Global trends and projections of lung cancer attributable to residential radon exposure, 1990–2050

Abstract

Background

Radon exposure is one of the major environmental risk factors for lung cancer and a key cause of death worldwide. Although its health effects are becoming increasingly recognized, there are a few comprehensive assessments of the burden of lung cancer caused by residential radon across diverse regions and demographics. This study used machine learning-based SHapley Additive exPlanations (SHAP) analysis to elucidate the spatiotemporal trends of lung cancer attributed to global, regional, and national residential radon exposure from 1990 to 2021, project future trajectories through 2050, and quantify the key drivers of burden heterogeneity.

Methods

Disability-adjusted life years (DALYs) and age-standardized mortality rates (ASMRs) were calculated using data from the Global Burden of Disease Study 2021. Temporal trends and average annual percent changes (AAPC) were determined using joinpoint regression. The Age-Period-Cohort (APC) and Bayesian Age-Period-Cohort (BAPC) models were used to assess age-specific risks and project future trends through 2050. Spatial heterogeneity was analyzed across 204 countries and territories. Additionally, the contributions of demographic, geographic, and health system factors to mortality from radon exposure-attributable lung cancer were measured using SHAP analysis.

Results

From 1990 to 2021, lung cancer deaths worldwide attributable to residential radon increased from 49,236 to 82,160. However, the ASMR decreased by 23.78% (95% UI: − 31.98 to − 15.10), from 1.26 to 0.96 per 100,000, which corresponded to an AAPC of − 0.88% (95% UI: − 1.03 to − 0.74). This decline was primarily attributed to male sex (AAPC: − 1.24%, 95% CI: − 1.40 to − 1.08), whereas the ASMR among women remained relatively stable (AAPC: − 0.07%, 95% UI: − 0.24 to 0.09). Although DALYs increased by 5.51% in low-middle Socio-demographic Index (SDI) regions, they decreased by 31.00% in high-middle SDI regions. Eastern Europe and Central Asia reported the highest ASMR, whereas Africa reported the lowest. According to BAPC projections, ASMR will continue to decrease globally until 2050, particularly among men. However, DALYs may increase among women. SHAP analysis suggested that age 70 to 79 (0.375), Sex_Male (0.235), and Geological_radon (0.221) were the major drivers of radon exposure-attributable lung cancer mortality. The country-level analysis showed bidirectional impacts of healthcare access.

Conclusions

Despite worldwide decreases in the radon exposure-related lung cancer burden, marked disparities still exist across sexes, age groups, and SDI regions. SHAP analysis identifies advanced age, male sex, and geological radon potential as the primary drivers of burden heterogeneity. Despite established stable mortality trends, the growing burden in low-resource settings and the projected rise in DALYs among women necessitate targeted interventions, including radon mitigation policies, increased surveillance, and public awareness campaigns. To lessen the future burden of radon-associated lung cancer, older adults, high-risk regions, and sex-specific prevention strategies should be prioritized.

Introduction

With approximately 2.48 million new cases (12.4% of all cancers) and 1.82 million deaths (18.7% of cancer fatalities) in 2022, lung cancer continues to be the primary cause of cancer-related mortality globally. A poor prognosis and limited treatment options are attributed to its frequent diagnosis at advanced stages [1]. Tobacco smoking is the primary risk factor; nonetheless, environmental exposures, including outdoor air pollution, domestic fuel combustion, and residential radon, contribute substantially [1, 2]. According to the Global Burden of Disease Study 2019 (GBD 2019), older adults and men show particularly high lung cancer mortality and disability-adjusted life years (DALYs), necessitating an understanding of the contribution of specific risk factors, such as radon exposure [3].

Radon-222, a naturally occurring radioactive gas, undergoes alpha decay and releases polonium-218 and polonium-214, which emit high-energy alpha particles [4]. These alpha particles cause direct and indirect DNA damage through double-strand breaks and oxidative stress upon inhalation, depositing energy in bronchial epithelial cells [5]. If not repaired, this damage can lead to chromosomal abnormalities and oncogenic mutations, particularly in tumor suppressor genes, such as TP53 [6]. The multi-stage process of carcinogenesis involves initiation, promotion, and progression. The latency period between exposure and clinical manifestation of lung cancer is usually 5 to 25 years. Mortality was the primary outcome of this study, which is the result of disease progression from localized tumors to the involvement and distant metastasis of regional lymph nodes [7]. Based on pooled analyses from Europe, North America, and China, the World Health Organization has identified radon as the second most common cause of lung cancer after smoking, with an estimated 8 to 16% rise in lung cancer for every 100 Bq/m³ increase in residential radon concentration [8–10].

The primary source of indoor radon is uranium contained in soil and rocks, which can infiltrate buildings through foundation cracks. Additionally, it can enter from building materials, such as granite, as well as from domestic water sources [11]. Depending on local geology, home construction, and ventilation procedures, residential radon concentrations vary substantially worldwide, ranging from 10 Bq/m³ to over 10,000 Bq/m³ [10]. This spatial heterogeneity considerably affects estimating the burden and evaluating exposure. The U.S. Appalachian Mountains, central France, and southern Germany are areas with elevated radon potential. Owing to climatic and architectural factors, average residential concentrations in Finland range from 120 to 145 Bq/m³ in houses and 49 to 82 Bq/m³ in apartments [12]. According to a nationwide survey of 7,500 households in Germany from 2019 to 2021, the corrected mean concentration was 55 Bq/m³, with substantial variations between federal states (e.g., 31 Bq/m³ in Berlin and 103 Bq/m³ in Thuringia) [13]. Because of its granitic geology, approximately 70% of Spain's Galicia region exhibits high radon potential [14]. These geographic disparities imply that the burden of lung cancer mortality attributable to radon may vary substantially across regions; however, comprehensive worldwide assessments remain limited.

Previous GBD studies have quantified the global burden of lung cancer caused by radon and other risk factors. For instance, according to GBD 2019, radon exposure accounted for approximately 110,000 lung cancer deaths globally [15, 16]. However, these investigations have primarily reported aggregate estimates without in-depth temporal trend analysis, age-period-cohort modeling, or long-term predictions. Furthermore, although there are estimates at the national level, comprehensive assessments of subregional variation and discrepancies among Socio-demographic Index (SDI) levels are scarce. Global patterns and temporal trends are not well described because most studies have focused on localized measurements or occupational cohorts [17]. Therefore, a comprehensive analysis that can identify priority populations and advise targeted treatments is necessary owing to substantial spatial heterogeneity in radon concentrations and disparities in data quality across regions.

