Urbanization compound pathways of global lung cancer incidence risk under proximal and distal interactions

Summary

Lung cancer burden varies across countries undergoing different stages of urbanization, but the interacting roles of behavioral and macro-environmental factors remain difficult to quantify. Using GBD 2021 and World Bank data from 187 countries between 1991 and 2021, we combined GWRF-SHAP to examine nonlinear and spatially heterogeneous interactions between urbanization, proximal behavioral factors, and distal socio-environmental contexts. Lung cancer incidence shifted from traditionally developed regions toward rapidly urbanizing areas. Urbanization-occupational exposure and urbanization-forest cover emerged as the dominant interaction pathways, ranking first in 103 and 33 countries, respectively. Air pollution tended to amplify risk in rapidly urbanizing countries, whereas forest-related interactions generally showed buffering patterns with regional exceptions. Consequently, when formulating geographically differentiated lung cancer prevention strategies, it is essential to consider the combined impact of occupational health management, pollution control, the transition to clean energy, and environmental planning.

Graphical abstract

Highlights

• •A GWRF-SHAP framework quantifies nonlinear urbanization-cancer interactions globally
• •Occupational exposure and forest cover dominate urbanization-related pathways
• •Dominant interaction factors shifted substantially across countries from 1991 to 2021
• •Findings support occupational protection, clean energy, and environmental health governance

Public health; Computational molecular modelling; Environmental health; Cancer; Machine learning

Introduction

Among the numerous challenges facing global public health, cancer remains a major disease threatening human life safety. According to GLOBOCAN 2022 cancer burden statistics, there were 2.48 million new lung cancer cases worldwide in 2022, accounting for 12.4% of all new cancer cases globally. Concurrently, lung cancer caused 1.8 million deaths, representing 18.7% of all cancer deaths worldwide.¹ By 2030, the global incidence of lung cancer is projected to increase by 38% compared to 2018, with global lung cancer deaths projected to reach 2.45 million by 2030.² Lung cancer has ranked as the leading cause of cancer deaths worldwide for consecutive years, becoming the most prevalent cancer globally in terms of incidence and mortality, imposing a severe disease burden. As the aging population trend intensifies, the incidence and mortality burden of lung cancer among the elderly far exceeds that of the working-age population.³ The high treatment costs impose a severe economic burden on families. According to research from Heidelberg University in Germany, lung cancer is projected to cause global losses of approximately $3.9 trillion from 2020 to 2050.⁴ This series of trends highlights the significant impact of lung cancer on global health and economic development, driving countries to establish response systems to improve the current situation.

Urbanization, as an inevitable trend in global socioeconomic development, is quietly transforming human living environments and health conditions. Its association with lung cancer incidence has become a significant topic of academic interest. According to data from the United Nations Human Settlements Program (UN-Habitat), although the 2019 novel coronavirus pandemic negatively impacted global urbanization, the growth of the world’s urban population is returning to its trajectory. By 2025, the global urban population is projected to exceed 5 billion, accounting for 55.3% of the total population. It is estimated that by 2050, the global urban population will increase by an additional 2.2 billion.⁵ Global urbanization drives industrial upgrading and transformation, with residents in affluent urban areas enjoying a life expectancy 20 years longer than those in impoverished regions.⁶ However, amid rapid global urbanization, air pollution has intensified, not only imposing a burden of respiratory diseases and fueling the prevalence of noncommunicable diseases but also widening health inequalities. Concurrently, the phenomenon of urban community segregation has precipitated severe mental health crises among residents.^7,8The development of lung cancer results from the combined effects of multiple factors. Current research analyzes associations with lung cancer from proximal factors such as lifestyle behaviors. By constructing lung cancer risk prediction models, it identifies links between high-risk factors in unhealthy lifestyles and lung cancer within populations, revealing associations between smoking, high alcohol intake, negative life events, and lung cancer occurrence.⁹ Additionally, other studies examine distant macro-level factors. Research indicates that changes in the built environment impact air quality, leading to air pollution (primarily particulate matter and SO₂) and exposure to pollutants, which may indirectly influence lung cancer development.¹⁰ Furthermore, the urban-rural imbalance in healthcare resource allocation affects lung cancer treatment outcomes. Survival rates for rural patients treated in urban institutions are significantly higher than those treated in rural hospitals, highlighting how unequal distribution of medical resources directly impacts the treatment effect of lung cancer.¹¹ Relatedly, spatial multilevel approaches have been employed to link urbanization-related contexts with chronic disease outcomes. For instance, Jiménez-Ortega’s study utilized a spatial multilevel framework to examine how urbanization and unhealthy food expenditures jointly influence type 2 diabetes risk in Mexico.¹² Such research suggests that integrating urbanization into a unified framework of multilevel determinants facilitates a more systematic understanding of regional disparities in chronic disease risk and their implications for governance strategies. However, existing lung cancer research still faces several limitations: While some studies have examined lung cancer drivers from a global perspective, they have not integrated finer-grained dimensions such as regional urbanization patterns and economic development. This makes it difficult to reveal more nuanced geospatial heterogeneity in lung cancer incidence risk patterns, thereby failing to support the development of differentiated global public health strategies. At the global level, existing studies predominantly employ linear and nonlinear models to examine the association between lung cancer and single factors. They lack analysis of the multidimensional interactions between proximal and distal factors influencing lung cancer, making it difficult to systematically reveal the impact of urbanization and its interacting factors across different countries and regions worldwide. Furthermore, most studies can only identify key drivers of lung cancer, while few can quantify the marginal contributions of these drivers across different country types.

