The urban physical exposome and leisure-time physical activity in early midlife: a FinnTwin12 study
===================================================================================================

* Zhiyang Wang
* Sari Aaltonen
* Roos Teeuwen
* Vasileios Milias
* Carmen Peuters
* Bruno Raimbault
* Teemu Palviainen
* Erin Lumpe
* Danielle Dick
* Jessica E. Salvatore
* Maria Foraster
* Payam Dadvand
* Jordi Júlvez
* Achilleas Psyllidis
* Irene van Kamp
* Jaakko Kaprio

## Abstract

Leisure-time physical activity is beneficial for health and is associated with various urban characteristics. Using the exposome framework, the totality of the environment, this study investigated how urban physical environments were associated with leisure-time physical activity during early midlife. A total of 394 participants (mean age: 37, range 34-40) were included from the FinnTwin12 cohort residing in five major Finnish cities in 2020. We comprehensively curated 145 urban physical exposures at residential addresses of participants and measured three leisure-time physical activity measures: (1) total leisure-time physical activity (total LTPA) and its sub-domains (2) leisure-time physical activity without commuting activity (LTPA) and (3) commuting activity. Using K-prototypes cluster analysis, we identified three urban clusters: “original city center,” “new city center,” and “suburban”. Results from adjusted linear regression models showed that participants in the “suburban” cluster had lower levels of total LTPA (beta: -0.13, 95% CI: -0.23, -0.03) and LTPA (beta: -0.17, 95% CI: -0.28, -0.05), compared to those in the “original city center” cluster. The eXtreme Gradient Boosting models ranked exposures related to greenspaces, pocket parks, and road junctions as the top important factors influencing outcomes, and their relationships with outcomes were largely non-linear. More road junctions and more pocket parks correlated with higher total LTPA and LTPA. When the all-year normalized difference vegetation index within a 500 m buffer fell below 0.4, it correlated with higher levels of total LTPA, whereas above 0.4, it correlated with lower levels. To conclude, our findings revealed a positive correlation between urbanicity and physical activity in Finnish cities and decomposed this complexity into crucial determinants. Importance rankings and nonlinear patterns offer valuable insights for future policies and projects targeting physical inactivity.

Keywords
*   exercise
*   machine learning
*   urbanization

## 1 Introduction

Regular physical activity has been widely demonstrated to prevent multiple non-communicable diseases and reduce the risk of premature death1. The economic and health burden arising from physical inactivity is substantial and continually rising, costing public health care systems an estimated USD 47.6 billion globally every year2. Since previous studies show a strong contribution of environmental factors to physical activity3–5, interventions targeting the environment may be a good entry point to promote physical activity.

Urbanization stands as a transformative trend, with more than half the world’s population currently residing in urban areas6. Many reviews have summarized the salient link between the urban environment and physical activity7–9. The exposome offers a theoretical framework with an umbrella perspective to depict the totality of the environment that people experience10 and examines health effects from the real-world urban environment, of which the urban physical component plays an important role. The exposome studies have the potential to unveil more comprehensive non-genetic predictors through large-scale characterization of the environment. Gorman et al. have outlined the bidirectional effects between the exposome and physical activity but pointed out the uncertainty in mechanism and interactions11. The urban physical exposome is ubiquitous and multifaceted, which makes it a complex entity to study.

Every environmental factor contributes to this complex totality of exposures, and no factor is isolated. Urban regeneration projects are a good example, usually designed to improve public health by implementing structural and risk-minimizing solutions. They often yield collateral effects on other aspects, such as bringing economic, social, and cultural benefits, within the city’s complex system12–14. For example, an urban riverside park regeneration project in Barcelona, Spain was estimated to attract over five thousand adult users daily to perform different types of physical activity15. Beyond the project’s basic objectives, an open-air museum will be built there, transforming social and built environments. Nowadays, regeneration projects around the world are often multi-component and intersectoral. In another Barcelona regeneration program, aiming to improve living conditions in the most disadvantaged neighborhoods (involving, for example, social services, green spaces, and household support), researchers found that the neighborhood with a bigger project budget was associated with a higher frequency of physical activity among residents16. Previous studies relying on single exposures or limited sets were relatively inadequate to depict the broader urban environment and its health effects.

In this study, we aimed to comprehensively study the impact of the urban physical component of the exposome on the level of leisure-time physical activity during early midlife through two objectives: 1) clustering people with heterogeneous urban environments and comparing physical activity levels between clusters, as well as sex-specific effect and 2) ranking urban physical exposures by importance on leisure-time physical activity, examining non-linear relationships, and detecting pairwise interactions between exposures.

## 2 Material and methods

The flow chart of this study is presented in Figure 1.

![Figure 1](http://medrxiv.org/https://www.medrxiv.org/content/medrxiv/early/2024/06/10/2024.06.09.24308658/F1.medium.gif)

[Figure 1](http://medrxiv.org/content/early/2024/06/10/2024.06.09.24308658/F1)

Figure 1 
Study flow

### 2.1 Participants

Participants were from the FinnTwin12 cohort, which is a nationwide prospective cohort of all Finnish twins born between 1983 and 1987. Briefly, at baseline (1994-1999), 5522 12-year-old twins were invited to participate and 87% of them agreed to take part. There were four follow-ups: age 14, age 17, young adulthood (mean age 24), and early midlife (mean age 37), with retention rates of 92%, 75%, 66%, and 41%, respectively. A recent study has detailed the latest follow-up of the cohort17. In this study, we included individual twins who lived in five large cities of Finland, namely Helsinki, Tampere, Espoo, Oulu, or Jyväskylä, in 2020. A

### 2.2 Measures

#### 2.2.1 Leisure-time physical activity

Our study focuses on early midlife leisure-time physical activity, which is performed at the person’s discretion along with essential daily living activity or work-related tasks18. This type of physical activity is considered one of the most effective ways to increase overall physical activity levels19. It was measured through structured and validated questions on the frequency, mean duration, and mean intensity of participants’ leisure-time physical activity sessions, as well as a question on their commuting activity20,21. Based on these structured questions, we quantified mean metabolic equivalent of task (MET) hours per day, which expresses the energy cost of physical activities in the form of the resting metabolic rate22. Its calculation formula was the following: physical activity frequency × physical activity duration × physical activity intensity23. The MET values for activity intensity were: 4 for intensity corresponding to walking, 6 for intensity corresponding to vigorous walking to jogging, 10 for intensity corresponding to jogging, and 13 for the intensity corresponding to running. All types of leisure time physical activities were considered when MET hours per day were calculated. We assumed that commuting activity was done on 5 days per week and on the intensity of walking. The questions are listed in Supplemental Note 1.

The primary measure, *total leisure-time physical activity (total LTPA)*, was the sum of leisure-time and commuting-related physical activities. These two sub-domains were secondary measures: 1) *leisure-time physical activity without commuting activity (LTPA)* and 2) *commuting activity*. Participants with over mean 45 MET hours/day of total LTPA were identified as outliers and removed. This threshold corresponds to, for example, approximately 3.5 hours of fast running daily, which is likely unrealistic24. The distributions of all three measures are shown in Supplemental Figure 1, and due to the skewness, we log-transformed them.