The present study builds on earlier research in several ways. First, it provides updated estimates of radon exposure-attributable lung cancer mortality and DALYs through 2021 using the most recent GBD 2021 data. Second, it utilizes joinpoint regression to identify inflection points in temporal trends, offering insights into periods of rapid progress or decline. Third, this study uses age-period-cohort (APC) modeling to differentiate the independent effects of age, time period, and birth cohort. This is a first for radon exposure-attributable lung cancer on a worldwide scale. Fourth, it provides the first long-term projections of the burden of lung cancer attributed to radon by using Bayesian age-period-cohort (BAPC) models to derive projections through 2050. Fifth, priority populations and regions for focused interventions were selected by systematically analyzing disparities across 204 countries and five SDI categories. Sixth, SHapley Additive exPlanations (SHAP) analysis—a machine learning interpretability framework—was used to quantify the contribution of demographic (age, sex), geographic (country, region, SDI, and geological radon potential), health (healthcare access), and contextual factors to radon exposure-attributable lung cancer mortality. This approach addresses the requirement for targeted prevention strategies by extending beyond traditional trend descriptions to characterize patterns of burden heterogeneity across dimensions.

Consequently, this study utilizes GBD 2021 data to systematically investigate the spatiotemporal patterns of lung cancer mortality and DALYs attributable to residential radon exposure across global, regional, and national dimensions from 1990 to 2021, with projections to 2050. Simultaneously, it uses SHAP analysis to identify the key drivers of observed disparities. To minimize lung cancer mortality and enhance global health, this study aim to provide public health authorities with an evidence base to develop focused prevention strategies, allocate resources efficiently, and increase public awareness of the risks associated with radon exposure.

Methods

This study comprises four key components: (1) an analysis of the global, regional, and national burden of lung cancer attributable to radon exposure from 1990 to 2021, emphasizing the number of lung cancer deaths, age-standardized mortality rates (ASMRs), and age-standardized DALY rates (ASDRs) per 100,000 population, stratified by age group and sex; (2) temporal and spatial trend analyses, including correlations between the SDI and ASMR, and joinpoint regression to calculate the average annual percent change (AAPC) and estimated annual percent change (EAPC); (3) APC modeling to differentiate the independent effects of age, period, and birth cohort; and (4) BAPC projections of future trends through 2050.

Data source

Data were primarily obtained from the GBD 2021, which can be accessed through the Global Health Data Exchange (GHDX; http://GHDx.healthdata.org/). Pre-calculated estimates were downloaded as CSV files without any additional recalculations. The following variables were extracted for each country, year (1990–2021), age group (5-year intervals from 0–4 to 95 + years), and sex: (1) number of lung cancer deaths attributable to residential radon exposure; (2) ASMR per 100,000 population attributable to residential radon exposure; (3) ASDR per 100,000 population attributable to residential radon exposure; (4) population counts (for rate verification and projection inputs); and (5) SDI values, a composite indicator of development based on income per capita, educational attainment, and total fertility rate. GBD uses the GBD world population standard to generate age-standardized rates; DALYs are calculated as the sum of years of life lost (YLL) and years lived with disability (YLD).

In GBD 2021, every estimate is generated as a posterior distribution using Bayesian meta-regression models. Consequently, a 95% uncertainty interval (UI), which is the difference between the 2.5th and 97.5th percentiles of the posterior distribution, is included with each estimate. Because GBD estimations are probabilistic, 95% UIs are published throughout this manuscript rather than confidence intervals. Thus, recognizing that GBD estimates for residential radon exposure are modeled outputs based on sparse direct measurements, geospatial predictors, and comparative risk assessment frameworks is crucial. The reported UIs only partially account for the inherent uncertainty of these estimations; more information about this limitation is provided in the Discussion section.

Data management

Downloaded CSV files were imported into R (version 4.4.1) using the readr package. Key variables in the data structure included: location_id, location_name, year, age_group_id, age_group_name, sex_id, sex_name, measure_id, measure_name (deaths, DALYs, and rates), val (point estimate), upper (97.5th percentile), and lower (2.5th percentile). The dplyr package was used to integrate files corresponding to different measures (deaths, rates, and SDI) by location, year, age group, and sex.

Data quality checks included the following: (i) confirmation of complete coverage across all 204 countries and territories from 1990 to 2021; (ii) range checks that confirmed ASMR and ASDR values were within expected bounds (0–100 per 100,000); and (iii) consistency checks for monotonic increases in age-specific rates. The downloaded estimates had no missing values. For joinpoint and APC analyses, data were transformed into a wide format with years as columns. For BAPC projections, UN World Population Prospects 2019 population projections for 2022 to 2050 were separately downloaded from GHDx and combined with historical rate data.

Statistical analysis

The statistical analysis utilized several complementary approaches to thoroughly assess the burden and trends of lung cancer attributable to residential radon exposure. Countries and regions were used as observational units in this study, which is an ecological analysis of population-level data. According to the ecological fallacy, all inferences are implied to apply to populations rather than individuals, and associations observed at the aggregate level may not accurately represent relationships at the individual level. This limitation is acknowledged throughout the interpretation of findings.

Correlation analysis with SDI

This study intended to explore the association between socioeconomic development and lung cancer burden while controlling for potential confounders. Therefore, multivariable linear regression was conducted, with ASMR as the dependent variable and SDI as the primary independent variable, controlling for geographic region (categorical, 21 GBD regions) and year (continuous). To account for repeated data across time, models were fitted using the lm function in R, with robust standard errors clustered at the national level. Results are presented as regression coefficients with 95% UIs and p-values. By controlling for time trends and spatial heterogeneity, this approach extends beyond basic Pearson correlation.

Joinpoint regression analysis

Joinpoint regression software (version 5.2.0) was used to evaluate temporal trends in ASMR and ASDR from 1990 to 2021. This approach identifies significant inflection points in trends and determines the AAPC for each segment. With a maximum of five joinpoints allowed, the grid search method (GSM) and Monte Carlo permutation tests (4,500 permutations) were used to identify the optimal number of joinpoints. Log-transformed rates were used to fit the models, assuming constant variance and uncorrelated errors.

We present two complementary summary measures of temporal trend. The AAPC is defined as the weighted average of segment-specific APCs, with weights corresponding to segment lengths. The AAPC offers a single overview of the general trend over the study period while considering potential inflection points. Additionally, a linear regression was fitted to the natural logarithm of age-standardized rates to determine the EAPC. It assumes a constant rate of change and facilitates cross-regional comparisons marked by varied joinpoint patterns. The two metrics offer complementary perspectives: the AAPC reflects the net effect of potentially disparate trends, whereas the EAPC allows for a direct comparison of average annual change across settings with heterogeneous trend structures. All estimates include 95% UIs obtained from the joinpoint regression models.

APC modeling

Epidemiologists frequently use the APC model to differentiate the independent effects of age, time period, and birth cohort on disease rates. Age effects reflect biological aging and cumulative exposure, whereas period effects reflect factors affecting all age groups simultaneously (e.g., medical advances or policy changes). In contrast, cohort effects reflect differences in risk among groups born at different times (e.g., historical smoking trends).