Existing research has begun expanding from traditional statistical regression to spatial statistical analysis, such as using geographically weighted regression (GWR) to study the effects of air pollution and urbanization rates on lung cancer mortality rates.¹³ However, GWR cannot capture complex nonlinear relationships. To reveal nonlinear relationships and capture spatial heterogeneity, scholars have combined GWR with machine learning models, proposing the Geographically Weighted Random Forest (GWRF) model.¹⁴ The geographically weighted random forest (RF) model combines the predictive power and ability to handle factor interactions of RF with the spatial sensitivity of GWR. This overcomes GWR’s limitation in capturing complex nonlinear relationships, enabling it to capture spatial heterogeneity and flexibly fit nonlinear relationships. It effectively addresses the shortcomings of previous studies that were limited to single factors. Simultaneously, the SHAP machine learning model is employed to perform local interpretability decomposition on GWRF results. This quantifies the marginal contributions of individual drivers and their interaction terms to lung cancer incidence, making it suitable for heterogeneity analysis. Ultimately, integrating GWRF with SHAP establishes a spatial machine learning framework that combines spatial heterogeneity capture capabilities with interpretability tools.¹⁵

Therefore, this study analyzes multi-source data from GBD2021 and the World Bank spanning 1991–2021. Employing the GWRF-SHAP model, it focuses on the interactive effects of multiple drivers. From the perspective of urbanization interactions, it dissects the dual and triple coupling mechanisms linking lung cancer to lifestyle behaviors (proximal factors), and levels of social development (distal factors). This aims to reveal the heterogeneous patterns across countries and regions at the global scale, thereby identifying high-vulnerability areas and populations with pronounced cumulative risks. The findings provide quantitative evidence for targeted interventions and evidence-based policy development at the subregional level.

Results

Descriptive statistics and spatial visualization

Descriptive statistics

The baseline characteristics of the variables included in urbanization, proximal factors, and distal factors are shown in Table 1.

Table 1.

Descriptive characteristics of 31-year average variables included in the main GWRF-SHAP analysis across 187 countries and regions

Variable	Mean	Std	Min	Max	IQR
Urban	0.3512	0.1818	0.0009	0.8371	0.2707
Per_GDP	12421.1758	19970.5492	189.2822	167809.2127	11654.973
Forest	32.5898	24.0812	0	98.551	39.3758
Employment	56.2686	11.3281	22.634	87.518	12.909
Haqi	49.2758	21.9591	9.318	94.2311	35.9247
Air_pollution	142.2746	101.3146	2.347	437.3718	172.6378
Tobacco	94.4651	51.5004	15.4048	336.7884	72.5613
High_alcohol	30.1695	21.2288	0.6293	152.5752	25.7977
High_BMI	68.2427	40.3437	9.3234	293.0997	47.3879
Low_activity	10.3283	6.4232	1.6715	41.7069	7.516
Occupational_matter	14.6876	9.1431	2.9818	67.9071	9.0014

Global distribution of lung cancer risk

As shown in Figure 1, at the national level, Greenland recorded the highest ASIR in 1991 (106.41 per 100,000), while Nigeria had the lowest (1.35 per 100,000). By 2021, Monaco had the highest ASIR (64.05 per 100,000), while Nigeria had the lowest (1.52 per 100,000). At the regional level, high ASIR values for lung cancer in 1991 were primarily concentrated in North America, Australia, and high-latitude developed regions of Eurasia. By 2021, these high values had shifted to small southern European countries, China in East Asia, and Greenland, while low values remained persistently concentrated in South Asia, the Middle East, and the African continent.

Figure 1.

Spatial distribution of lung cancer incidence rates in 187 countries worldwide in 1991 and 2021

(A) Country-level ASIR of lung cancer in 1991.

(B) Country-level ASIR of lung cancer in 2021.

Urbanization—Identifying global SHAP pathways through proximal-distal synergistic coupling

Key features of global SHAP

Using the GWRF-SHAP model, we ranked the absolute SHAP interaction contributions of the proximal, distal, and triple interactions between urbanization and lung cancer incidence across 187 countries and regions worldwide (Figure 2). Among these, the top four proximal interactions were urbanization with occupational matter, high BMI, high alcohol, and low activity; the top four distal interactions were urbanization with forest, air pollution, per_GDP, and employment; The top four triple interactions were urbanization with high alcohol and forest, high BMI and forest, tobacco and air pollution, and high BMI and air pollution. Among overall interaction factors, the most influential interactions on lung cancer incidence were urbanization with occupational matter and forest. Regarding the top interaction characteristics by country, the interactions driving lung cancer outcomes most frequently were urbanization with occupational matter (103 countries) and forest (33 countries). Countries driven by triple interactions were relatively scattered, with only a few driven by urbanization with high BMI (3 countries), air pollution (1 country), and per_GDP (1 country). Among major economically developed regions, most of North America and Europe are driven by the interaction between urbanization and occupational matter, while high-latitude and tropical regions are more frequently driven by the interaction between urbanization and forest.

Figure 2.

Contribution patterns of lung cancer incidence under urbanization interactions and distribution of leading interaction factors across 187 countries

(A) Summary of interaction-feature contributions identified by the GWRF-SHAP model. Bars show the contribution magnitude of each interaction feature to lung cancer incidence across countries, and overlaid dots represent country-specific SHAP values. Dot colors indicate the relative feature value. The dot matrix below identifies the variables included in each interaction feature, with filled circles denoting included variables.

(B) Spatial distribution of the dominant interaction feature across 187 countries and regions. Countries are colored according to the interaction feature with the largest absolute SHAP contribution, highlighting the geographically leading urbanization-related interaction pathway associated with lung cancer incidence.

Temporal shifts in dominant urbanization-related interaction factors: A benchmark-year comparison

To strengthen the temporal interpretation beyond the time-aggregated interaction maps, we conducted a benchmark-year comparison for 1991 and 2021 (Figure 3). For each year, we identified the dominant interaction factor for each country as the interaction feature with the largest absolute SHAP interaction contribution. The benchmark-year comparison shows that dominant interaction patterns are not static: 177 out of 187 countries exhibited a change in the dominant interaction factor between 1991 and 2021.

