	0.0001	−0.0001	−0.0001	117

∗p < 0.05; STD: standard deviation. Spatial kernel: Adaptive bisquare; Criterion for optimal bandwidth: AICc, Number of varying coefficients: 13.

Figure 4.

Spatial distribution of the significant coefficients from the GWR model

(A–C) showed two variables’ mean coefficients, which are significantly correlated and the intercept, the confidence level is 95%. Data are represented as the mean.

MGWR results

In contrast to the GWR single BW of 117, the BW of the MGWR variables varies, which can be classified into two groups, one is the spatial distribution of variables with no obvious heterogeneity, close to or global scale (BW is or close to the sample size), such as the BH, the ARN, RND, the BSD and SSD, and NDVI, GVI, DKDRF, HP, and its STD is relatively small, and local scale variables with relatively small BW, which have a small impact on the spatial distribution of GDM, such as the BD and SVF, and therefore a large STD.

Table 5 reflects the correlation of the built-environment variables and the density of GDM. The intercept is significant at a localized scale of 43, which indicates a noticeable spatial variation at the surface (STD = 0.457), with coefficients ranging from −0.456 to 1.044 and a mean value of 0.025. BD (BW = 45) exhibiting a stronger degree of spatial heterogeneity, with coefficients lying between −0.278 and 0.616, shows a transition from negative to positive correlation as it moves from Baoshan to Pudong, from northwest to east. LUD, with coefficients ranging from −0.028 to 0.033, displays a transition from strong negative to strong positive correlation in its impact on GDM, from Baoshan to Pudong. Notably, the Jing’an district and parts of the surrounding areas exhibit the strongest negative correlation, while most of the Pudong District shows positive correlation. DKDRF with coefficients varying from −0.013 to 0.011, exhibits a gradual transition from negative to positive correlation with GDM as it moves from districts like Yangpu to Pudong, from north to south. In contrast, HP shows a reverse trend, transitioning from negative to positive correlation as it moves from Pudong to Yangpu and other districts, from south to north.

Table 5.

Summary statistics for MGWR parameter estimates (N = 117)

Variable	MGWR
	Mean	STD	Min.	Max.	BW	Effective scale	Level
Intercept	0.025∗	0.457	−0.457	1.044	43	Local	District
BH	−0.119	0.009	−0.132	−0.093	116	Global	Whole area
BD	0.044∗	0.211	−0.278	0.616	45	Local	District
SVF	−0.316∗	0.034	−1.159	−0.058	44	Local	District
LUD	−0.010	0.015	−0.028	0.033	116	Global	Whole area
RND	0.065	0.016	0.039	0.103	116	Global	Whole area
ARN	−0.093	0.021	−0.142	−0.064	116	Local	Whole area
BSD	0.032	0.006	0.012	0.037	116	Global	Whole area
SSD	0.015	0.003	0.014	0.024	116	Global	Whole area
NDVI	0.080	0.010	0.067	0.106	116	Global	Whole area
GVI	−0.128∗	0.043	−0.220	−0.068	109	Local	District
DKDRF	−0.000	0.006	−0.013	0.011	116	Global	Whole area
HP	−0.010	0.009	−0.131	−0.093	116	Global	Whole area

∗p < 0.05. Number of iterations used: 51, score of change (SOC) type: smoothing f, termination criterion for MGWR: 0.00001, effective scale indicates the extent to which the related variable has a significant association with the density of GDM.

BH demonstrates a gradually intensifying negative correlation trend as it moves from Baoshan to Pudong, from northwest to southeast. ARN, completely differently, exhibits a gradually increasing negative correlation from Putuo, Changning, and Xuhui toward Pudong, from west to east. GVI, with a BW of 109 approaching the global scale, displays notable spatial variation (STD = 0.043), with its coefficient ranging from −0.068 to −0.220, and its impact on GDM gradually intensifies from west to east, from Xuhui and Changning toward Pudong, in terms of negative correlation. In the meantime, SVF, as a local variable with a BW of 44, also exhibits significant spatial variation (STD = 0.034). It presents a negative correlation that is weaker at the Shanghai Outer Ring Road edges and stronger at the center of the study area for GDM, with Hongkou District showing the strongest negative correlation.