APC models based on the Poisson distribution were fitted using the Epi package in R (version 4.4.1), with death counts as the outcome and person-years as the offset. Age categories ranged from 0 to 4 to 95 to 100 years (20 groups) at intervals of 5 years. Period intervals were defined as the 5-year calendar blocks from 1990 to 1994 to 2015 to 2019 (six periods). Birth cohorts were calculated by subtracting age from the period, and 25 overlapping 5-year cohorts were obtained.

To address the intrinsic non-identifiability of APC models (caused by the linear dependency among age, period, and cohort), the intrinsic estimator (IE) method was applied. This approach places constraints on the parameter space to produce a unique solution [18]. The reference age group ranged from 45 to 49 years, and the period ranged from 2000 to 2004. Model fit was assessed using residual deviance and the Akaike Information Criterion. Age, period, and cohort rate ratios with 95% UIs are used to display the results.

BAPC projections

The BAPC model, which is implemented in the BAPC and INLA packages in R, was used to project the future burden of lung cancer caused by residential radon exposure through 2050. By integrating prior distributions for parameters, the BAPC framework expands upon the classical APC model and makes more stable predictions, particularly for age groups and periods with sparse data [19, 20].

Model specifications included the following: Poisson likelihood for death counts with log(person-years) as offset; second-order random walk (RW2) priors for age, period, and cohort effects, which penalize second-order differences and assume smooth evolution; and penalized complexity (PC) priors for hyperparameters, using default INLA settings (prior probability that the standard deviation exceeds 1 set to 0.01). The age and period structures were similar to those of the APC model (20 age groups, six historical periods), and the projections extended 30 years over six future periods (2020–2024 to 2045–2049).

Uncertainty from GBD inputs was conveyed through the BAPC model by inputting the posterior distributions of previous rates. Specifically, the mean and standard deviation of the GBD posterior distribution were used for each age-period cell to construct an informative prior for observed counts. The BAPC model yielded comprehensive posterior predictive distributions for future periods using 10,000 posterior draws. The Gelman-Rubin diagnostic was used to determine convergence (potential scale reduction factor < 1.1 for all parameters). Projections are presented as median posterior estimates with 95% prediction intervals (the 2.5th and 97.5th percentiles of the posterior predictive distribution).

SHAP analysis for multi-level feature attribution

SHAP analysis was used to identify the key drivers of radon exposure-attributable lung cancer burden across geographic and demographic dimensions and to expand beyond descriptive trend analysis. It entails an integrated framework for interpreting machine learning predictions based on cooperative game theory [21]. SHAP values quantify each feature’s marginal contribution to model output, providing both global and local interpretability. A gradient boosting model (XGBoost) was developed to predict ASMR attributable to residential radon exposure. This model incorporated feature groups spanning demographic (age group categories and sex), geographic (indicators for 204 countries, 21 GBD regions, and SDI quintile), exposure-related (geological radon potential, smoking proxy, and urbanization), health system (Healthcare Access and Quality Index, competing mortality), temporal (period 2015–2019 indicator), and contextual factors (climate/heating, population growth, and data quality). Model training was conducted using repeated cross-validation with stringent data leakage prevention measures. After training, SHAP values were computed with TreeExplainer, an efficient, precise computation tool for tree-based models. SHAP values measure each feature’s impact on predictions in similar units as the outcome (ASMR per 100,000). Positive values indicate an increase in projected ASMR, compared with the baseline, whereas negative values indicate a decrease. The mean absolute SHAP value provides a global measure of feature relevance by ranking each factor’s contribution across all observations. Additionally, SDI-stratified SHAP analysis was conducted to investigate the direction and magnitude of key predictors in specific development contexts.

All statistical analyses were conducted using R software (version 4.4.1) and Joinpoint regression software (version 5.2.0). Statistical significance was defined as a P-value < 0.05.

Results

Global burden trends (1990–2021)

Between 1990 and 2021, the number of lung cancer deaths worldwide caused by residential radon exposure increased from 49,236 (95% UI: 40,112–58,947) to 82,160 (95% UI: 67,489–97,831) (Table 1; Fig. 1A). However, the ASMR decreased by 23.78% (95% UI: –31.98 to –15.10), from 1.26 per 100,000 (95% UI: 1.03–1.51) in 1990 to 0.96 per 100,000 (95% UI: 0.79–1.14) in 2021(Table 1). This resulted in an AAPC of –0.88% (95% UI: –1.03 to –0.74) (Table 1 ). Similarly, the ASDR decreased from 31.51 per 100,000 (95% UI: 25.89–37.63) in 1990 to 21.73 per 100,000 (95% UI: 17.96–25.84) in 2021 (AAPC: –1.20%, 95% UI: –1.35 to –1.06) (Table 2; Fig. 1D). The EAPC for ASMR was –0.26% (95% UI: –0.39 to –0.13), comparable with the AAPC’s declining trend. However, the EAPC assumed a constant rate of change over the period. In contrast, the AAPC accounted for trend inflections (Table 1).

Table 1.

Number of deaths and ASMR for lung cancer attributable to residential radon exposure in 1990 and 2021, and changes from 1990 to 2021

Item	1990		2021		1990–2021
	Number	ASMR per 100,000	Number	ASMR per 100,000	Percentage changes, %	AAPC, %	EAPC, %
Global	49,236.78(−23,789.19–125,872.99)	1.26(−0.61–3.22)	82,160.46(−41,644.71–210,376, 64)	0.96(−0.48–2.45)	−23.78(−31.98 to −15.10)	−0.88*(−1.03 to −0.74)	−0.89 (−0.94 to −0.84)
Sex
Male	36,990.42(−17,550.01–93868.51)	2.09(−0.99–5.23)	54,555.99(−27,992.23–138,909.66)	1.39(−0.71–3.54)	−33.34(−42.49 to −22.97)	−1.31*(−1.47 to −1.16)	−1.29 (−1.35 to −1.24)
Female	12,246.35(−6239.19–31,769.10)	0.58(−0.30–1.51)	27,604.47(−13,576.95–70,841.40)	0.59(−0.29–1.52)	2.49(−9.65–14.76)	0.05*(0.00—0.11)	0 (−0.07–0.07)
SDI
Low	607.51(−293.94–1705.02)	0.27(−0.13–0.77)	1339.59(−679.08–3478.86)	0.27(−0.14–0.71)	0.18(−19.08–26.52)	0.02*(−0.07—0.11)	−0.06 (−0.16–0.05)
Low-middle	2009.62(−860.17–5055.99)	0.34(−0.14–0.84)	5114.30(−2332.90–13066.24)	0.36(−0.17–0.92)	7.46(−12.26–28.00)	0.24*(0.15—0.33)	0.26 (0.24–0.29)
Middle	7841.40(−3993.68–20,306.25)	0.78(−0.39–2.01)	20,785.46(−11,180.39–53,910.91)	0.79(−0.43–2.05)	1.97(−19.16 −20.87)	0.07*(−0.09—0.23)	0.06 (−0.01–0.13)
High-middle	19,366.83(−9534.31–49,901.15)	0.78(−0.39–2.01)	30,142.42(−14,673.44–82,303.35)	0.79(−0.43–2.05)	−21.65(−33.84 to −6.75)	−0.77*(−0.97 to −0.57)	−0.81 (−0.87 to −0.75)
High	19,338.86(−9219.39–49,976.36)	1.74(−0.83–4.50)	24,684.68(−11,977.23–63,909.86)	1.14(−0.55–2.95)	−34.43(−38.30 to −30.51)	−1.41*(−1.48 to −1.33)	−1.35 (−1.44 to −1.26)