Figure 3.

Benchmark-year comparison of dominant interaction factors by country in 1991 and 2021

(A) Dominant urbanization-related interaction factors across countries in 1991.

(B) Dominant urbanization-related interaction factors across countries in 2021.

In 1991, dominant interaction factors were more concentrated, with the triple interaction of urbanization, high BMI, and HAQI present in 82 countries (43.9%). This was followed by the triple interaction of urbanization, tobacco, and HAQI (49 countries, 26.2%) and the triple interaction of urbanization, high BMI, and air pollution (20 countries, 10.7%). By 2021, the distribution showed diversification. The triple interaction of urbanization with tobacco and HAQI dominated in 30 countries (16.0%). The interaction of urbanization with occupational exposure dominated in 23 countries (12.3%). The triple interaction of urbanization with low physical activity and HAQI accounted for 22 countries (11.8%). The interaction of urbanization with forest cover and the triple interaction of urbanization with high BMI and air pollution each accounted for 17 countries (9.1%).

Lifestyle patterns beneath the surface of urbanization: The overlapping effects of proximal factors

Among the interactions between proximal factors and urbanization (Figure 4), the combined effect of occupational exposure and urbanization exhibited the most pronounced regional variation in lung cancer burden. The interaction effects of other behavioral factors, however, generally show a suppressive effect and are accompanied by local positive clustering. Specifically, the SHAP value range for the Urban and Occupational_matter interaction was the widest (−0.2358 to 0.1454), with positive hotspots predominantly distributed across Eastern Europe and negative values concentrated in Central Africa, Southeast Asia, and northern Eurasia. This indicates that occupational exposure under overlapping urbanization exerts the strongest influence on lung cancer burden with the most pronounced regional differentiation. The interaction between Urban and High_BMI or High_alcohol showed negative effects in most countries, with positive effects observed only in a few mid-to-high latitude nations. The significant negative zone for Urban and High_BMI was located in Southeast Asia and Africa, potentially linked to lower nutritional and metabolic levels. The interaction between Urban and Low_activity predominantly showed negative effects globally, with positive values primarily observed in island nations of Oceania and mid-to-high latitude regions such as Mongolia, Russia, and Canada. This suggests regional variations are more likely influenced by multiple factors, including differing stages of urban development and transportation patterns. Overall, Occupational_matter shows the strongest spatial heterogeneity driven by urbanization. The interaction between High_BMI, High_alcohol, and Low_activity with urbanization predominantly shows negative suppression, but there are also local positive aggregations.

Figure 4.

Spatial pattern of lung cancer incidence under interactions between urbanization and proximal factors, showing the country-level SHAP interaction contributions of urbanization with occupational exposure, high BMI, high alcohol intake, and low physical activity

Social development behind urbanization: The moderating effect of distal factors

In the interaction outcomes between distal factors and urbanization (Figure 5), forests and air pollution exhibit significant directional differences in their interactions with urbanization across varying climatic and developmental contexts, while socioeconomic factors demonstrate relatively weaker marginal effects. Specifically, the SHAP value range for the Urban and Forest interaction is the widest (−0.0390 to 2.3681), predominantly showing weak negative effects across most regions, interspersed with locally strong positive effects in mid-to-high latitudes. In high-latitude regions, the urbanization and forest interaction shows localized positive SHAP contributions, indicating an interaction-attribution pattern that differs from that observed in most countries. A plausible contextual interpretation, consistent with prior evidence from cold-climate settings, is that residential wood burning for heating and, in some cases, open burning practices may increase particulate matter exposure, which is compatible with the observed risk-enhancing interaction pattern. However, because solid fuel use and energy structure variables were not directly included in the present model, this interpretation should be regarded as a literature-informed hypothesis rather than a data-driven mechanistic conclusion. In the Urban and Air_pollution interaction, risk elevation primarily occurred in rapidly urbanizing nations across Oceania and East Asia, while some economically developed countries and less urbanized African regions exhibited reduced risk, reflecting stage-specific differences in urbanization. In contrast, the overall impact of Urban and Per_GDP through Employment is relatively weak, with marginal contributions ranging from approximately −0.02 to 0.08 and −0.03 to 0.04. This suggests that socioeconomic “upstream factors” exert influence primarily as underlying moderators rather than core drivers.

Figure 5.

Spatial pattern of lung cancer incidence under interactions between urbanization and distal factors, showing the country-level SHAP interaction contributions of urbanization with forest cover, air pollution, per capita GDP, and employment

Lifestyle-social development coupling pathway: The triple interaction of urbanization with proximal and distal factors

In the triple-interaction GWRF-SHAP model (Figure 6), forests and pollution exhibit distinct effects categorized as “modulation” and “amplification.” Notably, the triple interaction between Urban, High_alcohol, and Forest reveals a clear hemispheric divergence: Hotspots predominantly cluster in North America’s mid-north, Southern Europe, and the Middle East, while most of Eurasia exhibits cold spots. In Russia, characterized by abundant forest resources and cold climates, this interaction tends to be negative, suggesting forest-related factors may reverse the risks associated with “alcohol culture.” The negative impact of the triple interaction between Urban, High_BMI, and Forest is concentrated in forested nations across North and South America and Southeast Asia, while positive effects are more prevalent in forested-scarce regions of Africa and central Eurasia. The triple interaction between Urban, Tobacco, and Air-pollution exhibits positive synergistic amplification in East Asia (China), Greenland, and Canada, forming spatially scattered hotspots. This suggests that in urbanized contexts where pollution and tobacco consumption coexist, their synergistic positive effects are amplified. Cold spots for the triple interaction of Urban, High_BMI, and Air-pollution are concentrated in Africa, potentially linked to lower development levels and nutritional deficiencies in African nations. In contrast, regions undergoing rapid urbanization, higher BMI, and heavier pollution—such as Eurasia, Oceania, and North America—exhibit risk patterns driven by urbanization stages, manifesting as gradient differentiation. Overall, forest resources often mitigate risk, particularly in forested regions such as the Americas and Southeast Asia, where reduced risk is common. However, in some areas, the effect reverses: For instance, the impact becomes negative in Russia (forested and cold), while elevated risks persist in North America’s central-northern regions, Southern Europe, and the Middle East. Air pollution, conversely, tends to synergistically amplify lung cancer risk, exhibiting gradient differences aligned with urbanization development.