RND exhibits a gradually increasing positive correlation with GDM as it moves from Pudong to Baoshan and Jiading, from east to northwest. Similarly, BSD displays a gradually intensifying positive correlation trend from Pudong to Putuo, Changning, and other districts, from east to west. SSD, meanwhile, shows a gradually strengthening positive correlation trend from Xuhui to Baoshan, from south to north. Additionally, the influence of NDVI on GDM gradually intensifies from Xuhui toward Yangpu and Pudong, from southwest to northeast, in terms of positive correlation.

In summary, BH, SVF, LUD, ARN, GVI, and HP are negatively correlated with GDM, while RND, BSD, SSD, and NDVI are positively correlated with GDM. BD and DKDRF show positive correlations in some regions and negative correlations in others, however, only BD, SVF and GVI exhibit statistical significance, while the other indicators are statistically insignificant (Figure 5), and this phenomenon is mainly concentrated in the center and east of the study area such as Hongkou and Pudong. Based on the mean coefficient of each variable, it can be inferred that for every 0.1 increase in BD, the probability rises by 0.45%, for every 0.1 increase in SVF, the probability of GDM diagnosis decreases by 3%, and for every 0.1 increase in GVI, the probability of GDM diagnosis may decrease by 1%.

Figure 5.

Spatial distribution of the significant coefficients from the MGWR model

(A–L) showed the MGWR mean coefficients, the confidence level is 95%, and the gray area indicates that the coefficient is not significant. Data are represented as the mean.

Robustness analysis

To assess the extent to which unmeasured confounders interfere with the core results, this study adopted the E-value (E-value = RR + sqrt [RR∗(RR-1]) (RR > 1)) to test the robustness of the results.⁵⁰ For BD, the regression coefficient for GDM ranged from 0.134 to 0.164, with a calculated E-value of 1.55–1.64. This means that an unmeasured confounder must have an association strength of ≥1.55 with both BD and GDM to offset their correlation. However, a study on the population in eastern Massachusetts, USA, showed that pregnant individuals living in the highest density tertile had a higher risk of GDM compared to those in the lowest density tertile (500 m: odds ratio (OR): 1.17⁵¹) (lower than 1.55), it is relatively difficult for such a strongly associated confounder to exist, indicating that the results related to BD have a certain degree of robustness. For SVF, the regression coefficient ranged from −1.159 to −0.058, with an E-value of 2.04–5.82. An unmeasured confounder must have a dual association strength of ≥2.04 to offset the negative correlation between SVF and GDM, demonstrating extremely strong result robustness. For GVI, the regression coefficient ranged from −0.187 to −0.220, with an E-value of 2.01–2.21. An unmeasured confounder must have a dual association strength of ≥2.01 to offset its correlation, indicating strong result robustness. In conclusion, the robustness of the results related to BD is relatively weak, but it is highly difficult for strongly associated unmeasured confounders to exist. The E-values of SVF and GVI are at relatively high levels, indicating extremely strong robustness of their respective results.

Discussion

Key findings

Amidst China’s rapid urbanization and the rise of mega-cities, lifestyle factors such as motor vehicle-dominated urban transportation, unhealthy dietary habits, and inadequate physical activity have triggered clinical risk factors for GDM, posing a significant threat to the health of pregnant women, a special population group. This study focuses on this critical health threat, delving into the relationship between high-density built environment and GDM. From the perspective of urban public space governance, it is crucial to precisely identify and determine the key built environment variables associated with the disease and their spatial distribution. Based on these findings, government policymakers and urban planners can propose targeted public health prevention strategies for urban renewal and urban design, and create a more equitable built environment to further enhance the living conditions and quality of life for urban residents. Moreover, conducting research at the community level, where the basic research units correspond to the management units of local urban governments, can facilitate each urban community in adopting targeted strategies tailored to their specific conditions.