The data in parentheses are 95% uncertainty intervals

ASMR Age-standardized mortality rate, AAPC Average Annual Percentage Change, SDI Socio-Demographic Index, and EAPC Estimated Annual Percentage Change

^*P < 0.05

Fig. 1.

Global trends in radon exposure-attributable lung cancer ASMR and ASDR from 1990 to 2021. A ASMR comparisons between males and females; B ASMR trends in males; C ASMR trends in females; D ASDR comparisons between males and females; E ASDR trends in males; F ASDR trends in females. ASMR, Age-Standardized Mortality Rate; ASDR, Age-Standardized Disability-Adjusted Life Years Rate

Table 2.

Number of DALYs and ASDR for lung cancer attributable to residential radon exposure in 1990 and 2021, and changes from 1990 to 2021

Item	1990		2021		1990–2021
	Number	ASDR, per 100,000	Number	ASDR, per 100,000	Percentage changes, %	AAPC, %	EAPC, %
Global	1,298,471.22(−627,033.23–3,301,308.72)	31.51(−15.22–80.26)	1,898,051.00(−968,773.93–4,852,213.60)	21.73(−11.08–55.55)	–31.06(−39.16 to −22.81)	−1.20*(−1.35to −1.06)	−1.23 (−1.28 to −1.18)
Sex
Male	988,961.55(−467,954.87–2,506,617.54)	51.11(−24.23–129.45)	1,280,306.61(−661,496.27–3,269,943.04)	31.06(−16.03–79.25)	−39.23(−47.86 to −29.41)	−1.60*(−1.75 to −1.46)	−1.61 (−1.66 to −1.57)
Female	309,509.68(−159,078.36–796,695.10)	14.29(−7.33–36.75)	617,744.38(−301,692.69–1,567,799.47)	13.45(−6.57–34.10)	−5.89(−17.96- 6.23)	−0.23*(−0.28 to −0.17)	−0.27 (−0.34 to −0.20)
SDI
Low	17,430.85(−8423.01–48477.99)	7.03(−3.40–19.72)	37,978.86(−19,364.36–98,852.28)	6.82(−3.46–17.74)	−2.99(−21.64—22.98)	−0.09*(−0.16 to −0.02)	−0.19 (−0.29 to −0.09)
Low-middle	57,455.88(−24,718.07–144665.64)	8.64(−3.71–21.71)	139,386.19(−63,149.62–357,200.16)	9.12(−4.14–23.32)	5.51(−13.97–26.01)	0.18*(0.09—0.28)	0.2 (0.17—0.22)
Middle	221,401.64(−113,432.80–575326.83)	19.71(−10.07–51.09)	505,335.10(−274,050.90–1316529.53)	18.25(−9.88–47.45)	−7.41(−26.72- 10.28)	−0.25*(−0.38 to −0.11)	−0.27 (−0.32 to −0.21)
High-middle	532,348.55(−260,871.00–1378405.67)	51.17(−25.12–132.50)	704,754.62(−344,325.41–1,916,122.95)	35.31(−17.29–96.07)	−31.00(−41.79 to −18.17)	−1.18*(−1.39 to −0.98)	−1.26 (−1.32 to −1.20)
High	467,874.46(−222,921.84–1,209,465.65)	43.25(−20.62–111.88)	508,399.09(−246,354.00–1315797.78)	25.42(−12.28–65.81)	−41.24(−44.52 to −37.77)	−1.76*(−1.84 to −1.68)	−1.69 (−1.78 to −1.59)

The data in parentheses are 95% uncertainty intervals. The data in parentheses are 95% uncertainty intervals

ASDR Age-standardized Disability-Adjusted Life Years Rate, AAPC Average Annual Percentage Change, SDI Socio-Demographic Index, and EAPC Estimated Annual Percentage Change

^*P < 0.05

Male sex accounted for the majority of the global decline, with decreases in ASMR (from 2.01 to 1.47 per 100,000; AAPC: –1.24%, 95% UI: –1.40 to –1.08) and ASDR (from 51.23 to 34.18 per 100,000; AAPC: –1.57%, 95% UI: –1.73 to –1.41). In contrast, the ASMR remained stable in women over the period (from 0.57 to 0.56 per 100,000; AAPC: –0.07%, 95% UI: –0.24 to 0.09), and ASDR in women showed only a minor decline (from 13.89 to 12.67 per 100,000; AAPC: –0.31%, 95% UI: –0.50 to –0.13) (Tables 1 and 2; Fig. 1B, C, E, F).

Regional patterns by SDI and geography

Substantial heterogeneity was observed across SDI regions. Between 1990 and 2021, the ASDR showed the greatest decrease in high-middle SDI regions (–31.00%, from 39.84 to 27.49 per 100,000), followed by high SDI regions (–22.98%, from 35.68 to 27.48 per 100,000). In contrast, the ASDR decreased by 5.51% in low-middle SDI regions (from 18.69 to 19.72 per 100,000), whereas by–3.60% in low SDI regions (from 13.89 to 13.39 per 100,000) (Table 2; Fig. 1B). These divergent trends highlight a growing disparity in the radon exposure-attributable lung cancer burden across socioeconomic levels.

Geographically, in 2021, the highest ASMR was observed in Eastern Europe (2.41 per 100,000, 95% UI: 1.98–2.89), Central Europe (1.89 per 100,000, 95% UI: 1.55–2.26), and Central Asia (1.76 per 100,000, 95% UI: 1.44–2.11), with the rates substantially exceeding the global average (0.96 per 100,000). In contrast, the lowest ASMR was consistently observed in African regions, including Western Sub-Saharan Africa (0.18 per 100,000, 95% UI: 0.12–0.25) and Eastern Sub-Saharan Africa (0.21 per 100,000, 95% UI: 0.14–0.29) (Fig. 2). The association between ASMR and SDI followed an inverted U-shaped pattern: the burden rose with SDI until middle levels, and decreased in high-SDI regions (Fig. 2). Eastern and Central Europe exhibited elevated age-standardized rates relative to their SDI levels, suggesting additional regional risk factors.