Figure 6.

Spatial pattern of lung cancer incidence under triple interactions involving urbanization, proximal factors, and distal factors, showing the country-level SHAP interaction contributions of urbanization with high alcohol intake and forest cover, high BMI and forest cover, tobacco and air pollution, and high BMI and air pollution

Discussion

This study delineates the spatiotemporal evolution of global lung cancer ASIR from 1991 to 2021 at both national and regional levels. Using GWRF-SHAP, it quantifies the contribution patterns of multi-level interactions between urbanization, proximal lifestyle behaviors, and distal macro-socioeconomic factors to lung cancer risk. First, regarding the spatiotemporal trajectory, the spatial center of high-risk areas gradually shifted from traditionally developed nations to rapidly urbanizing regions between 1991 and 2021. In contrast, low-risk zones remained stable, predominantly distributed in countries with slower urbanization. Second, in terms of interaction significance, the global lung cancer risk pattern exhibits dual dominant pathways, primarily driven by the interactions between urbanization and occupational exposure, and between urbanization and forest cover. Finally, regarding interaction categories, the proximal overlapping pattern globally demonstrates urbanization-occupational exposure interactions as the dominant driver. This is supplemented by urbanization-interactions with high BMI, high alcohol consumption, and low physical activity, which suppress risks in multiple regions while locally elevating risks in mid-to-high latitudes. In the distal regulatory pattern, urbanization-forest interactions exhibit weak negative buffering effects and counteract lifestyle-related risks within triple interactions. The interaction between urbanization and air pollution diverges across stages of urbanization: it elevates risks in rapidly urbanizing countries while reducing them in high-income or low-urbanization nations. Within the triple interaction, forests exhibit both regulatory buffering and synergistic amplification effects on pollution.

Spatiotemporal evolution pathways of the lung cancer burden

In 1991, the high burden of lung cancer was primarily concentrated in North America, Australia, and developed countries in the high latitudes of Eurasia. These regions had generally undergone early industrialization and large-scale urbanization, compounded by factors such as high smoking rates, concentrated occupational exposures, and rising levels of air pollution.¹⁶ This made these regions become the major high-risk concentration areas in the world at that time. By 2021, the geographic distribution of high lung cancer burden had partially shifted globally. The center of high ASIR gradually moved from traditional high-latitude regions in Europe and North America toward southern European countries, East Asia (particularly China), and Greenland. This shift primarily resulted from signs of reduced lung cancer burden in some traditional high-incidence countries and regions, driven by advancements in healthcare systems, tobacco control policies, and environmental governance.¹⁷ In contrast, regions such as South Asia, the Middle East, and sub-Saharan Africa exhibit lower lung cancer ASIRs. This reflects their nascent stages of urbanization and industrialization, lower population life expectancy, and relatively limited impacts from smoking and occupational exposures.^18,19

Therefore, from a global perspective, given regional differences in urbanization stages, lifestyles, and macro-environmental governance, lung cancer prevention and control efforts must prioritize sustained governance in high-burden countries while proactively planning risk mitigation in rapidly urbanizing regions. This approach avoids repeating the governance challenges faced by early industrialized nations.

Global SHAP importance landscape under interaction features

The results of this study reveal that the interaction between urbanization and occupational exposure, as well as the interaction between urbanization and forest cover, dominate the global lung cancer risk landscape. These interactions emerged as the primary drivers in 103 and 33 countries, respectively, highlighting their pivotal role in shaping lung cancer burden across diverse urbanization processes and environmental backgrounds.

First, the interaction between urbanization and occupational exposure exhibits clustering patterns in most North American and European countries. These nations typically experienced rapid urban expansion and industrial development in their early stages, where prolonged exposure to silica, dust, asbestos, and harmful gases in urban environments amplified lung cancer risks over time. For instance, widespread asbestos use in multiple European countries around the mid-20th century led to persistently elevated lung cancer risks among workers in construction, shipbuilding, and manufacturing sectors.^20,21 Studies in North America also demonstrate a positive association between occupational dust exposure and lung cancer risk, particularly pronounced in highly urbanized areas.²²

Second, the interaction between urbanization and forest cover dominates in certain high-latitude and tropical countries, including Russia, Japan, Greenland, and regions near the equator. Notably, while this interaction remains dominant in both high-latitude and tropical areas, its effect is generally associated with negative SHAP interaction contributions across most regions, with only isolated exceptions such as Greenland showing a positive correlation. This pattern likely reflects forests' buffering role against urbanization risks in most high-latitude and tropical regions.^23,24 Extensive forest cover aids pollutant dispersion and dilution, thereby reducing overall environmental risks compounded by urbanization.²⁵ Greenland stands as a rare exception, where the interaction between urbanization and forests positively elevated lung cancer incidence in the region. A literature-informed contextual interpretation is that this may relate to the coexistence of historically high smoking prevalence and additional particulate exposure from cold-climate combustion practices (e.g., residential wood burning and open burning). Because variables on household solid-fuel use and energy structure were not directly included in the present model, this interpretation should be regarded as a plausible explanatory pathway rather than a data-driven mechanistic conclusion.^{26,27,28,29,30}

Overall, countries dominated by urbanization and occupational exposure are predominantly traditional industrialized regions in North America and Europe, where lung cancer burden is closely linked to workplace environmental exposures. Conversely, countries characterized by urbanization and forest cover are concentrated in high-latitude and equatorial regions (such as Russia, Greenland, Canada), where lung cancer risk is more associated with interactions between energy use and climatic conditions. Therefore, for urbanization-occupational exposure dominant countries, efforts should focus on enhancing workplace ventilation and personal protective measures, phasing out high-pollution, high-exposure industries, and transitioning toward low-carbon and clean energy sectors. For urbanization-forest cover dominant countries, strategies should include increasing vegetation coverage while reducing residential wood burning and open burning practices.