#### 2.2.2 Urban physical exposome

We assigned 145 indicators of urban physical exposures to the residential address of each study participant. Detailed description and summary statistics of these indicators are presented in Supplemental Table 1. The urban physical exposome set comprehensively depicted the urban environment including aspects such as traffic, streets, land use, green (i.e. parks, forests, and fields) and blue (i.e. lakes and seas) spaces, and so on. The computing and enriching process was on the geocode level and derived from multiple open sources, which is described in Supplemental Note 2 and elsewhere25–28. Most urban physical exposures were measured or modelled in 2018 and 2023, and the percentage of area covered by trees was measured in 2015. We used the residential history provided by the Digital and Population Data Services Agency, Finland between birth and 2021 to merge the urban physical exposures by EUREF-FIN geocodes. Exposures available in 2018 or 2015 were merged with residential addresses of participants in 2018 or 2015, while exposures available in 2023 were merged with residential addresses in 2020.

#### 2.2.3 Other measures

Five sociodemographic variables were identified a priori: sex (categorical, female vs. male), age (continuous, year), work (categorical, not working or other situation vs. currently work), education (categorical, post-secondary or lower vs. bachelor/equivalent or above), and marital status (categorical, married, steady relationship, or living together vs. no). The latter three were self-reported at the early midlife follow-up. Sex was based on the register information obtained when the cohort was established, while age was computed from the difference between the date of response and the date of birth. There were another three behavioral variables: illicit substance use (categorical, never vs. at least once), ever smoker (smoked over 100 cigarettes lifetime) (categorical, no vs. yes), and alcohol drinking (categorical, monthly or less or even never vs. 2-4 times a month or more), inquired also at the early midlife follow-up. Adult leisure-time physical activity was associated to most of the sociodemographic and behavioral variables, as shown in previous research29–31.

To depict the social environment, four neighborhood social variables at the postal code level were derived from Statistics Finland in 2018: the proportion of resident living alone (single household), of residents with the lowest education level, of residents with the lowest income quartile, and of unemployed residents. A neighborhood deprivation score was generated from the latter three social variables32. We first standardized the three variables to z-scores, and their mean value is the deprivation score. Using a median split, we then categorized neighborhoods where participants lived in 2018 into two levels: low- and high-deprived. Thus, there were two neighborhood social variables: the proportion of resident living alone and deprivation level, which were merged via residential history in 2018 too.

### 2.3 Analysis

#### 2.3.1 Data processing

After excluding those people who did not have information on leisure-time physical activity, sociodemographic, behavioral, and neighborhood-level social variables, 394 twin individuals resident in these urban areas were included in this study. Given that there were only 44 twin pairs with both cotwins satisfying the inclusion criteria, we did not consider zygosity as a covariate and did not perform any pairwise twin analysis. The distribution of sociodemographic and behavioral variables among included and excluded participants are presented in Supplemental Table 2, respectively. There were significant differences between included and excluded participants in education, illicit substance use, and alcohol drinking.

#### 2.3.2 Clustering analysis

The k-prototypes cluster analysis was employed to distinguish distinct patterns in the urban environment. It combines dissimilarity measures from both k-means and -modes algorithms for mixed types of exposures, and has shown to have a good performance33,34. Continuous exposures were standardized by standard deviation (SD). All 145 urban physical exposures were included in the clustering algorithms. The Silhouette method was used to pre-specify the number of clusters35. One-step imputation within the algorithm was applied for missing values36. Since k-prototypes cluster analysis is sensitive to outliers, the principal component analysis of mixed data was conducted before. Participants whose first or second principal components fell outside the range of five standard deviations were identified as outliers37, as a practical way, and excluded from the cluster analysis; three participants were excluded.

Next, hierarchical linear regression was performed for the relationship between the urban cluster and leisure-time physical activity measures with three adjustment plans for covariates: 1) sociodemographic variables, 2) sociodemographic and behavioral variables, and 3) sociodemographic, behavioral, and neighborhood social variables. The cluster effect of sampling based on families of twin pairs was controlled by the robust standard error. We also performed the sex-stratified analysis.

#### 2.3.3 Machine learning analysis

Before exploring the complexity within the urban environment via a pluralistic analysis platform, generalized linear regression models with the robust standard error were repeatedly performed between each leisure-time physical activity measure (total LTPA, LTPA, and commuting activity) and each urban physical exposure (missing values were imputed). The *a priori* significant threshold of 0.01 was used to identify noteworthy candidates. Dimensional reduction increases the model stability of subsequent analysis.

Then, we performed the eXtreme Gradient Boosting (XGBoost) model to assess the importance of urban physical exposures on each physical activity measure, uncover interactions, and identify nonlinear relationships38. It is an optimized distributed gradient boosting library designed for efficient and scalable training of machine learning models, with gradient-boosted decision trees algorithm38. The hyperparameters were tuned through the 5-fold cross-validation grid search39. The participants were randomly split into training and testing subsets in a ratio of 3:1. The model performance was evaluated by root-mean-square error (RMSE). All urban physical exposures, sociodemographic, behavioral, and neighborhood social variables were included in the model. After hyperparameter tuning, the model was repeated two additional times with different seeds for result robustness.

To increase model transparency, the SHapley Additive exPlanations (SHAP) value was used to interpret and visualize the results from the XGboost model, which features the exposures’ importance on the outcome based on the cooperative game theory40. Its direction suggests the direction of impact on prediction, leading the model to predict either a higher or lower value of outcomes. Its magnitude is a measure of how strong the effect is. We quantified pairwise interaction SHAP values between included variables and summed their absolute value of all participants, with a high value indicating a strong interaction and synergistic effect41. Additionally, Group-Lasso INTERaction-NET was performed for interaction to compare with the XGBoost’s result42.

#### 2.3.4 Sensitivity analysis

Due to missing values in urban physical exposures, we additionally performed sensitivity analyses of K-prototype cluster analysis and repeated generalized linear regression models between each urban physical exposure and each leisure-time physical activity measure, after removing participants with missing values (n=13).

## 3 Results

### 3.1 Description of participants

Of the 394 included participants (mean age: 37, SD: 1.5) (Table 1), more individuals were female (55%). Altogether, 87%, 79%, and 75% of participants were employed, had at least bachelor-level education, and were married or in a stable relationship, respectively. In their early midlife, more than half of the participants drank alcohol at least 2-4 times a month (58%), but fewer had smoked over 100 cigarettes (45%) or had used illicit substances such as marijuana at least once (48%). Before log-transformation, the means of total LTPA, LTPA, and commuting activity (unit: MET hours/day) were 5.4 (SD: 4.7), 4.3 (SD: 4.4), and 1.1 (SD: 1.0), respectively. After log-transformation, Spearman correlations between total LTPA and LTPA, between total LTPA and commuting activity, and between LTPA and commuting activity were 0.9, 0.3, and 0.1, respectively

View this table:
[Table 1:](http://medrxiv.org/content/early/2024/06/10/2024.06.09.24308658/T1)

Table 1: Characteristics of sociodemographic, behavior, and neighborhood social variables (participants n=394)

### 3.2 Results from clustering and hierarchical regression

The Silhouette method identified the optimal number of clusters to be three (largest Silhouette index, total within-cluster sum of squares: 42323.69). Using the map of Helsinki and Espoo and the spatial layer of centers and shopping areas in 2019 from the community structure monitoring system, Finnish Environment Institute43, we visually classified Cluster 1, 2, and 3 as the “Original city center”, “New city center”, and “Suburban” clusters, respectively, based on the participants’ residence in 2018, as the urban cluster variable (Figure 2).