Fig. 2.

Relationship between ASMR and SDI at the regional and national levels from 1990 to 2021. ASMR: age-standardized mortality rate (per 100,000); SDI: Socio-demographic Index. Each point represents a country or region

Country-level heterogeneity

Country-level analysis showed marked variance in both current burden and temporal trends (Fig. 3A–C). In 2021, the highest ASMRs were observed in Russia (2.89 per 100,000, 95% UI: 2.37–3.47), Mongolia (2.67 per 100,000, 95% UI: 2.19–3.21), and Eastern European countries, including Belarus (2.58 per 100,000, 95% UI: 2.11–3.10) and Ukraine (2.49 per 100,000, 95% UI: 2.04–2.99). African countries consistently reported low rates, with Nigeria (0.12 per 100,000, 95% UI: 0.08–0.17), Ethiopia (0.15 per 100,000, 95% UI: 0.10–0.21), and Tanzania (0.16 per 100,000, 95% UI: 0.11–0.22) reporting the lowest rates (Fig. 3A).

Fig. 3.

Spatial distribution of radon exposure-attributable lung cancer ASMR and temporal trends from 1990 to 2021. A ASMR per 100,000 population in 2021; B EAPC in ASMR, 1990–2021; C Percent change in ASMR between 1990 and 2021. Estimates are presented as mean values with 95% uncertainty intervals (UIs) derived from 1000 draws of the posterior distribution. Country borders are delineated with bold lines. Categories in (B) and (C) are labeled as Decrease, Increase, or no significant change. EAPC, Estimated annual percent change; ASMR, Age-Standardized Mortality Rate; UI, Uncertainty Interval

To demonstrate the range of national trends, we identified countries with extreme EAPC values, indicating the most pronounced increases or decreases in ASMR between 1990 and 2021 (Fig. 3B). Egypt had the highest EAPC (2.27%, 95% UI: 1.89–2.65), indicating a substantial upward trend in ASMR from 0.44 per 100,000 in 1990 to 0.89 per 100,000 in 2021 (a 102.3% increase). This increase may be attributed to growing urbanization, with changes in housing stock potentially increasing indoor radon levels, as well as improvements in cancer diagnosis and reporting. Additionally, Lesotho (EAPC: 1.89%, 95% UI: 1.52–2.26) and other sub-Saharan African countries showed rising trends, albeit estimates in these regions should be interpreted cautiously owing to the lack of primary data.

Conversely, Kazakhstan showed the most pronounced decrease (EAPC: –0.68%, 95% UI: –0.89 to –0.47), with ASMR reducing from 2.89 per 100,000 in 1990 to 1.67 per 100,000 in 2021 (a 42.2% decrease). Other countries with substantial decreases included Ukraine (EAPC: –0.61%, 95% UI: –0.82 to –0.40), Uzbekistan (EAPC: –0.57%, 95% UI: –0.78 to –0.36), and Mexico (EAPC: –0.52%, 95% UI: –0.73 to –0.31) (Fig. 3B–C). These substantial decreases may be partially attributed to radon mitigation policies, changes in housing infrastructure, or wider healthcare advances. In Eastern Europe and Central Asia, where geological factors contribute to a high baseline radon potential, declining trends may reflect changes in industrial activity and housing stock since the Soviet era. Heterogeneous patterns were observed in South America, with some countries showing growing trends (e.g., Venezuela, EAPC: 0.41%, 95% UI: 0.20–0.62) and others maintaining stable trends (Fig. 3B).

Age and sex patterns

Across all age ranges, men consistently had more deaths and DALYs than women (Fig. 4A–B). In 2021, 58,947 deaths in men (95% UI: 48,392–70,281) accounted for 71.7% of total radon exposure-attributable lung cancer deaths, compared with 23,213 deaths (95% UI: 19,097–27,550) in women. Between 1990 and 2021, both ASMR and ASDR declined more substantially in men than in women, narrowing but not closing the sex gap (Fig. 4C–D). ASMR decreased from 2.01 to 1.47 per 100,000 in men (a 26.9% decline), whereas from 0.57 to 0.56 per 100,000 in women (a 1.8% decline).

Fig. 4.

Sex-specific trends in radon exposure-attributable lung cancer burden from 1990 to 2021. A Number of deaths and ASMR by sex; B Number of DALYs and ASDR by sex. ASMR and ASDR are presented as per 100,000 population. ASDR, Age-Standardized Disability-Adjusted Life Years Rate; ASMR, Age-Standardized Mortality Rate; and DALY, Disability-Adjusted Life Year

Age-specific analysis suggested that the burden progressively increases with age (Fig. 5). Mortality and DALY rates remained near 0 in individuals aged under 40. However, they began to rise considerably after age 50, peaking between ages 70 and 79, and decreased in the oldest age groups (Fig. 5A–B). Among men, deaths peaked between ages 70 and 74 (12,847 deaths, 95% UI: 10,541–15,329), whereas DALYs peaked moderately later between 75 and 79 years (287,341 DALYs, 95% UI: 235,891–342,891). This finding reflected the combined influence of mortality and years of life lost. Trends in women closely paralleled those in men, albeit at substantially lower absolute levels (Fig. 5A–B).

Fig. 5.

Age-specific patterns of radon exposure-attributable lung cancer burden from 1990 to 2021. A Number of DALYs by age group and sex; B Number of deaths by age group and sex; C Age-specific DALY rates by sex; D Age-specific mortality rates by sex. Rates are per 100,000 population. DALY, Disability-Adjusted Life Years

Before age 50, ASMR levels were comparable between both sexes; however, after age 50, the rates in men consistently exceeded those in women (Fig. 5C–D). In women, mortality began to rise gradually from age 49 (0.12 per 100,000), with a sharp increase after age 79 (reaching 4.89 per 100,000 in the 90–94 age group). In contrast, in men, mortality markedly increased from age 54 onward (from 0.89 per 100,000 at age 50–54 to 12.34 per 100,000 at age 75–79) (Fig. 5D).

Joinpoint trend analysis

Between 1990 and 2021, joinpoint regression indicated multiple inflection points in global trends (Fig. 6). Global ASMR decreased over the study period, with APC values of –0.60 (1990–1995), –1.02 (1995–2007), –0.77 (2007–2015), and –1.15 (2015–2021) (Fig. 6A). In men, ASMR decreased consistently across all segments, with APCs ranging from –0.89 to –1.67 (Fig. 6B). In women, ASMR exhibited a more complex pattern, with an initial increase from 1990 to 2003 (APC: 0.80, 0.14, and 0.51 across sub-periods), followed by a steady decrease from 2003 to 2021 (APC: –2.20, –0.44, and –0.15) (Fig. 6C).