Global spatial differentiation patterns under proximal, distal, and triple interaction types

In the interaction between proximal factors and urbanization, occupational exposure exhibits the strongest spatial heterogeneity. Specifically, the SHAP values for the interaction between urbanization and occupational exposure show the widest range, with positive hotspots concentrated in Eastern European countries and negative regions distributed across Central Africa, Southeast Asia, and northern Eurasia. This spatial distribution reflects differences in industrial development and urbanization stages across regions. Eastern Europe consists primarily of traditionally industrialized nations where occupational exposure factors such as dust and toxic gases have accumulated over prolonged urbanization. For instance, despite the EU ban on all forms of asbestos,³¹ workers in construction, electricians, and carpenters may still encounter high asbestos concentrations during renovation, maintenance, and demolition, increasing lung cancer risks.³² In contrast, Central Africa and Southeast Asia exhibit lower urbanization levels and industrial structures dominated by agriculture and light industry, resulting in generally limited occupational exposure and thus a negative, inhibitory interaction effect.³³ Regarding the interaction between urbanization and high BMI/alcohol consumption, most countries exhibit negative effects, with positive interactions observed only in a few high-income, mid-to-high latitude nations. Regarding the interactions between urbanization and high BMI or alcohol consumption, most countries show negative SHAP interaction contributions, with positive contributions observed only in a few high-income, mid-to-high-latitude countries. For BMI, prior large-scale cohort evidence suggests that the BMI-lung cancer association is non-trivial. A pooled analysis of multiple prospective cohorts reported an overall inverse association between BMI and lung cancer risk, while also highlighting heterogeneity across subgroups.^34,35 Meanwhile, cross-country differences may reflect variation in the prevalence of obesity and alcohol use, correlated behavioral risk clustering, and differences in urbanization contexts across development stages.^36,37 Similarly, the interaction between low physical activity and urbanization shows a negative trend globally. However, positive interactions are observed in mid-to-high latitude regions such as Mongolia, Russia, and Canada. This may occur because in cold, high-latitude areas, the cold climate inhibits physical activity, leading to increased health risks when low physical activity combines with an urban environment.³⁸

In the interaction between distal factors and urbanization, the moderating effect of forest cover stands out most prominently. The interaction between urbanization and forests exhibits a predominantly weak negative SHAP interaction contributions globally, indicating a broadly distributed risk-attenuating interaction pattern in many regions. However, the interaction between air pollution and urbanization shows marked heterogeneity across stages of urbanization. In rapidly urbanizing nations such as East Asia and Oceania, where fossil fuels dominate energy structures, this interaction amplifies risks synergistically. Conversely, in some high-income countries and less urbanized African regions, emission controls, clean energy adoption, or slower industrialization progress mitigate risks through this interaction.³⁹

Within the triple-interaction domain, the interplay between forests, pollution, and lifestyle further reveals complex regional differentiation patterns. In forested, cold Russia, the interaction association of the urbanization-alcohol consumption-forest triple interaction tends to be negative, suggesting forests may exert an inverse regulatory effect on risks associated with high alcohol consumption. This indicates that nature exposure may mitigate the burden of heavy drinking by reducing stress and other pathways.⁴⁰ The urbanization-smoking-air pollution triple interaction shows positive risk-enhancing contributions in regions including China, Greenland, and parts of Canada. In East Asia, this pattern is consistent with long-term high particulate matter exposure.⁴¹ In high-latitude regions (e.g., Greenland and Canada), this pattern may also be related to cold-season combustion-related exposures reported in prior studies.⁴² To avoid repetition, the detailed contextual discussion of high-latitude combustion-related explanations and the corresponding interpretation boundary is provided above. Overall, forests mitigate risks in most regions, while air pollution tends to synergistically amplify lung cancer risk, exhibiting gradient differentiation across different stages of urbanization.

Therefore, lung cancer prevention strategies must be tailored to local conditions. For instance, in Eastern Europe and some industrialized nations, emphasis should be placed on strengthening occupational exposure regulation and urban pollution control. In rapidly urbanizing developing countries, attention should focus on controlling pollutant emissions to avoid the synergistic increase in risks from air pollution. In high-latitude cold regions, while increasing vegetation coverage, special emphasis should be placed on transforming energy utilization methods, gradually transitioning to clean energy sources. Concurrently, attention should be given to developing indoor activity spaces to prevent reduced physical activity levels caused by severe cold.

Taken together, our findings indicate that global lung cancer incidence risk is spatially structured by heterogeneous urbanization-related interaction pathways. Between 1991 and 2021, high-incidence patterns increasingly involved rapidly urbanizing settings, while the interactions of urbanization with occupational exposure and forest cover emerged as the leading pathways across countries. These findings support geographically differentiated prevention priorities: strengthened occupational exposure control and screening in traditionally industrialized settings; clean-energy transition and proactive pollution control in rapidly urbanizing regions; and environmental planning together with reductions in combustion-related exposures in forest-rich or cold-climate settings. Given the ecological and attribution-based nature of the present analysis, these implications should be interpreted as policy-relevant evidence rather than causal effects.