This study found that BD, RND, BSD, SSD, and HP have a positive correlation with GDM, while BH, SVF, ARN, and GVI have a negative correlation with GDM. ARN is a commonly used variable in the study and analysis of the association mechanism between the built-environment and human health. The results of the MGWR model indicate a negative correlation between the ARN and the risk of GDM. Specifically, the lower the ARN value, which means the higher the road network accessibility, the higher the risk of residents being diagnosed with GDM. This finding seems to contradict the common understanding that high accessibility promotes physical activity, thereby reducing the risk of metabolic diseases.^19,52,53,54 However, a reasonable explanation can be derived by considering the characteristics of Shanghai’s urban development. Unlike some European cities where active transportation (walking and cycling) dominates,⁵⁵ Shanghai’s high accessibility relies more on the improvement of the motor vehicle transportation system. Travel modes in Shanghai are still dominated by motor vehicles,^29,30 which has not significantly increased the level of physical activity among residents. Moreover, areas with high accessibility are often accompanied by greater motor vehicle traffic volume, which may increase residents' exposure to environmental hazards such as air pollution and traffic noise.²¹ Such exposures have been proven to be associated with an elevated risk of GDM. In contrast, areas in Shanghai’s suburban regions with lower accessibility and higher ARN values, despite being slightly less convenient in terms of transportation, typically possess more abundant green space resources. Additionally, environmental pressures such as traffic pollution and noise disturbance in these areas are significantly lower than those in the central urban areas. These favorable environmental conditions may, to a certain extent, offset the limitations imposed by low accessibility on physical activity, thereby exerting a relatively better protective effect against the risk of GDM.

It is worth noting that the relationship between NDVI and GVI and GDM in this study, the NDVI shows a positive correlation, which is consistent with the results of some previous studies. Some studies have suggested that there appeared to be no association between green space and GDM,^21,56 and some studies found that the green space had a positive impact on GDM,³⁰ while others showed a negative impact between NDVI and GDM,⁵ however, it was observed that GVI demonstrated the most significant negative correlation with GDM in Pudong. This finding may be attributed to the fact that NDVI alone cannot fully account for the proximity and accessibility of green spaces.⁵⁷ On the other hand, GVI, as a crucial factor that quantifies the spatial proportion of visible green vegetation within a person’s horizontal field of view, reflects the quality of urban green space planning and the psycho-physical experiences derived from being immersed in green landscapes. GVI is directly related to residents' perception and evaluation of the urban green space environment, shaping their overall well-being and potentially influencing health outcomes.⁵⁸

The present study has uncovered a significant negative correlation between SVF and the density of GDM in districts such as Pudong and Hongkou. A relevant study had demonstrated that SVF exerts a notable influence on microclimate and thermal comfort.⁵⁹ Furthermore, the degree of visual openness within residential areas has been shown to have potential impacts on health. Consequently, urban planners must pay attention to controlling building heights and densities in areas such as Hongkou and Pudong, to provide residents with a more expansive visual environment.

The findings suggest a significant impact on HP, with pregnant women living in low-housing communities more likely to develop GDM, consistent with the findings of existing studies,⁶⁰ which demonstrate the important impact of socioeconomic deprivation on adverse pregnancy outcomes. This is because lower levels of health care and unhealthy behaviors (smoking, drinking, and unhealthy diets) are more prevalent in low-economic communities.⁶¹ In addition, studies have shown that people living in communities of lower socioeconomic status face inequalities in the level of health care,⁵⁰ further increasing the development and incidence of GDM.

This study found that the MGWR model can reveal the spatial scale association effect between built-environment variables and the density of GDM (Table 5). Compared with the GWR model, the intercept coefficient of the MGWR model has a larger range and a higher standard deviation, indicating that the spatial heterogeneity is higher. At the same time, a smaller BW was used for BD and SVF, which means that it can capture more refined spatial variations, necessitating interventions at the community level. Variables with larger BWs, such as HP and NDVI, support the formulation of city-wide policies, which means that they captured more granular spatial variations. There is a significant negative correlation between the ARN and the density of GDM, which implies that higher accessibility is associated with a higher risk of GDM according to the definition, and it is obviously affected by geographical location. Only in Huangpu District, Hongkou District, and Yangpu District, such as Lujiazui Street, Daqiao Street, Wujiaochang Town, and other community units, the correlation was presented. Variables such as RND, DKDRF, HP, BH, BSD, SSD, and BD also show different degrees of geographical differences. For example, the small range and standard deviation of the SVF estimate indicate that the impact of the indicator tends to be consistent across different geographical locations. The relationship between BD, LUD, RND, BSD, SSD, NDVI, GVI, DKDRF, and the density of GDM did not show statistical significance. This does not mean that these variables are not related to GDM, just that they have less impact than other variables. In addition to build-environment variables, pregnant women’s personal characteristics, such as age, family medical history, and habits such as smoking and drinking alcohol, can also play a role in GDM.^3,62

Based on the above research findings, it is recommended to implement a differentiated spatial governance strategy. For high-risk areas such as the Hongkou District and the Pudong New District, it is necessary to strengthen the control of architectural planning. Strictly limit the height and density of buildings. Through scientifically setting the upper limit of the plot ratio and optimizing the building setback standards, systematically increasing the open spaces in the city and effectively improve the SVF value of the area. In terms of building layout, attention should be paid to creating a well-arranged spatial form to avoid the formation of a closed and oppressive skyline landscape, so as to effectively improve the visual environment and psychological feelings of residents. In addition, in community renewal projects, it is necessary to simultaneously promote the project of improving the quality of public spaces. According to local conditions, facilities such as pocket parks, open-air squares, and fitness trails should be added to create diversified outdoor leisure and sports spaces for residents. This not only reduces exposure to road traffic pollution and noise but also further lowers the risk of GDM through environmental intervention.