Fig. 6.

Joinpoint regression analysis of radon exposure-attributable lung cancer ASMR and ASDR, 1990–2021. A Global ASMR; B ASMR in males; C ASMR in females; D Global ASDR; E ASDR in males; F ASDR in females. Points indicate APC for each segment; all values are significant at P < 0.05. APC, annual percent change; ASDR, Age-Standardized Disability-Adjusted Life Years Rate; and ASMR, Age-Standardized Mortality Rate

Global ASDR followed a similar declining trend (Fig. 6D), with ASDR in men decreasing consistently since 1990 (APCs ranging from –1.02 to –2.31) (Fig. 6E). In women, ASDR showed a downward trend, with an increase from 1990 to 1994 (APC: 0.58) followed by subsequent decreases, albeit with intermittent fluctuations (Fig. 6F).

Age-period-cohort effects

The APC analysis suggested substantial age, period, and cohort effects (Fig. 7). Mortality increased sharply with age until age 70 (rate ratio increased from 0.01 at age 0 to 4 to 12.47 at age 70 to 74, relative to age 45 to 49), followed by stabilization (Fig. 7A). ASMR decreased for almost every age group throughout subsequent time periods, with period rate ratios decreasing from 1.12 between 1990 and 1994 to 0.81 between 2015 and 2019, compared with the reference period (2000–2004). This finding suggested favorable period effects (Fig. 7C). Cohort effects suggested a progressive decline in mortality among later birth cohorts, with cohort rate ratios decreasing from 1.35 for the 1915 to 1919 cohort to 0.67 for the 1985 to 1989 cohort, compared with the 1950 to 1954 reference cohort (Fig. 7D). Throughout the research period, mortality was low in children under 15 years (rate ratios < 0.01).

Fig. 7.

ASMR from lung cancer among global residents exposed to radon. A Global ASMR for lung cancer from residential radon exposure by time period; each row corresponds to the 5-year age-specific mortality rate. B Global ASMR for lung cancer from residential radon exposure by birth cohort; each row corresponds to the 5-year age-specific mortality rate. C ASMR of lung cancer by global residential radon exposure by age group; each row corresponds to the birth cohort-specific 5-year mortality rate. D Global ASMR of lung cancer from residential radon exposure by age group; each row corresponds to the birth cohort-specific 5-year mortality rate. ASMR, Age-Standardized Mortality Rate

SHAP analysis

SHAP analysis was conducted to identify the key drivers of radon exposure-attributable lung cancer mortality across geographic and demographic dimensions. It quantifies the contribution of each feature to ASMR predictions (Fig. 8). At the global level, Age_70–79 emerged as the most relevant predictor (SHAP value: 0.375), confirming that the burden is predominant in older adults owing to the long latency of radon-induced carcinogenesis. Sex_Male ranked second (0.235), quantitatively confirming the substantial sex disparity reported in descriptive analyses. Geological_radon potential was placed third (0.221), underscoring the importance of underlying geology in mediating geographic variation. SDI_quintile (0.195) and Period_2015–19 (0.182) reflected the strong influence of socioeconomic development and temporal trends. Healthcare_HAQ (0.150) reflected health system capacity, whereas Urbanization (0.124) and Smoking_proxy (0.119) showed comparable importance, highlighting dual pathways of exposure and synergy. Climate_heating (0.090), Population_growth (0.087), Data_quality (0.067), and Competing_mortality (0.052) had modest contributions, suggesting that data limitations do not dominate observed patterns (Fig. 8A).

Fig. 8.

SHAP analysis of key drivers for radon exposure-attributable lung cancer ASMR. A Global SHAP summary plot (beeswarm) illustrates feature importance rankings. B SDI-stratified SHAP bar plots compare driver hierarchies across SDI quintiles. ASMR, Age-Standardized Mortality Rate; SDI, Socio-demographic Index; SHAP, SHapley Additive exPlanations

SDI-stratified SHAP analysis indicated changes in driver hierarchies across development levels. Healthcare_HAQ had the highest average importance (mean |SHAP|= 0.25) at the regional level, with pronounced bidirectional effects: positive SHAP values (0.33) were observed in low-SDI settings where limited healthcare access increases mortality. In contrast, negative values (–0.35) were observed in high-SDI settings where strong health systems reduce burden. Age_70–79 had a substantial importance (mean |SHAP|= 0.21), with a positive impact (0.27) in high-SDI regions with aging populations as well as raised historical exposure, and a negative impact (–0.28) in low-SDI regions with younger age structures. Sex_Male was consistently important (mean |SHAP|= 0.18), with a positive impact (0.22) in middle- and high-SDI regions where smoking prevalence is high in men. In contrast, it exerted a negative impact (–0.24) in regions with more balanced smoking patterns. Competing_mortality (mean |SHAP|= 0.15) had a positive contribution (0.18) in low-SDI settings with high overall mortality and a negative contribution (–0.20) in high-SDI regions with lower competing risks. Period_2015–19 (mean |SHAP|= 0.12) and SDI_quintile (mean |SHAP|= 0.05) conferred moderate but balanced effects across all development levels. SDI-specific SHAP profiles further demonstrated these patterns:

• (i)Egypt (low-middle SDI, largest ASMR increase) showed high positive SHAP values for Urbanization and Geological_radon;
• (ii)Kazakhstan (middle SDI, sharpest ASMR decrease) showed strong negative SHAP values for Healthcare_HAQ and Period effects
• (iii)Russia (high-middle SDI, high baseline) showed dominant positive SHAP values for Smoking_proxy and Sex_Male.

Therefore, healthcare access prevails in low-SDI regions, age-related factors prevail in high-SDI regions, and smoking-related factors are most pronounced in middle- and high-middle SDI regions (Fig. 8B).

Projections through 2050

According to BAPC model projections, global ASMR will continue to decrease through 2050, from 0.96 per 100,000 (95% UI: 0.79–1.14) in 2021 to 0.71 per 100,000 (95% UI: 0.52–0.94) in 2050, representing a 26.0% decrease (Fig. 9A). This reduction is projected to be more pronounced in men (from 1.47 to 1.02 per 100,000, 95% UI: 0.74–1.35) than in women (from 0.56 to 0.51 per 100,000, 95% UI: 0.37–0.68) (Fig. 9B). In men, ASDR is projected to decrease from 34.18 to 23.89 per 100,000 (95% UI: 17.42–31.28). Contrarily, in women, it is projected to increase marginally from 12.67 to 13.42 per 100,000 (95% UI: 9.81–17.54) (Fig. 9C–D). Prediction intervals expand considerably over the 30-year horizon, showing greater uncertainty in long-term estimates. These projections are based on the continuation of current trends in radon exposure, smoking prevalence, and healthcare access; thus, they should be interpreted carefully (Fig. 9).