Limitations of the study

This study also has several limitations. First, the accuracy and robustness of GBD estimates depend heavily on the quality and quantity of data used for modeling. Some countries lack high-quality mortality data, particularly in regions with lower SDI.⁴³ Second, the use of a country-level ecological design entails a relatively coarse spatial resolution, which may mask substantial within-country spatiotemporal heterogeneity in exposures, urbanization processes, and lung cancer burden. Therefore, our results should be interpreted as evidence of cross-national ecological patterns rather than within-country local relationships, and they may be subject to ecological fallacy and the modifiable areal unit problem. Future studies should integrate subnational or city-level longitudinal data to better characterize micro-scale spatial non-stationarity and local exposure-health relationships. Third, the use of annual average air pollution levels fails to distinguish between different pollutants and seasonal effects. Fourth, methodologically, GWRF-SHAP provides only an interactive model for attributing associated factors, rather than causal identification. This inherently limits causal inference to some extent. Fifth, the study did not systematically incorporate process variables such as energy structure, urban form, and open burning. Future research should integrate these factors to clarify causal relationships. Concurrently, temporal analysis could be introduced by constructing an LSTM-GWRF coupled framework to characterize the dynamic evolution and lag effects of exposure-health relationships. Integrating finer-grained urban-scale data would enhance the ability to identify micro-scale spatial non-stationarity.

Resource availability

Lead contact

Further information and requests for resources should be directed to and will be fulfilled by the lead contact, Baoguo Shi, Department of Economics, School of Economics, Minzu University of China, Beijing, China. Email: bgshi2008@hotmail.com.

Materials availability

This study did not generate new unique reagents, biological materials, or physical materials.

Data and code availability

• •Data: This paper analyzes existing publicly available data. Lung cancer incidence and risk-factor estimates were obtained from the Global Burden of Disease Study 2021 through the GBD Results Tool (https://vizhub.healthdata.org/gbd-results/). Socioeconomic and environmental indicators were obtained from World Bank Open Data (https://data.worldbank.org/). The source data used in this study are publicly accessible through these official data platforms, as also listed in the key resources table.
• •Code: All original code used for data processing, GWRF-SHAP modeling, and model evaluation has been deposited at Zenodo and is publicly available under Zenodo: https://doi.org/10.5281/zenodo.20463710. The source code repository is also available at GitHub: https://github.com/Liboxi1999/global-lung-cancer-urbanization-gwrf-shap. This accession information is also provided in the key resources table.
• •Additional information: Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Acknowledgments

We thank the Global Burden of Disease collaborators and the World Bank Open Data team for making their datasets publicly accessible. This work was supported by the National Natural Science Foundation of China (grant no. 72174047) and the Hangzhou Normal University Research Project (grant no`. 4265C50224204063).

Author contributions

Conceptualization, Y.L. and B.S.; methodology, Y.L., Y.S., B.L., and F.Q.; data curation, Y.S., B.L., and F.Q.; formal analysis, Y.S., B.L., and F.Q.; software, Y.S., B.L., and F.Q.; investigation, Y.S., B.L., and F.Q.; validation, Y.L., B.S., and Y.C.; visualization, Y.L. and X.Z.; writing – original draft, Y.S., B.L., Y.C., and F.Q.; writing – review and editing, Y.L., B.S., Y.L.; supervision, Y.L. and B.S.; project administration, Y.L. and B.S.

Declaration of interests

The authors declare no competing interests.

Declaration of generative AI and AI-assisted technologies in the writing process

During the preparation of this work, the authors used AI tools to improve the clarity and language of the manuscript. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Deposited data
lung cancer incidence and risk-factor estimates	Global Burden of Disease Study 2021	https://vizhub.healthdata.org/gbd-results/
World Bank Open Data: socioeconomic and environmental indicators	World Bank	https://data.worldbank.org/
Software and algorithms
Analysis code for data processing, GWRF-SHAP modeling, and model evaluation	This paper	Zenodo: https://doi.org/10.5281/zenodo.20463710
Source code repository	GitHub	GitHub: https://github.com/Liboxi1999/global-lung-cancer-urbanization-gwrf-shap
Geographically weighted random forest model	This paper	Implemented using the deposited analysis code
Python	Python Software Foundation	Version 3.12
R	R Foundation for Statistical Computing	Version 4.3.2

Experimental model and study participant details

This study was an ecological analysis based on publicly available country-level aggregated data. No individual-level human participant data, animal experiments, cell lines, or clinical interventions were involved. Therefore, ethical approval and informed consent were not required. The analytical units were 187 countries and regions worldwide from 1991 to 2021. Lung cancer incidence was measured using age-standardized incidence rates.

Method details

Data sources and indicator construction

The data presented and analyzed in this study are sourced from the Global Burden of Disease (GBD) Study (https://vizhub.healthdata.org/gbd-results/) and World Bank Open Data (https://data.worldbank.org/). This study obtained lung cancer incidence data, urbanization indicators, and lifestyle behavior data for 187 countries and regions worldwide from 1991 to 2021 from the GBD 2021. World Bank Open Data provided annual socioeconomic and demographic indicators for each country from 1991 to 2021. This study uses age-standardized incidence rates for lung cancer as the outcome measure. The study assembled annual country-level data for 187 countries and regions from 1991 to 2021. For the main GWRF-SHAP analysis, each outcome, urbanization indicator, proximal factor, and distal factor was averaged over the 31-year study period within each country or region. Therefore, the final dataset used for the main spatial modeling analysis consisted of 187 country-level observations, with each observation representing the long-term average profile of one country or region. In addition, a benchmark-year comparison was conducted using country-level data from 1991 to 2021 to examine temporal changes in dominant urbanization-related interaction factors. The countries and regions included in the study are listed in Appendix Table S1, and the relevant disease codes are detailed in Appendix Table S2.