![Figure 2:](http://medrxiv.org/https://www.medrxiv.org/content/medrxiv/early/2024/06/10/2024.06.09.24308658/F2.medium.gif)

[Figure 2:](http://medrxiv.org/content/early/2024/06/10/2024.06.09.24308658/F2)

Figure 2: 
Twin participants’ residence in the Helsinki and Espoo area in 2018 colored by cluster. Note: The gray layer shows centers and shopping areas in 2019.

After fully adjusting for sociodemographic, behavioral, and neighborhood social variables, compared to participants who lived in the “original city center” cluster, participants who lived in the “suburban” cluster were associated with significantly lower log-transformed scores of total LTPA (beta: -0.13, 95% CI: -0.23, - 0.03) and LTPA (beta: -0.17, 95% CI: -0.28, -0.05) (Table 2). The effect sizes did not change substantially after adjustment of sociodemographic variables only and adjustment of both sociodemographic and behavioral variables. Regardless of adjustment plans, there was no significant association between the urban cluster and commuting activity (Table 2). There was no significant difference in any outcome between participants who lived in the “suburban” and “new city center” clusters. The powers of full-adjusted models of total LTPA, LTPA, and commuting activity were all 1.0.

View this table:
[Table 2:](http://medrxiv.org/content/early/2024/06/10/2024.06.09.24308658/T2)

Table 2: Results of the linear regression between the urban cluster and physical activity measures

After stratifying the analyses based on sex, we observed that, in males, the result pattern and effect sizes were like the overall results between the urban cluster and total LTPA, while the association between the urban cluster and LTPA became null after full adjustment (Supplemental Table 3). However, in females, after additionally adjusting for behavioral variables only or for both behavioral and neighborhood social variables, no significant association of the urban cluster with total LTPA and LTPA was seen (Supplemental Table 3).

### 3.3 Results from XGBoost

Based on the repeated generalized linear regression, there were 25 urban physical exposures significantly associated with total LTPA and 24 with LTPA (Supplemental Table 4). No urban physical exposure met the threshold p-value of 0.01 for association with commuting activity (Supplemental Table 4), so there was no XGBoost analysis for it.

In the XGBoost model of total LTPA including all urban physical exposures, sociodemographic, behavioral, and neighborhood social variables, the top three important urban physical exposures were the count of any type of road junctions within a 500 m buffer (ints\_500), the total area of all interconnected pocket parks within an 800 m walking distance (sumarea\_pocketparks\_800), and the 5-years moving average of Normalized Difference Vegetation Index (NDVI), an indicator of general greenness, within a 500 m buffer around the home during whole year (ndvi_5yrs_all_500) (Figure 3A). In dependence plots, SHAP values positively correlated with both the count of any type of road junctions within a 500 m buffer (Figure 3B) and the total area of all interconnected pocket parks within an 800 m walking distance (Figure 3C). When the two urban physical exposures were within a certain range, SHAP values remained constant, which this type of non-linearity made these two exposures look like threshold variables. When the count of any type of road junctions within a 500 m buffer was in the range of 1-40, 40-50, and over 50, the SHAP values were approximately -0.003, 0, and 0.003, respectively. When the total area of all interconnected pocket parks within an 800 m walking distance was in the range of 0-0.005, 0.005-0.01, and over 0.01 km2, the SHAP values were approximately -0.002, 0.001, and 0.004, respectively. In Figure 3D, the 5-years moving average of NDVI within a 500 m buffer during whole year also showed a pattern of a binary threshold variable. When it was below 0.4, the SHAP value was around 0.0012. When it was over 0.4, the SHAP value was around -0.0012.

![Figure 3:](http://medrxiv.org/https://www.medrxiv.org/content/medrxiv/early/2024/06/10/2024.06.09.24308658/F3.medium.gif)

[Figure 3:](http://medrxiv.org/content/early/2024/06/10/2024.06.09.24308658/F3)

Figure 3: 
Results of XGBoost models for total leisure-time physical activity (total LTPA) and leisure-time physical activity without commuting activity (LTPA)

Note: The SHAP bar plots show the influence of each variable: total LTPA (a) and LTPA (e). The SHAP dependence plots show how a single individual influences the XGboost prediction on total LTPA (b, c, d) and LTPA (f, g, h). ints\_500 is the count of any type of road junctions within a 500 m buffer;sumarea\_pocketparks\_800 is the total area of all interconnected pocket parks within an 800 m walking distance; ndvi\_5yrs\_all_500 is the 5-years moving average of Normalized Difference Vegetation Index within a 500 m buffer during whole year; count_pocketparks_800 is the count of pocket parks within an 800 m walking distance. Abbreviation: leisure-time physical activity (LTPA); SHapley Additive exPlanations (SHAP)

In the XGBoost model of LTPA (Figure 3E), the most important urban physical exposures were the count of pocket parks within an 800 m walking distance (count_pocketparks_800), the total area of all interconnected pocket parks within an 800 m walking distance remained the second most important, and the count of any type of road junctions within a 500 m buffer became the third most important. The SHAP value revealed a switch in predictions from lower (negative SHAP values) to higher (positive SHAP values) log-transformed LTPA when the number (count) of pocket parks within an 800 m walking distance was more than two (Figure 3F). Interestingly, each count led to two different but close SHAP values. The patterns of the total area of all interconnected pocket parks within an 800 m walking distance (Figure 3G) and the count of any type of road junctions within a 500 m buffer (Figure 3H) were similar to the model of total LTPA.

Supplemental Figure 2 displays pairwise SHAP interaction values in the XGBoost model of total LTPA, and there was no pairwise interaction between urban physical exposures. Similarly, the XGBoost model of LTPA (Supplemental Figure 3) indicates slightly some interactions but with very low values (<0.001). Group-Lasso INTERaction-NET models also did not capture any strong pairwise interaction for the two physical activity measure analyses.

For the XGBoost model of total LTPA, the RMSE is 0.27 in the training subset and 0.29 in the testing subset. Comparing the reported test with two extra tests, the importance rank varied, but the count of any type of road junctions within a 500 m buffer was always the most or second most important (Supplemental Table 5). For the XGBoost model of LTPA, the RMSE is 0.32 in both training and testing subsets. The importance rank also varied between reported results and two extra tests, but the most important urban physical exposure was the count of pocket parks within an 800 m walking distance, being always in the top two (Supplemental Table 6).

### 3.4 Sensitivity analysis for missing value in urban physical exposures

After excluding 13 participants with missing values in some urban physical exposures, the Silhouette method identified two clusters. In the following fully adjusted linear regression models, no significant differences in any of the physical activity measures were found between the clusters. Repeated generalized linear regression analyses revealed 25 urban physical exposures significantly associated with total LTPA, consistent with the analysis using imputed data, and 26 exposures with LTPA, two more than the analysis with imputed data. Still, no urban physical exposure reached the 0.01 P-value threshold for association with commuting activity.