Advantages of the study

The aim of this study was to investigate the association between different built-environment variables and the density of GDM at different levels and spatial scales, and to find out whether these screening study variables were positively or negatively correlated with the density of GDM. MGWR was used to analyze the spatial scale effects of multidimensional built-environment variables associated with GDM. Compared with existing studies, a certain degree of innovation has been made in the screening of built-environment variables, the selection of research regions, and even the research methods.

• 1In this study, a more multivariate and granular selection of built-environment variables was considered. In the existing field, there have been many achievements in the study of the built-environment and human health, but there are few detailed and in-depth studies on the built-environment of GDM, which is a specific disease, and most of the architecture and street form factors and socioeconomic variables are considered when studying the pathogenesis association mechanism of complications during pregnancy, land use, road traffic, and green open space, covering as many as 12 variables, instead of only selecting a few variables such as NDVI or accessibility when studying the association mechanism of pregnancy complications in the existing field, HP are rarely considered and are added as supplements to socioeconomic variables. The selection of built-environment variables has been expanded compared with previous studies.
• 2The selection of the area within Shanghai’s inner city as the research region is based on its characteristics such as high population density, large building volume, limited green space, high housing costs, prevalent private car travel mode, and a significantly high-sugar and high-oil diet habit. It is a typical super metropolis in China with a complex built environment. Studying the relationship between this unique built-environment and GDM aims to provide targeted evidence for formulating scientific and effective public health policies.
• 3In this study, several methods were used to explore the correlation between the built-environment and GDM. First, in terms of research methods, MGWR was used for analysis, which had a higher degree of fit. Second, in this study, the spatial scale effect of the association between the built environment and GDM in MGWR was studied, and a more accurate spatial regression model was established by making full use of the comprehensive case address information. These provide fresh perspectives and methodologies for utilizing NPIs to mitigate the risk of GDM.

In a nutshell, a more comprehensive and detailed set of built-environment variables based on multi-source data quantification greatly promotes the study of the correlation mechanism between the built-environment and public health, and provides a certain reference value for public health intervention. Furthermore, results of the study contribute to the development of improved urban health governance strategies from the perspective of urban spatial layout and planning, and offer certain reference value insights for promoting a more reasonable and efficient sustainable urban pattern.

Conclusion

This study illustrates the intrinsic association mechanism between the high-density urban built environment and GDM by taking the Shanghai inner city as a case study.

• 1Multi-source data are aggregated to measure and calculate the built-environment variables, which greatly advances the study of the interaction between the built environment and public health.
• 2The global Moran’ I obtained in this study ranged from 0.3664 to 0.7456, indicating that the variables in the study area showed a significant positive spatial autocorrelation, that is, different built-environment variables had a significant impact on the spatial non-stationary characteristics of GDM, and there were significant differences in different geographical locations at specific scales.
• 3MGWR method was used to measure the spatial scale effect of the relationship between the built environment and GDM. Compared with GWR and OLS statistical models, MGWR (Adjust R² = 0.643) showed a better fit for the relationship between the built-environment and GDM, indicating that MGWR is an effective method to analyze the relationship between the built-environment and GDM at multiple levels.
• 4In terms of spatial distribution, BD and SVF have a small spatial scale, which is a local variable (BW is relatively small), and have a large impact on the distribution of GDM in different regions. BH, LUD, ARN, RND, BSD, SSD, NDVI, GVI, DKDRF, and HP have a large scale, with GVI and SVF having a significant negative correlation with GDM.
• 5Based on the mean coefficients of each variable in the MGWR model, it can be inferred that within a certain range, for every 0.1 increase in SVF, the probability of a GDM diagnosis decreases by 3%, and for every 0.1 increase in GVI, the probability of a GDM diagnosis may decrease by 1%, especially in Hongkou and Pudong. Other built-environment variables have a certain impact on the GDM, which are insignificant.