In terms of indicator construction, this study explores the impact of interactions among proximal lung cancer risk factors, distal lung cancer risk factors, and triple interactions between proximal and distal factors on lung cancer burden from an urbanization interaction perspective. Urbanization level is defined as the proportion of urban population to total population. Proximal factors encompass lifestyle and behavioral determinants influencing lung cancer incidence, including tobacco consumption, high alcohol intake, obesity rates, insufficient physical activity, and occupational exposure. Distal factors primarily represent macro-socioeconomic determinants, comprising per_GDP, employment rate, healthcare accessibility index, forest coverage, and air pollution levels. Note that Per_GDP and employment are included as distal socioeconomic context indicators; they are not components of the urbanization measure, which is defined solely as the proportion of urban population. (see Key indicators and data sources below).

Key indicators and data sources.

Dimension	Variable	Data source
Outcome Indicator	age-standardized incidence rate of lung cancer	Global Burden of Disease Study 2021; GBD Results Tool
Urbanization Indicators	the proportion of the urban population relative to the total population	Global Burden of Disease Study 2021; GBD Results Tool
Proximal Factors	Tobacco	Global Burden of Disease Study 2021; GBD Results Tool
	High_alcohol	Global Burden of Disease Study 2021; GBD Results Tool
	High_BMI	Global Burden of Disease Study 2021; GBD Results Tool
	Low physical activity	Global Burden of Disease Study 2021; GBD Results Tool
	Occupational exposure	Global Burden of Disease Study 2021; GBD Results Tool
Distal Factors	Per_GDP	World Bank Open Data
	Employment	World Bank Open Data
	HAQI	Global Burden of Disease Study 2021; GBD Results Tool
	Forest	World Bank Open Data
	Air_pollution	Global Burden of Disease Study 2021; GBD Results Tool

Quantification and statistical analysis

Analytical sample and descriptive analysis

The analytical unit of the main ecological spatial analysis was the country or region. Although annual data were collected for 187 countries and regions from 1991 to 2021, the main GWRF-SHAP analysis was conducted using the 31-year average value of each study variable for each country or region. Accordingly, the final sample size for the main model was n = 187 countries and regions, where n represents the number of country-level spatial units included in the analysis rather than country-year observations.

Continuous variables included in the main analysis were summarized using the mean, standard deviation (SD), minimum, maximum, and interquartile range (IQR), as reported in Table 1. All predictor variables were standardized using z-scores before model fitting to improve comparability across variables measured on different scales.

Geographically weighted random forest (GWRF)

This study employs the Geographically Weighted Random Forest (GWRF) model for modeling. GWRF combines the predictive power of ensemble learning (via random forests) with the spatial sensitivity of Geographically Weighted Regression (GWR), while accommodating spatial heterogeneity, nonlinear effects, and interactions among variables. Each decision tree in the Random Forest (RF) training process utilizes a random subsample of features for splitting, which helps reduce overfitting and enhances model robustness.⁴⁴ However, due to spatial heterogeneity across regions, the Geographically Weighted Random Forest (GWRF) extends the RF framework by incorporating the spatial weighting principle of Geographically Weighted Regression (GWR) to capture this local variability. GWRF enables exploration of spatial variations in the relationship between the target variable and its predictors. It accounts for the fact that predictor effects may differ by location. Furthermore, GWRF adaptively selects the importance of local features at each geographic location, aiding in revealing differences in the marginal effects of variables across countries and regions.⁴⁵ This study employs geographic locations as the core unit for local modeling. First, data undergoes standardized processing to ensure comparability in scale and type, followed by constructing a feature set for spatial learning. Subsequently, at each geographic sample point u_ij=(x_i,y_i), its local neighborhood is determined based on spherical distance, and a spatial weight matrix is constructed accordingly, granting higher weights to spatially closer samples. Let d_ij denote the spherical distance between location u_i and training sample u_j, and b_i denote the adaptive bandwidth. The weight of location u_i for sample u_j is defined as:

$${\mathrm{w}}_{\text{ij}}=\exp \left[-{\left(\frac{{\mathrm{d}}_{\text{ij}}}{{\mathrm{b}}_{\mathrm{i}}}\right)}^2\right]$$

This formulation embodies the concept of kernel weighting with monotonically decreasing weighting by distance, where the bandwidth bi controls the effective range of the local neighborhood. By training a local random forest on the weighted sample sets {(x_i,y_i,w_ij)} at location u_i, we obtain a location-specific prediction function:

$$\hat{\mathrm{y}}\left({\mathrm{u}}_{\mathrm{i}}\right)={\mathrm{f}}_{\text{RF}}\left(\mathrm{X};\mathrm{W}\left({\mathrm{u}}_{\mathrm{i}}\right)\right)=\frac{1}{\mathrm{M}}\sum _{\mathrm{m}=1}^{\mathrm{M}}{\mathrm{T}}_{\mathrm{m}}\left(\mathrm{X};\mathrm{W}\left({\mathrm{u}}_{\mathrm{i}}\right)\right)$$

W(u_i) = {w_ij}j denotes the weight set at that position, T_m represents the mth regression tree, and M is the number of trees. Tree growth and splitting are based on the principle of minimizing weighted impurity. Under weighted variance conditions, the objective can be expressed as:

$$\min \text{split}\sum _{\mathrm{j}\in \mathrm{\delta }\mathrm{L}}{\mathrm{w}}_{\text{ij}}{\left({\mathrm{y}}_{\mathrm{i}}-{\overline{\mathrm{y}}}_{\mathrm{L}}\right)}^2+\sum _{\mathrm{j}\in \mathrm{\delta }\mathrm{R}}{\mathrm{w}}_{\text{ij}}{\left({\mathrm{y}}_{\mathrm{j}}-{\overline{\mathrm{y}}}_{\mathrm{R}}\right)}^2$$

δL and δR denote the left and right subsets generated by candidate splits, and ${\overline{\mathrm{y}}}_{\mathrm{L}}$ and ${\overline{\mathrm{y}}}_{\mathrm{R}}$ represent the corresponding weighted means. By adjusting the spatial scale of the neighborhood, weight decay intensity, and forest complexity, this model flexibly captures nonlinear relationships and spatial heterogeneity at local scales. This process is repeated point-by-point across the study area, generating a continuous spatial prediction surface and location-specific local interpretations to reveal spatial patterns and mechanistic differences in global lung cancer risk.