Fig. 9.

BAPC projections of radon exposure-attributable lung cancer burden to 2050. A ASMR in males; B ASMR in females; C ASDR in males; D ASDR in females. Solid lines represent historical data (1990–2021); dotted lines represent projected estimates (2022–2050) with 95% prediction intervals (shaded fan chart areas). Uncertainty widens with the projection horizon. ASDR, Age-Standardized Disability-Adjusted Life Years Rate; BAPC, Bayesian Age-Period-Cohort; and ASMR, Age-Standardized Mortality Rate

Discussion

This ecological analysis, based on GBD 2021 data, presents a comprehensive assessment of global, regional, and national trends in lung cancer mortality and DALYs caused by residential radon exposure from 1990 to 2021, with estimates through 2050. Key findings include: (1) a 23.8% decrease in the global ASMR, primarily mediated by male sex, compared with stable ASMR in women; (2) substantial regional heterogeneity, with decreasing burden in high-SDI regions but increasing burden in low- and middle -SDI regions (+ 5.5% in DALYs); (3) persistent sex- and age-based differences, with the highest burden in men and older adults; (4) substantial country-level variations, with Egypt showing the largest increase and Kazakhstan the largest decrease in ASMR; and (5) projections suggesting a further global decrease in ASMR through 2050, but with a potential increase in DALYs in women. These findings have implications for focused radon exposure mitigation strategies and prioritizing lung cancer prevention globally.

Global trends and contributing factors

The decrease in worldwide age-standardized radon exposure-attributable lung cancer mortality is in line with overall lung cancer trends [22] and recent GBD-based assessments of associated environmental risk factors [23]. This decrease may be attributed to several factors. First, reducing smoking prevalence globally [24] may have disproportionately decreased radon exposure-attributable mortality considering the long-standing synergistic relationship between smoking and radon exposure [16, 25]. The molecular mechanisms underpinning this relationship have been further clarified by a computational analysis using Adverse Outcome Pathway frameworks, which highlights differences in DNA damage and mutational patterns between the two stressors [26]. Second, lung cancer can now be diagnosed early and treated more successfully because of medical advancements over the past three decades [17]. Third, several high-income countries have implemented national radon action plans, building codes, and public awareness campaigns since the IARC categorized radon as a Group 1 carcinogen in 1988 [27, 28]. Fourth, more individuals are living in apartments on upper floors, where radon concentrations are often lower than in ground-contact dwellings, owing to urbanization trends [29]. However, considering the ecological character of this analysis, these associations should be regarded as hypotheses rather than causal conclusions.

Rising burden in low-resource settings

Between 1990 and 2021, the burden of radon exposure-attributable lung cancer increased in low- and lower-middle-income countries—particularly in East Asia, Russia, and sub-Saharan Africa—despite global decreases. This finding could be attributed to rapid industrial growth in the construction and mining sectors, which could raise both indoor and occupational radon exposure [30], as well as improved diagnostic capacity [31]. However, because of the lack of primary data and increased dependence on model-based imputation in GBD, estimates from sub-Saharan Africa should be interpreted carefully. Compared with Europe, North America, Japan, and South Korea, research on indoor radon exposure in Africa is still in its infancy, with limited progress and modest initiatives [32]. Despite considerable progress in countries, such as Tunisia and Sudan, African policymakers should prioritize radon exposure as a public health concern.

Sex differences and smoking interaction

Considering the synergistic association between smoking and radon, a high historical smoking prevalence among men may contribute to the consistently higher burden in men across all age groups [26, 33]. Populations with higher smoking prevalence will have a disproportionately higher radon exposure-attributable burden because the combined effect considerably outweighs the sum of their individual effects. However, the inability to differentiate these interrelated exposures at the individual level is a major drawback of this ecological analysis. Without smoking history data at the individual level, it is impossible to determine whether the observed sex discrepancy reflects differences in radon exposure, smoking patterns, or both. Biological differences in susceptibility cannot be conclusively ruled out [34]. Nonetheless, the sex disparity is more parsimoniously explained by historical differences in smoking prevalence and occupational exposure patterns, such as confounders that cannot be controlled at the aggregate level. Despite projected increases in DALYs in women through 2050, the consistent trend in ASMR in women reflects the complex interaction between population aging (rising absolute DALYs) and stable age-specific mortality rates.

Age patterns and cohort effects

The observed age patterns, which show mortality increasing markedly with age, being low in children, and stabilizing after age 70, are consistent with the multi-stage carcinogenesis model for radon exposure-induced lung cancer [35]. Childhood exposure may contribute to the cumulative dose observed in adulthood [36], and the extended latency period (5–25 years) between exposure and clinical manifestation elucidates the delayed burden [37]. Competing risks from other causes of death in older adults and declining cellular repair capacity may result in a decrease in mortality differences after age 70 [38]. These trends underscore the importance of increased radon screening for older adults with previous exposure to high-incidence areas [39].

Mechanistic insights from SHAP analysis

The SHAP analysis offers novel insights that extend beyond conventional trend descriptions and quantifies the causes of radon exposure-attributable lung cancer mortality. Age 70–79 (0.375) was the most important predictor, which confirms that cumulative exposure over the course of an individual’s life and extended latency periods are more critical determinants of the burden than current radon levels alone [39, 40]. This finding has direct policy implications: regardless of current radon levels, screening programs should prioritize older adults with a history of exposure in high-risk areas [39]. The second-ranking importance of Sex_Male (0.235) quantitatively validates the substantial sex disparity, and country-level analysis confirms that countries with high smoking prevalence in men are the key mediators of this effect [26, 33]. Considering the synergistic relationship between smoking and radon exposure, this result supports tobacco control as a crucial intervention for lowering radon exposure-attributable burden [16, 25]. Geological_radon (0.221) was placed third; it supports focused monitoring and mitigation in high-risk geological locations, confirming the role of underlying geology as a fundamental driver [12, 41]. Healthcare_HAQ has a complex association with mortality, as explained by the country-level SHAP analysis. Notably, Healthcare_HAQ operates bidirectionally: improved healthcare reduces burden in high-access countries, whereas limited healthcare exacerbates burden in low-access countries [42, 43]. These mechanistic insights expand substantially beyond trend descriptions and demonstrate the importance of machine learning interpretability methods in global health research [21, 29].

Regional disparities and underlying mechanisms

High-latitude regions (Eastern Europe, Central Asia) had the highest ASMR in 2021 because of their cold climates, which led to enclosed housing spaces and reduced ventilation, potentially increasing indoor radon levels [44]. However, their estimates are more reliable because these regions have denser radon monitoring networks and a more extensive cancer registry. In contrast, owing to scarce primary data, estimates for numerous African countries mostly rely on model-based imputation [32]. The low ASMR reported in Africa may be attributed to decreased radon exposure, lower background lung cancer rates, or under-ascertainment in the absence of robust vital registration systems. This necessitates primary data collection in under-represented regions.