These findings provide scientific evidence for urban planners and government policymakers to formulate community-scale urban renewal and urban design optimization strategies from the perspective of urban public space governance, thereby creating a more equitable and sustainable urban built-environment and enhancing the living conditions and quality of life for urban residents.

Limitations of the study

There are some limitations in this study, and the direction of future research is proposed.

Firstly, this study did not take into account other variables that may affect GDM, such as air pollution, noise, temperature, and also the characteristics of the pregnant women. Future studies can supplement relevant data to improve the reliability of the results.

Secondly, this study only revealed the correlation between built environment and GDM, without analyzing the mediating effects of variables.

Thirdly, due to the scarcity of maternal data, comprehensive GDM data from National Health Commission of the People’s Republic of China and other sources were not obtainable; instead, data from a single hospital was used as a substitute for the overall GDM level within Shanghai inner city, which poses certain limitations.

Finally, this study focuses only on GDM. Future research could use a variety of health data to analyze the relationship between the built-environment and physiological indicators or multiple disease conditions.

Resource availability

Lead contact

Requests for further information and resources should be directed to and will be fulfilled by the lead contact, Liqiang Zheng (liqiangzheng@126.com).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •The data reported in this study cannot be deposited in a public repository due to privacy and ethical restrictions. To request access, contact Xinhua Hospital Affiliated to Shanghai Jiao Tong University School of Medicine. In addition, processed datasets derived from these data have been deposited at Document S1. Data are publicly available as of the date of publication.
• •This paper does not report original code.
• •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Acknowledgments

This study was partly funded by the Collaborative Innovation Program of Shanghai Municipal Health Commission (2020CXJ001) and the Xinhua Hospital Early Life Plan project, Shanghai Jiao Tong University School of Medicine. All individuals provided informed consent to participate in this study and approval was provided by the Ethics Committee of the Xinhua Hospital Affiliated to Shanghai Jiao Tong University School of Medicine (Document S2. The Xinhua Hospital Ethics Committee Chinese) and (Document S3. Xinhua Hospital Ethics Committee English).

The views expressed in this article are those of the authors. The funder did not play any role in study design, data collection and analysis, interpretation of the results, the decision to publish, or preparation of the manuscript.

Author contributions

Conceptualization, F.G., N.L., R.P., and L.Z.; methodology, F.G., N.L., and L.Z.; software, N.L., X.X., B.W., Y.C., and F.G.; validation, F.G., R.P., L.Z., and D.Z.; formal analysis; R.P., H.J., L.Z., D.Z., and Q.Z.; investigation, R.P., L.Z., Q.Z., and H.J.; data curation; R.P., Q.Z., L.Z., and H.J.; original draft preparation; F.G., N.L., B.W., Y.C., and R.P., review and editing, F.G., L.Z., D.Z., Q.Z., and N.L.; visualization, N.L., F.G., and R.P.; supervision, L.Z., D.Z., Q.Z., H.J., and F.G.; project administration, L.Z., D.Z., Q.Z., F.G., and H.J.; funding acquisition, L.Z., D.Z., Q.Z., H.J., and F.G. All authors have read and agreed to the published version of this article.

Declaration of interests

The authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Software and algorithms
QGIS	QGIS Development Team	https://www.qgis.org/
GWR	GWR4 development team	https://gwr.maynoothuniversity.ie/gwr4-software/
MGWR	Professor Stewart Fotheringham of the School of Geographical Sciences and Urban Planning, Arizona State University, Arizona State, U. S.	https://sgsup.asu.edu/sparc/software
Other
The land use type and Building height	Zenodo	https://zenodo.org/
Road Network Data	Open Street Map (OSM)	http://www.osm.org
POI data and building footprint data	Bigemap	www.bigemap.com
NDVI	Google earth engine	https://earthengine.google.com/
GVI and SVF	Baidu Map	https://map.baidu.com
House price data	Beike	https://www.ke.com

Experimental model and study participant details

This study employs a prospective cohort design to dynamically track maternal health status throughout pregnancy. The data was from the ELP of Xinhua Hospital Affiliated to Shanghai Jiao Tong University School of Medicine, and ELP is a multi-center prospective observational study, including cohort studies and case-controlled studies. A total of 10685 pregnant women data from 2016 to 2024 were obtained. This study focuses on the impact of the Built-environment on GDM in pregnant women at the community and street scales, centering on the macro spatial environment without incorporating individual characteristics of pregnant women such as age and ethnicity. In the “robustness analysis” section, we discuss the potential impact of unmeasured confounding factors on the results and verify the stability of the core conclusions, indicating that the current research findings still have good robustness and reliability.