To improve reproducibility, we explicitly report the key hyperparameter settings used in the GWRF model. All predictor variables were standardized using z-scores prior to model fitting to improve comparability across variables with different scales and units. The GWRF was implemented using an adaptive k-nearest-neighbor (kNN) bandwidth with k = 50, and spatial weights were assigned using a Gaussian kernel, where the local bandwidth for each target country was defined as the distance to its 50th nearest neighbor. For the main spatial GWRF-SHAP analysis, n = 187 countries and regions were included. Accordingly, the adaptive bandwidth of k = 50 corresponded to approximately 27% of the included country-level spatial units contributing, with distance-based weighting, to each local model.

The bandwidth was specified using a bias–variance trade-off principle commonly used in geographically weighted modeling: a bandwidth that is too small may lead to unstable local estimation due to insufficient effective observations, whereas a bandwidth that is too large may oversmooth local variation and make the model approach a global estimator. The selected bandwidth (k = 50) was chosen to balance spatial heterogeneity capture and local estimation stability. Model performance and robustness were evaluated through 5-fold cross-validation, comparison with alternative machine-learning models, and comparison of prediction errors between the training and testing sets. Predictive performance was assessed using the coefficient of determination (R²), mean absolute error (MAE), and root mean squared error (RMSE), with detailed results reported in Tables S3–S5.

For the main GWRF estimation, the local random forest used 100 trees (n_estimators = 100) with the default tree-depth setting. Considering that each local model was estimated using approximately 50 weighted observations, this forest size provided a stable ensemble-averaging effect while maintaining computational feasibility. In preliminary testing, increasing the number of trees beyond 100 produced only marginal performance gains but substantially increased computational cost, whereas smaller forests led to higher variance in local predictions.

GWRF-SHAP

Although the GWRF model effectively identifies the importance of each predictor variable, it inherently fails to reveal the direction of these variables' influence. Therefore, this study further incorporates the SHAP method to perform a locally interpretable decomposition of the GWRF model results. SHAP employs game theory to assign a Shapley value to each predictor variable, quantifying both the magnitude and direction of its influence on the dependent variable.⁴⁶ For any country i, its GWRF model prediction can be expressed as:

${\hat{\mathrm{y}}}_{\mathrm{i}}=\text{fGWRF}\left({\mathrm{x}}_{\mathrm{i}1},{\mathrm{x}}_{\mathrm{i}2},\dots ,{\mathrm{x}}_{\text{iK}}\right)$ x_ik denotes the kth explanatory variable, and K represents the total number of variables. The SHAP method defines the average contribution of a single variable x_k to the prediction outcome by calculating its marginal contribution across all possible feature combinations:

$${\mathrm{\varphi }}_{\mathrm{k}}=\sum _{\mathrm{S}\subseteq \mathrm{N}\setminus \left\{\mathrm{k}\right\}}\frac{\left|\mathrm{S}\right|!\left(\left|\mathrm{N}\right|-\left|\mathrm{S}\right|-1\right)!}{\left|\mathrm{N}\right|!}\left[{\mathrm{f}}_{\mathrm{S}\cup \left\{\mathrm{k}\right\}}\left({\mathrm{x}}_{\mathrm{S}\cup \left\{\mathrm{k}\right\}}\right)-{\mathrm{f}}_{\mathrm{S}}\left({\mathrm{x}}_{\mathrm{S}}\right)\right]$$

N denotes the entire feature set, S represents the subset excluding feature k, and f_S(x_S) indicates the expected output when the model uses only feature S. This equation reflects the average marginal contribution of variable x_k to the model’s prediction, while the weighting coefficient ensures fairness and additivity across all feature combinations. The model’s overall prediction result can be expressed as:

$$\mathrm{f}\left({\mathrm{x}}_{\mathrm{i}}\right)={\mathrm{\varphi }}_0+\sum _{\mathrm{k}=1}^{\mathrm{K}}{\mathrm{\varphi }}_{\mathrm{k}}$$

Among these, ϕ₀ represents the baseline value in the absence of feature input, while ϕ_k denotes the SHAP value for feature k, indicating its individual contribution to the prediction outcome.

Building upon the locally specific characteristics of the GWRF model, SHAP values are computed within the local forest model at each geographic location. This yields a locally distributed function of variable importance across the spatial dimension:

$${\mathrm{\varphi }}_{\mathrm{k}}\left({\mathrm{u}}_{\mathrm{i}}\right)=\text{SHAP}\left({\mathrm{f}}_{\text{GWRF}}\left({\mathrm{u}}_{\mathrm{i}}\right),{\mathrm{x}}_{\text{ik}}\right)$$

SHAP analysis not only provides a comprehensive understanding of how each independent variable influences the dependent variable, but also effectively determines the magnitude and direction of these effects for each sample. It quantifies the contribution of each feature to the model’s prediction, making it suitable for heterogeneity analysis. To ensure computational feasibility for local SHAP decomposition while maintaining interpretability, SHAP-based local explanations were generated using a reduced-complexity local forest configuration (20 trees, maximum depth = 8) for the explanation stage. This setting was used only for local attribution calculation and did not replace the main GWRF estimation settings reported above. Data processing, statistical analysis, and visualization in this study were conducted using Python 3.12 and R 4.3.2. The software packages and corresponding versions used in this analysis are listed in the key resources table.