## 4 Discussion

We used clustering analysis and XGBoost to simultaneously and comprehensively study the effect of 145 urban physical exposures on leisure-time physical activity in 394 Finnish adults in their early midlife. Three clusters were identified: “original city center”, “new city center”, and “suburban”. We found people living in suburban areas had a lower level of physical activity in leisure time compared to those living in the original city center. There was no difference between “original city center” and “new city center” clusters. The effects appeared more clearly in males, while behavioral and neighborhood social factors may account for the associations in females. XGBoost models revealed a complex relationship between the urban physical exposome and leisure-time physical activities, in which important exposures showed non-linearity and looked like threshold variables. Increased road junctions and more and bigger pocket parks correlated with higher levels of leisure-time physical activity. However, higher amounts of vegetation greenness (indicated by NDVI) were associated with low leisure-time physical activity levels. We did not find any considerable interaction between urban physical exposures contributing to leisure-time physical activities.

Previous research has documented the relationship between the degree of urbanization and physical activity but with inconsistent findings regarding the direction of effects. A cross-sectional study in Shanghai, China with 327 respondents (mean age: 40) similarly reported higher leisure-time physical activity among downtown residents compared to suburban dwellers, but in contrast to our results, significant results were also found for transportation activities44. Another Canadian study showed that the physical activity level was higher in urban than in suburban among adolescents from schools in lower socio-economic areas45. Nevertheless, a systematic review suggested that children and teenagers who live in suburban areas were more physically active than in rural and urban areas46, and, similar to the Shanghai study above, a nationwide study in China showed that rising urbanization correlates with longer commuting times among adults (mean age: 45)47. Sex-specific effects have also been also observed. In the US, only male adolescents living in urban areas engaged in more moderate-to-vigorous physical activity (MVPA) than those living in suburban areas48. Additionally, distinct patterns between sexes in the significance and direction of associations between urbanity level in different aspects and physical activity measures were noted in Mexico49. Socioeconomic status (SES) might explain the sex difference, as the association weakened to null in females after adjusting for behavioral and neighborhood social variables. Previous population studies have observed some interaction effects between sex and SES on physical activity50–52. The inconsistency between literature and our findings may be due to different population characteristics, sports cultures, country contexts, urban planning, or urbanicity definitions. Instead of a pre-definition of (sub)urban areas by governmental guidelines, we used an unsupervised data-driven clustering method to determine heterogeneous urban environments within urban areas reflecting real-life exposure modes and accounting for correlation, additive, and mixture effects53.

XGBoost models ranked the elements of pocket parks, road junctions, and greenspaces as strongly associated with leisure-time physical activities among early midlife adults. A natural experimental study in low-income American neighborhoods found increased leisure-time exercise among middle-aged residents after pocket parks were constructed54. Users of pocket parks, defined as living within a 0.5 mile (∼800 m) radius, had higher exercise levels than traditional park users54. Researchers further summarized that pocket parks were cost-effective for promoting physical activity in inner-city areas54. A study in Chongqing, China, utilizing interviews on conceptual understanding of park images, revealed that the environmental characteristics of pocket parks contributed to a restorative effect involving entertainment activities and relief55. Noteworthy, a recent Chinese study using Light Gradient-Boosting Machine model found that recreational facilities were the most important factor for walking behavior in old adults but the number of parks was the least important among 11 factors, highlighting the specific effect driven by the content inside parks or recreation areas56. For road junctions, a Finnish study found the density of intersections, defined as the junction of a minimum of three roads, was positively associated with the number of physical activity bouts and the level of moderate to vigorous physical activity among older adults57. Zang et al. used random forest models to identify the intersection density, as well as streetscape greenery, as the most important physical exposure contributing to light physical activity among older adults58. More intersections usually indicate a greater degree of connectivity, which creates a more convenient environment for people to walk or bike to their destinations. However, the relationship between street connectivity, involving the number of intersections, and physical activity in all age groups of adults varied across different buffer areas in urban environments, suggesting the complexity of urban living environments59. Where the association of greenspace with physical activity is relatively inconsistent60, our findings show an association in which surrounding greenness is positively associated with LTPA up until a threshold of 0.4 NDVI, with higher NDVI relating to lower LTPA. High levels of green space might reflect suburban living to some extent, and other greenspace indicators, such as accessibility, were not prominent. The relationship between greenspace and physical activity could be moderated by the level of urbanization60. Other studies have similar findings on the threshold effect. For example, the positive association of physical activity with multiple green space uses indicators reached to peak when indicators were within a 600 m buffer61. Zang et al. also found that streetscape greenery had a positive effect on light physical activity when it ranged from 0.12 to 0.15 point, corresponding to a low level of visible greenery58. Besides, another Chinese study also identified the 0.4 NDVI, corresponding to areas with sparse to moderate vegetation, as the turning point for its association with self-rated health among the old population62, and self-rated health closely correlated with physical activity63. This annotation added to current evidence has critical guidance on urban planning.

We selected different methods to depict the contour of association between the urban physical exposome and leisure-time physical activity, translating abstract characteristics into practical understanding. On one hand, clustering analysis has the advantage of providing insight into real-world scenarios and holding a high scalability to uncover hidden patterns. On the other hand, tree-based machine learning can be applied as a pluralistic analysis platform to synthesize evidence between a range of urban physical exposures and physical activity58,64,65. Comparing to conventional analyses, the XGBoost model enhances our assessments with several advantages: 1) unraveling nonlinear relationships through visualization, 2) disentangling complex interactions among multiple exposures, and 3) offering robust computation for multi-inference approaches66. By deepening the understanding of distinct and complex characteristics of the urban physical exposome, supported by detailed exposure profiling, policymakers can develop precise and cost-effective interventions and strategies to address the challenge of low physical activity levels.

Besides its strength, this study is not without limitations. First, the sample size was relatively small compared to other exposome studies. Although the sample size for K-prototype clustering (over 10 times the number of clusters) and subsequent regression seems to be adequate (but not for the sex-stratified analysis), inconsistency in additional tests highlights the need for a larger sample. Additionally, due to the complexity of the large-dimensional exposome set, the modest sample size made capturing relatively small interactions more challenging. Therefore, we should be cautious when interpreting results. Second, only participants from the five largest cities in Finland were included, limiting the generalizability. Besides, we did not include any participants living in rural areas. Not only the physical environment, but lifestyles may also differ between urban and rural areas. Therefore, the interpretation should be narrowed down to specific types of cities. Third, urban physical exposures were based on residential addresses, which overlook dynamic human behaviors outside the home, leading to measurement errors. In addition, the used residential geocodes corresponded to participants’ residences in 2017, 2018, or 2020, without accounting for how long they lived at those addresses. Measurement errors could skew our identification of key determinants, as exposures with larger errors might show weaker associations and be classified as less influential, even if they are actually more important than those identified as most influential. More granular and accurate estimations of exposure and behavior could facilitate the exploration in the dynamic interaction between the environment and human behavior12. Fourth, some exposures were available in 2023 but merged with the address in 2020, posing a temporality issue. The relatively slow urban renewal and construction in Finland reduced the concern67. Fifth, missing values in exposures may introduce bias. Excluding participants with missing values altered the optimal number of clusters, while the number of significant associations between exposures and outcomes remained similar to the number based on imputed data. Given that only about 3% of participants had missing values, the effect is likely modest, but caution is still warranted. Sixth, leisure-time physical activity was self-reported. The device-based measurement of leisure-time physical activity would have been more accurate. However, the validity of leisure-time physical activity questions used in Finnish twins has been demonstrated20,21.