Method details

The study area was selected as the central urban area of Shanghai (Shanghai inner city), that is, the area within the Shanghai Outer Ring Road.

This study employs a prospective cohort design to dynamically track maternal health status throughout pregnancy. We focused this analysis on singleton births to mothers aged 18 years or less than 44 years. The stillbirth or antepartum stillbirth were excluded. Women with an extreme preterm (<28 weeks gestation) or post term (>44 weeks gestation) birth were excluded, as were those with pre-existing diabetes (type 1 or type 2), and residing out of Shanghai inner city or missing home address, as evidenced by the questionnaire survey, leaving 4533 subjects in the final analysis, and 832 mothers were diagnosed with GDM. According to the WHO's international classification of diseases (ICD), the types of pregnancy complications in this study include GDM (O24.4).

Built-environment data primarily collected from OSM, Bigemap, Google earth engine, Baidu Map and Beike. The built-environment variable includes three categories: physical space, social-culture, and economy, and the physical space data include: BH, BD, FAR, SVF, LUD, RND, ARN, BSD, SSD, NDVI, GVI. Social-culture data includes: DKDRF. Economy data includes: HP. On this basis, the average of the plots is calculated in the units of community units at the street level and below.

Quantification and statistical analysis

This study employs geo-spatial models to explore the impact of the built environment on GDM at the street and community scales, the 11 administrative regions within the study area are divided into 117 community units (N = 117) as the research sample.

Moran’s index

Moran's index (Moran’s I) is a statistical indicator used to measure spatial auto-correlation. Spatial auto-correlation refers to the presence or absence of similar attribute values in adjacent or close areas of geographic space (Section the spatial auto-correlation analysis of built-environment variables). The global Moran’s I was used to measure the strength of spatial correlation, the Z score to quantify the extent to which the observed spatial pattern deviates from a random distribution, and the p-value to determine the statistical significance of the spatial correlation results.

The four spatial agglomeration patterns (“High-High, Low-Low, High-Low, Low-High”) of LISA were identified by the association between the attribute value of each spatial unit and its neighbors, and validated by significance tests to exclude random errors.

AICc and adjusted R² were used to compare the goodness of fit for different models.

OLS

OLS is a statistical method used to estimate parameters in a linear regression model by minimize the sum of squares residuals, which is the sum of squares of the error between the observed and predicted values. In linear regression, a linear relationship is assumed between the dependent and independent variables (Section OLS results). The OLS results is based on the robust standard error to correct for potential heteroscedasticity, and the coefficient, p-value, and t-value are used to quantify the effect size and statistical significance of the independent variables, while the Koenker (BP) test is employed to verify the homoscedasticity assumption of the residuals.

GWR and MGWR

GWR is a local spatial modeling technique that generates location-specific regression models across the study area. Unlike global methods such as OLS, GWR captures spatial heterogeneity and local variations in relationships between variables, thereby offering a more realistic representation of spatial processes (Section GWR results).

GWR incorporates local contextual effects of spatial objects, improving model accuracy for spatial data. However, it assigns the same BW (spatial scale) to all explanatory variables when defining spatial weights. MGWR relaxes this constraint by allowing each variable to have its own BW. This flexibility enables MGWR to model relationships at varying spatial scales, better reflecting the unique spatial influence of each explanatory variable and leading to more precise local coefficient estimates (Section MGWR results).

In these two models, the coefficients, calculated as the arithmetic average of local coefficients across all spatial units, reflected the overall average influence trend of the independent variable on the dependent variable. The Max. and Min. values of the coefficients jointly defined the spatial fluctuation range of the influence intensity of the independent variable, intuitively presenting the extreme differences in local effects. STD quantified the degree of dispersion of local coefficients around the mean and serves as a core indicator for measuring the intensity of spatial heterogeneity. A larger STD value indicated more significant spatial variation in the influence of the independent variable. The optimal BW, determined by minimizing the AICc, had a value that reflects the spatial scale characteristics of the independent variable’s influence.

Additional resources

MGWR Learning and Discussion Community, https://www.csdn.net/.

The Ethics Protocol approval number, XHEC-C-2016-016.

Published: November 29, 2025