## 5 Conclusion

This study employed two analytical approaches to explore the intricate impact of the urban physical exposome on leisure-time physical activity in early midlife in Finland. Clustering analysis revealed three heterogeneous patterns of urban environments. Living in suburban areas was associated with lower levels of leisure-time physical activity than in original city center areas. XGBoost models identified pocket parks, road junctions, and greenspaces as influential factors with non-linear relationships, which behaved like threshold variables. Given limitations in sample size, generalizability, and measurement granularity, we call for further studies in other settings to replicate our analyses. We still advocate presenting the evidence to stakeholders and policymakers to develop tailored interventions on some urban features to achieve higher cost-effectiveness by focusing on the most influential determinants and their optimal ranges in addressing the challenge of the physically inactive lifestyle in our rapidly urbanizing world.

## Supporting information

Supplementary Note 1-2, Figure 1-3 and Table 1-6 [[supplements/308658_file05.docx]](pending:yes)

## Data Availability

The FinnTwin12 data are not publicly available due to the restrictions of informed consent. However, the FinnTwin12 data are available through the Institute for Molecular Medicine Finland (FIMM) Data Access Committee (DAC) (fimm-dac@helsinki.fi) for authorized researchers who have IRB/ethics approval and an institutionally approved study plan. To ensure the protection of privacy and compliance with national data protection legislation, a data use/transfer agreement is needed, the content and specific clauses of which will depend on the nature of the requested data. Requests will be addressed in a reasonable time frame (generally two to three weeks), and the primary mode of data access is by either personal visit or remote access to a secure server.

## Author’s contribution

M.F., P.D., J.J., A.P., I.v.K., and J.K. conceived the exposome framework. Z.W. developed the research question and designed the analysis and other authors commented to refine it. S.A., D.D., J.S., and J.K. led the FinnTwin12 cohort. R.T., V.M., B.R., and M.F. enriched urban physical exposures and T.P. managed the FinnTwin12 data. Z.W. performed the analysis and wrote the original draft. All authors reviewed the draft and approved for the submission.

## Competing interests

The authors declare that they have no competing interests.

## Ethical requirement

The ethics committee of the Department of Public Health of the University of Helsinki (Helsinki, Finland) and the Institutional Review Board of Indiana University (Bloomington, Indiana, USA) approved the FinnTwin12 study protocol from the start of the cohort. The ethical approval of the ethics committee of the Helsinki University Central Hospital District (HUS) is the most recent and covers the most recent data collection (early midlife) (HUS/2226/2021, dated September 22, 2021). All participants and their parents/legal guardians gave informed written consent to participate in the study. The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975, as revised in 2008.

## Data and Code Availability

The FinnTwin12 data are not publicly available due to the restrictions of informed consent. However, the FinnTwin12 data are available through the Institute for Molecular Medicine Finland (FIMM) Data Access Committee (DAC) (fimm-dac@helsinki.fi) for authorized researchers who have IRB/ethics approval and an institutionally approved study plan. To ensure the protection of privacy and compliance with national data protection legislation, a data use/transfer agreement is needed, the content and specific clauses of which will depend on the nature of the requested data. Requests will be addressed in a reasonable time frame (generally two to three weeks), and the primary mode of data access is by either personal visit or remote access to a secure server.

Code for major analyses is available at [https://github.com/doge73/city\_urban\_PA](https://github.com/doge73/city_urban_PA).

## Acknowledgments

This research was partly funded by the European Union’s Horizon 2020 research and innovation program under grant agreement No 874724 (Equal-Life). Equal-Life is part of the European Human Exposome Network. Data collection in FinnTwin12 has been supported by the National Institute on Alcohol Abuse and Alcoholism (grants AA-12502, AA-00145, and AA-09203 to Richard J. Rose, and AA015416 to Danielle Dick and Jessica Salvatore) and the Academy of Finland (grants 100499, 205585, 118555, 141054, 264146, 308248, 312073, 336823, and 352792 to Jaakko Kaprio). Jaakko Kaprio acknowledges support by the Academy of Finland (grants 265240, 263278). ISGlobal acknowledges support from the grant CEX2018-000806-S funded by MCIN/AEI/10.13039/501100011033, and support from the Generalitat de Catalunya through the CERCA Program.

*   Received June 9, 2024.
*   Revision received June 9, 2024.
*   Accepted June 10, 2024.


*   © 2024, Posted by Cold Spring Harbor Laboratory

This pre-print is available under a Creative Commons License (Attribution-NonCommercial 4.0 International), CC BY-NC 4.0, as described at [http://creativecommons.org/licenses/by-nc/4.0/](http://creativecommons.org/licenses/by-nc/4.0/)

## Reference

1.  1.Booth FW, Roberts CK, Laye MJ. Lack of Exercise Is a Major Cause of Chronic Diseases. In: Comprehensive Physiology. 2012.p.1143–211. Available from: doi:10.1002/cphy.c110025 https://doi.org/10.1002/cphy.c110025.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1002/cphy.c110025&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=23798298&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F06%2F10%2F2024.06.09.24308658.atom) 

2.  2.Santos AC, Willumsen J, Meheus F, Ilbawi A, Bull FC. The cost of inaction on physical inactivity to public health-care systems: a population-attributable fraction analysis. Lancet Glob Heal 2023;11:e32–9. doi:10.1016/S2214-109X(22)00464-8.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/S2214-109X(22)00464-8&link_type=DOI) 

3.  3.Duncan GE, Goldberg J, Noonan C, Moudon AV, Hurvitz P, Buchwald D. Unique Environmental Effects on Physical Activity Participation: A Twin Study. PLoS One 2008;3:e2019.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1371/journal.pone.0002019&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=18414678&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F06%2F10%2F2024.06.09.24308658.atom) 

4.  4.Boomsma DI, Cherkas L, Cornes BK, Harris JR, Kaprio J, Kujala UM, et al. Variance Components Models for Physical Activity With Age as Modifier: A Comparative Twin Study in Seven Countries. Twin Res Hum Genet 2012/02/21. 2011;14:25–34. https://doi.org/DOI: 10.1375/twin.14.1.25.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1375/twin.14.1.25&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=21314253&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F06%2F10%2F2024.06.09.24308658.atom) 

5.  5.Carlin A, Perchoux C, Puggina A, Aleksovska K, Buck C, Burns C, et al. A life course examination of the physical environmental determinants of physical activity behaviour: A “Determinants of Diet and Physical Activity” (DEDIPAC) umbrella systematic literature review. PLoS One 2017;12:e0182083.
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=http://www.n&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F06%2F10%2F2024.06.09.24308658.atom) 

6.  6.Programme UNHS. World Cities Report 2022 [Internet]. United Nations, 2022 [cited 2024 Jan 23]. (World Cities Report). Available from: [https://www.un-ilibrary.org/content/books/9789210028592](https://www.un-ilibrary.org/content/books/9789210028592) doi:10.18356/9789210028592.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.18356/9789210028592&link_type=DOI) 

7.  7.Durand CP, Andalib M, Dunton GF, Wolch J, Pentz MA. A systematic review of built environment factors related to physical activity and obesity risk: implications for smart growth urban planning. Obes Rev 2011;12:e173–82. doi:10.1111/j.1467-789X.2010.00826.x.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1111/j.1467-789X.2010.00826.x&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=21348918&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F06%2F10%2F2024.06.09.24308658.atom) 

8.  8.Ding D, Sallis JF, Kerr J, Lee S, Rosenberg DE. Neighborhood Environment and Physical Activity Among Youth: A Review. Am J Prev Med 2011;41:442–55. doi:10.1016/j.amepre.2011.06.036.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/j.amepre.2011.06.036&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=21961474&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F06%2F10%2F2024.06.09.24308658.atom) 

9.  9.Kärmeniemi M, Lankila T, Ikäheimo T, Koivumaa-Honkanen H, Korpelainen R. The Built Environment as a Determinant of Physical Activity: A Systematic Review of Longitudinal Studies and Natural Experiments. Ann Behav Med 2018;52:239–51. doi:10.1093/abm/kax043.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1093/abm/kax043&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=http://www.n&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F06%2F10%2F2024.06.09.24308658.atom) 

10. 10.Wild CP. The exposome: from concept to utility. Int J Epidemiol 2012;41:24–32. doi:10.1093/ije/dyr236.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1093/ije/dyr236&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=22296988&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F06%2F10%2F2024.06.09.24308658.atom) 
    
    [Web of Science](http://medrxiv.org/lookup/external-ref?access_num=000302026800007&link_type=ISI) 

11. 11.Gorman S, Larcombe AN, Christian HE. Exposomes and metabolic health through a physical activity lens: a narrative review. J Endocrinol 2021;249:R25–41. doi:10.1530/JOE-20-0487.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1530/JOE-20-0487&link_type=DOI) 

12. 12.Sonnenschein T, Scheider S, de Wit GA, Tonne CC, Vermeulen R. Agent-based modeling of urban exposome interventions: prospects, model architectures, and methodological challenges. Exposome 2022;2:osac009. doi:10.1093/exposome/osac009.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1093/exposome/osac009&link_type=DOI) 

13. 13.Wang H, Liu N, Chen J, Guo S. The Relationship Between Urban Renewal and the Built Environment: A Systematic Review and Bibliometric Analysis. J Plan Lit 2021;37:293–308. doi:10.1177/08854122211058909.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1177/08854122211058909&link_type=DOI) 

14. 14.Chen Y, Liu G, Zhuang T. Evaluating the Comprehensive Benefit of Urban Renewal Projects on the Area Scale: An Integrated Method. Vol. 20, International Journal of Environmental Research and Public Health. 2023. doi:10.3390/ijerph20010606.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.3390/ijerph20010606&link_type=DOI) 

15. 15.Vert C, Nieuwenhuijsen M, Gascon M, Grellier J, Fleming LE, White MP, et al. Health Benefits of Physical Activity Related to an Urban Riverside Regeneration. Vol. 16, International Journal of Environmental Research and Public Health. 2019. doi:10.3390/ijerph16030462.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.3390/ijerph16030462&link_type=DOI) 

16. 16.Bartoll-Roca X, López MJ, Pérez K, Artazcoz L, Borrell C. Short-term health effects of an urban regeneration programme in deprived neighbourhoods of Barcelona. PLoS One 2024;19:e0300470.
    
    

17. 17.Cooke M, Lumpe E, Stephenson M, Urjansson M, Aliev F, Palviainen T, et al. Alcohol use in early midlife: Findings from the age 37 follow-up assessment of the FinnTwin12 cohort. OSF Prepr 2024. doi:10.31219/OSF.IO/A2N34.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.31219/OSF.IO/A2N34&link_type=DOI) 

18. 18.Caspersen CJ, Powell KE, Christenson GM. Physical Activity, Exercise, and Physical Fitness: Definitions and Distinctions for Health-Related Research. Public Heal Reports 1985;100:126–31.
    
    

19. 19.Borodulin K, Laatikainen T, Juolevi A, Jousilahti P. Thirty-year trends of physical activity in relation to age, calendar time and birth cohort in Finnish adults. Eur J Public Health 2008;18:339–44. doi:10.1093/eurpub/ckm092.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1093/eurpub/ckm092&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=17875578&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F06%2F10%2F2024.06.09.24308658.atom) 
    
    [Web of Science](http://medrxiv.org/lookup/external-ref?access_num=000256274200027&link_type=ISI) 

20. 20.Waller K, Kaprio J, Kujala UM. Associations between long-term physical activity, waist circumference and weight gain: a 30-year longitudinal twin study. Int J Obes 2008;32:353–61. doi:10.1038/sj.ijo.0803692.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1038/sj.ijo.0803692&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=17653065&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F06%2F10%2F2024.06.09.24308658.atom) 

21. 21.Leskinen T, Waller K, Mutikainen S, Aaltonen S, Ronkainen PHA, Alén M, et al. Effects of 32-Year Leisure Time Physical Activity Discordance in Twin Pairs on Health (TWINACTIVE Study): Aims, Design and Results for Physical Fitness. Twin Res Hum Genet 2009;12:108–17. https://doi.org/DOI: 10.1375/twin.12.1.108.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1375/twin.12.1.108&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=19210186&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F06%2F10%2F2024.06.09.24308658.atom) 

22. 22.Jetté M, Sidney K, Blümchen G. Metabolic equivalents (METS) in exercise testing, exercise prescription, and evaluation of functional capacity. Clin Cardiol 1990;13:555–65. doi:10.1002/clc.4960130809.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1002/clc.4960130809&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=2204507&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F06%2F10%2F2024.06.09.24308658.atom) 
    
    [Web of Science](http://medrxiv.org/lookup/external-ref?access_num=A1990DR94900005&link_type=ISI) 

23. 23.Kujala UM, Kaprio J, Sarna S, Koskenvuo M. Relationship of Leisure-Time Physical Activity and MortalityThe Finnish Twin Cohort. JAMA 1998;279:440–4. doi:10.1001/jama.279.6.440.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1001/jama.279.6.440&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=9466636&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F06%2F10%2F2024.06.09.24308658.atom) 
    
    [Web of Science](http://medrxiv.org/lookup/external-ref?access_num=000071839800031&link_type=ISI) 

24. 24.Ainsworth BE, Haskell WL, Herrmann SD, Meckes N, Bassett DRJR, Tudor-Locke C, et al. 2011 Compendium of Physical Activities: A Second Update of Codes and MET Values. Med Sci Sport Exerc 2011;43.
    
    

25. 25.Teeuwen R, Psyllidis A, Bozzon A. Measuring children’s and adolescents’ accessibility to greenspaces from different locations and commuting settings. Comput Environ Urban Syst 2023;100:101912. doi:10.1016/j.compenvurbsys.2022.101912.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/j.compenvurbsys.2022.101912&link_type=DOI) 

26. 26.Milias V, Psyllidis A. Assessing the influence of point-of-interest features on the classification of place categories. Comput Environ Urban Syst 2021;86:101597. doi:10.1016/j.compenvurbsys.2021.101597.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/j.compenvurbsys.2021.101597&link_type=DOI) 

27. 27.van Kamp I, Persson Waye K, Kanninen K, Gulliver J, Bozzon A, Psyllidis A, et al. Early environmental quality and life-course mental health effects: The Equal-Life project. Environ Epidemiol 2022;6.
    
    

28. 28.Wang Z, Zellers S, Whipp AM, Heinonen-Guzejev M, Foraster M, Júlvez J, et al. The effect of environment on depressive symptoms in late adolescence and early adulthood: an exposome-wide association study and twin modeling. Nat Ment Heal 2023. doi:10.1038/s44220-023-00124-x.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1038/s44220-023-00124-x&link_type=DOI) 

29. 29.Thompson TP, Horrell J, Taylor AH, Wanner A, Husk K, Wei Y, et al. Physical activity and the prevention, reduction, and treatment of alcohol and other drug use across the lifespan (The PHASE review): A systematic review. Ment Health Phys Act 2020;19:100360. doi:10.1016/j.mhpa.2020.100360.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/j.mhpa.2020.100360&link_type=DOI) 

30. 30.Poortinga W. Associations of physical activity with smoking and alcohol consumption: A sport or occupation effect? Prev Med (Baltim*)* 2007;45:66–70. doi:10.1016/j.ypmed.2007.04.013.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/j.ypmed.2007.04.013&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=17561247&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F06%2F10%2F2024.06.09.24308658.atom) 
    
    [Web of Science](http://medrxiv.org/lookup/external-ref?access_num=000248469200014&link_type=ISI) 

31. 31.Abu-Omar K, Messing S, Sarshar M, Gelius P, Ferschl S, Finger J, et al. Sociodemographic correlates of physical activity and sport among adults in Germany: 1997–2018. Ger J Exerc Sport Res 2021;51:170–82. doi:10.1007/s12662-021-00714-w.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1007/s12662-021-00714-w&link_type=DOI) 

32. 32.Kivimäki M, Batty GD, Pentti J, Shipley MJ, Sipilä PN, Nyberg ST, et al. Association between socioeconomic status and the development of mental and physical health conditions in adulthood: a multi-cohort study. Lancet Public Heal 2020;5:e140–9. doi:10.1016/S2468-2667(19)30248-8.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/S2468-2667(19)30248-8&link_type=DOI) 

33. 33.Preud’homme G, Duarte K, Dalleau K, Lacomblez C, Bresso E, Smaïl-Tabbone M, et al. Head-to-head comparison of clustering methods for heterogeneous data: a simulation-driven benchmark. Sci Rep 2021;11:4202. doi:10.1038/s41598-021-83340-8.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1038/s41598-021-83340-8&link_type=DOI) 

34. 34.Huang Z. Extensions to the k-Means Algorithm for Clustering Large Data Sets with Categorical Values. Data Min Knowl Discov 1998;2:283–304. doi:10.1023/A:1009769707641.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1023/A:1009769707641&link_type=DOI) 
    
    [Web of Science](http://medrxiv.org/lookup/external-ref?access_num=000077976300004&link_type=ISI) 

35. 35.Al-Zoubi MB, Rawi M al. An Efficient Approach for Computing Silhouette Coefficients. J Comput Sci 2008;4:252–5. doi:10.3844/jcssp.2008.252.255.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.3844/jcssp.2008.252.255&link_type=DOI) 

36. 36.Aschenbruck R, Szepannek G, Wilhelm AFX. Imputation Strategies for Clustering Mixed-Type Data with Missing Values. J Classif 2023;40:2–24. doi:10.1007/s00357-022-09422-y.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1007/s00357-022-09422-y&link_type=DOI) 

37. 37.Jolliffe IT. Outlier Detection, Influential Observations, Stability, Sensitivity, and Robust Estimation of Principal Components BT - Principal Component Analysis. In New York, NY: Springer New York, 2002.p.232–68. Available from: doi:10.1007/0-387-22440-8_10 https://doi.org/10.1007/0-387-22440-8_10.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1007/0-387-22440-8_10&link_type=DOI) 

38. 38.Chen T, Guestrin C. XGBoost: A Scalable Tree Boosting System. Proc 22nd ACM SIGKDD Int Conf Knowl Discov Data Min 2016. doi:10.1145/2939672.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1145/2939672&link_type=DOI) 

39. 39.Yu T, Zhu H. Hyper-parameter optimization: A review of algorithms and applications. arXiv Prepr arXiv2003056*89* 2020.
    
    

40. 40.Lundberg S, Lee S-I. A Unified Approach to Interpreting Model Predictions. 2017. doi:10.48550/arxiv.1705.07874.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.48550/arxiv.1705.07874&link_type=DOI) 

41. 41.Lundberg SM, Erion G, Chen H, DeGrave A, Prutkin JM, Nair B, et al. From local explanations to global understanding with explainable AI for trees. Nat Mach Intell 2020;2:56–67. doi:10.1038/s42256-019-0138-9.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1038/s42256-019-0138-9&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=32607472&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F06%2F10%2F2024.06.09.24308658.atom) 

42. 42.Lim M, Hastie T. Learning Interactions via Hierarchical Group-Lasso Regularization. J Comput Graph Stat 2015;24:627–54. doi:10.1080/10618600.2014.938812.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1080/10618600.2014.938812&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=26759522&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F06%2F10%2F2024.06.09.24308658.atom) 

43. 43.Viinikka A, Tiitu M, Heikinheimo V, Halonen JI, Nyberg E, Vierikko K. Associations of neighborhood-level socioeconomic status, accessibility, and quality of green spaces in Finnish urban regions. Appl Geogr 2023;157:102973. doi:10.1016/j.apgeog.2023.102973.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/j.apgeog.2023.102973&link_type=DOI) 

44. 44.Zhou R, Li Y, Umezaki M, Ding Y, Jiang H, Comber A, et al. Association between Physical Activity and Neighborhood Environment among Middle-Aged Adults in Shanghai. Rissel C, editor. J Environ Public Health 2013;2013:239595. doi:10.1155/2013/239595.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1155/2013/239595&link_type=DOI) 

45. 45.Shearer C, Blanchard C, Kirk S, Lyons R, Dummer T, Pitter R, et al. Physical Activity and Nutrition Among Youth in Rural, Suburban and Urban Neighbourhood Types. Can J Public Heal 2012;103:S55–60. doi:10.1007/BF03403836.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1007/BF03403836&link_type=DOI) 

46. 46.Sandercock G, Angus C, Barton J. Physical activity levels of children living in different built environments. Prev Med (Baltim*)* 2010;50:193–8. doi:10.1016/j.ypmed.2010.01.005.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/j.ypmed.2010.01.005&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=20083131&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F06%2F10%2F2024.06.09.24308658.atom) 
    
    [Web of Science](http://medrxiv.org/lookup/external-ref?access_num=000276000800007&link_type=ISI) 

47. 47.Zhu Z, Li Z, Liu Y, Chen H, Zeng J. The impact of urban characteristics and residents’ income on commuting in China. Transp Res Part D Transp Environ 2017;57:474–83. doi:10.1016/j.trd.2017.09.015.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/j.trd.2017.09.015&link_type=DOI) 

48. 48.Moore JB, Beets MW, Morris SF, Kolbe MB. Comparison of Objectively Measured Physical Activity Levels of Rural, Suburban, and Urban Youth. Am J Prev Med 2014;46:289–92. doi:10.1016/j.amepre.2013.11.001.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/j.amepre.2013.11.001&link_type=DOI) 

49. 49.Hermosillo-Gallardo ME, Jago R, Sebire SJ. Association between urbanicity and physical activity in Mexican adolescents: The use of a composite urbanicity measure. PLoS One 2018;13:e0204739.
    
    

50. 50.Wardle J, Waller J, Jarvis MJ. Sex Differences in the Association of Socioeconomic Status With Obesity. Am J Public Health 2002;92:1299–304. doi:10.2105/AJPH.92.8.1299.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.2105/AJPH.92.8.1299&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=12144988&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F06%2F10%2F2024.06.09.24308658.atom) 
    
    [Web of Science](http://medrxiv.org/lookup/external-ref?access_num=000177109800026&link_type=ISI) 

51. 51.Brodersen NH, Steptoe A, Boniface DR, Wardle J. Trends in physical activity and sedentary behaviour in adolescence: ethnic and socioeconomic differences. Br J Sports Med 2007;41:140 LP – 144. doi:10.1136/bjsm.2006.031138.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1136/bjsm.2006.031138&link_type=DOI) 

52. 52.Van Dyck D, Cerin E, De Bourdeaudhuij I, Salvo D, Christiansen LB, Macfarlane D, et al. Moderating effects of age, gender and education on the associations of perceived neighborhood environment attributes with accelerometer-based physical activity: The IPEN adult study. Health Place 2015;36:65–73. doi:10.1016/j.healthplace.2015.09.007.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/j.healthplace.2015.09.007&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=http://www.n&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F06%2F10%2F2024.06.09.24308658.atom) 

53. 53.Guillien A, Cadiou S, Slama R, Siroux V. The Exposome Approach to Decipher the Role of Multiple Environmental and Lifestyle Determinants in Asthma. Int J Environ Res Public Health 2021;18. doi:10.3390/ijerph18031138.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.3390/ijerph18031138&link_type=DOI) 

54. 54.Cohen DA, Marsh T, Williamson S, Han B, Derose KP, Golinelli D, et al. The Potential for Pocket Parks to Increase Physical Activity. Am J Heal Promot 2014;28:S19–26. doi:10.4278/ajhp.130430-QUAN-213.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.4278/ajhp.130430-QUAN-213&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=24380461&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F06%2F10%2F2024.06.09.24308658.atom) 

55. 55.Peng H, Li X, Yang T, Tan S. Research on the Relationship between the Environmental Characteristics of Pocket Parks and Young People’s Perception of the Restorative Effects—A Case Study Based on Chongqing City, China. Sustainability 2023;15. doi:10.3390/su15053943.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.3390/su15053943&link_type=DOI) 

56. 56.Yang L, Yang H, Cui J, Zhao Y, Gao F. Non-linear and synergistic effects of built environment factors on older adults’ walking behavior: An analysis integrating LightGBM and SHAP. *Trans Urban Data*, Sci Technol 2024:27541231241249864. doi:10.1177/27541231241249866.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1177/27541231241249866&link_type=DOI) 

57. 57.Keskinen KE, Gao Y, Rantakokko M, Rantanen T, Portegijs E. Associations of Environmental Features With Outdoor Physical Activity on Weekdays and Weekend Days: A Cross-Sectional Study Among Older People. Front Public Heal 2020;8.
    
    

58. 58.Zang P, Qiu H, Xian F, Yang L, Qiu Y, Guo H. Nonlinear Effects of the Built Environment on Light Physical Activity among Older Adults: The Case of Lanzhou, China. Int J Environ Res Public Health 2022;19. doi:10.3390/ijerph19148848.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.3390/ijerph19148848&link_type=DOI) 

59. 59.McGinn AP, Evenson KR, Herring AH, Huston SL, Rodriguez DA. Exploring Associations between Physical Activity and Perceived and Objective Measures of the Built Environment. J Urban Heal 2007;84:162–84. doi:10.1007/s11524-006-9136-4.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1007/s11524-006-9136-4&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=17273926&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F06%2F10%2F2024.06.09.24308658.atom) 
    
    [Web of Science](http://medrxiv.org/lookup/external-ref?access_num=000245345300004&link_type=ISI) 

60. 60.Browning MHEM, Rigolon A, McAnirlin O, Yoon H (Violet). Where greenspace matters most: A systematic review of urbanicity, greenspace, and physical health. Landsc Urban Plan 2022;217:104233. doi:10.1016/j.landurbplan.2021.104233.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/j.landurbplan.2021.104233&link_type=DOI) 

61. 61.Cardinali M, Beenackers MA, van Timmeren A, Pottgiesser U. The relation between proximity to and characteristics of green spaces to physical activity and health: A multi-dimensional sensitivity analysis in four European cities. Environ Res 2024;241:117605. doi:10.1016/j.envres.2023.117605.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/j.envres.2023.117605&link_type=DOI) 

62. 62.Huang B, Yao Z, Pearce JR, Feng Z, James Browne A, Pan Z, et al. Non-linear association between residential greenness and general health among old adults in China. Landsc Urban Plan 2022;223:104406. doi:10.1016/j.landurbplan.2022.104406.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/j.landurbplan.2022.104406&link_type=DOI) 

63. 63.Guan M. Associations of fruit & vegetable intake and physical activity with poor self-rated health among Chinese older adults. BMC Geriatr 2022;22:10. doi:10.1186/s12877-021-02709-6.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1186/s12877-021-02709-6&link_type=DOI) 

64. 64.Ping WX, Yan LZ, Meng Z, Yong LH, Ping WX, Yan LZ, et al. Machine-learning-assisted Investigation into the Relationship between the Built Environment, Behavior, and Physical Health of the Elderly in China. Biomed Environ Sci 2023, Vol 36, Issue 10, Pages 987-990 2023;36:987–90. doi:10.3967/BES2023.125.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.3967/BES2023.125&link_type=DOI) 

65. 65.Lee K, Wang J, Heo J. How the physical inactivity is affected by social-, economic- and physical-environmental factors: an exploratory study using the machine learning approach. Int J Digit Earth 2023;16:2503–21. doi:10.1080/17538947.2023.2230944.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1080/17538947.2023.2230944&link_type=DOI) 

66. 66.Ohanyan H, Portengen L, Huss A, Traini E, Beulens JWJ, Hoek G, et al. Machine learning approaches to characterize the obesogenic urban exposome. Environ Int 2022;158:107015. doi:10.1016/j.envint.2021.107015.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/j.envint.2021.107015&link_type=DOI) 

67. 67.FIEC - Statistical Report 2023 [Internet]. [cited 2023 Nov 22]. Available from: [https://fiec-statistical-report.eu/](https://fiec-statistical-report.eu/)