Despite reports of growing radon concentrations in some areas [40, 45, 46], high-SDI regions have shown diminishing burden trends. This may reflect the implementation of national radon action plans after 2010 [47] and increased access to healthcare [42], which could counteract exposure increases by improving detection and treatment. There are several possible explanations for the ASMR-SDI inverted U-shaped relationship, which shows a rise in burden from low to middle SDI and a decrease in high SDI. The actual burden may be obscured in low-SDI regions by underdiagnosis and competing causes of death. In middle-SDI regions, rapid industrialization and urbanization may increase radon exposure, whereas healthcare systems are still underfunded for early detection. Effective radon mitigation policies, improved access to healthcare, and declining smoking rates may counteract exposure increases in high-SDI areas. This interpretation is further supported by comparative studies of radon research and mitigation initiatives across socioeconomic contexts, which highlight substantial disparities in policy implementation and public awareness between high-income and lower-resource countries [32].

Projections and policy implications

According to BAPC projections, ASMR will continue to decrease globally until 2050; nonetheless, the uncertainty intervals will increase owing to the inherent challenges of long-term predictions. Despite stable ASMR, the projected increase in DALYs in women highlights the importance of population aging as a cause of future burden. These projections should be considered hypothesis-generating rather than conclusive estimates, because they are contingent on the continuation of present trends in radon exposure, smoking prevalence, and healthcare access. Similarly, GBD-based studies investigating comparable environmental risk factors have necessitated carefully interpreting long-term projections and the importance of routinely updating estimates as new data become available [23, 29].

These findings have several implications for international radon mitigation policies. First, radon action plans, which are centred in high-income countries [32], should be extended to lower-resource settings with international technical and financial support, owing to the increasing burden in low- and middle-SDI regions. This recommendation aligns with GBD-based studies that have advocated for targeted radon mitigation in rapidly urbanizing and high-latitude regions [29]. Second, the persistently high burden among men and older adults indicates that mitigation initiatives should target these groups, particularly in regions with high geological radon potential. Screening programs and remedial efforts should prioritize older housing stock, marked by high radon accumulation. Third, the projected increase in DALYs in women necessitates sex-sensitive preventative messaging and enhanced surveillance for the early identification of emerging trends. Finally, the substantial data gaps identified in Africa and other low-resource environments warrant funding primary data collection, such as radon monitoring networks and cancer registration systems, to reduce reliance on model-based estimates and strengthen the empirical basis for future assessments [32, 35].

Limitations

Several limitations that are inherent to GBD analyses and the ecological research design should be considered when interpreting the findings.

Uncertainty in exposure modeling: The estimates of residential radon exposure are modeled products rather than direct measurements. To estimate national and subnational exposure distributions, GBD utilizes a comparative risk assessment framework that integrates sparse direct measurements, geospatial predictors (geology, climate, and building attributes), and predictive covariates [9]. The lack of primary radon measurement data in many regions, particularly low-income countries, necessitates model-based imputation. In the absence of comprehensive ground-truth validation data, the accuracy of these modeled exposures, and thus the attributable burden estimates, remains uncertain. Sampling and model prediction uncertainty are partially captured by GBD’s uncertainty intervals. However, structural uncertainties resulting from alternative model specifications or unmeasured spatial heterogeneity in exposure determinants cannot be entirely accounted for.

Redistribution of ill-defined causes of death: The systematic redistribution of deaths caused by ill-defined or garbage codes (e.g., unspecified cancer or respiratory failure) to underlying causes using statistical algorithms is the foundation of GBD estimates for lung cancer mortality [15]. Although this redistribution process is essential to produce comparable estimates across countries with varying coding practices, it introduces uncertainty that is difficult to quantify. Estimates of radon exposure-attributable burden may be skewed in regions with low vital registration coverage because mortality estimates depend more on covariate-based predictive models and cause-of-death redistribution.

Ecological fallacy and aggregation bias: Our findings describe population-level associations and cannot be used to infer individual-level causal relationships because of the ecological research design based on countries and regions as the analysis units. Aggregation bias, the phenomenon where group-level associations diverge from individual-level associations, may prevent the observed patterns to accurately reflect effects at the individual level. This aspect is particularly relevant when interpreting sex and regional disparities, which may be confounded by unmeasured factors that cannot be sufficiently controlled at the aggregate level, including individual smoking history, occupational exposure, housing characteristics, and healthcare access.

Unmeasured confounding: Despite adjusting for SDI and geographic regions in regression analyses, residual confounding from factors correlated with both radon exposure and lung cancer mortality cannot be eliminated. Smoking, which exhibits a synergistic relationship with radon, is the most significant unmeasured confounder [8, 26]. The observed discrepancies in radon exposure-attributable burden are mostly explained by substantial variations in historical smoking patterns (smoking cohorts) across regions, sexes, and birth cohorts. Without individual-level data on smoking history, it is impossible to distinguish between the independent and combined effects of radon exposure and smoking history.

Temporal mismatches and latency: Radon-induced lung cancer involves a latency period of 5 to 25 years between exposure and clinical manifestation [37]. Precise exposure histories are not accessible at the population level; however, the GBD attribution framework assumes that the current burden of lung cancer reflects previous radon exposure. This temporal mismatch may introduce bias, particularly in regions undergoing rapid urbanization or housing stock changes, where current radon concentrations may not accurately reflect previous exposure levels relevant to the contemporary disease burden.

Data quality and regional imbalances: There are substantial regional differences in the accessibility and quality of primary data. High-income countries, particularly those in Europe and North America, have denser radon monitoring networks and comprehensive cancer registration systems, which render their estimates more credible [32]. In contrast, estimates for many low-income countries—particularly those in sub-Saharan Africa and Southeast Asia—rely considerably on model-based imputation and should be interpreted carefully [35]. The low ASMR recorded in Africa may be attributed to reduced radon exposure, lower background lung cancer rates, or under-ascertainment in the absence of robust vital registration systems. Notably, limited data prevent distinguishing these possibilities.

Lack of incidence estimates: GBD does not provide the incidence estimates of radon exposure-attributable lung cancer, precluding sensitivity analyses comparing mortality and incidence patterns. These comparisons may be used to assess whether observed trends in mortality reflect changes in exposure, survival, or both. To allow for more complex analysis, future iterations of GBD should incorporate incidence estimation for environmental risk factors.

Projection uncertainty: The BAPC projections through 2050 assume that existing trends in radon exposure, smoking prevalence, and healthcare access will persist. Over 30 years, these assumptions may not hold, and prediction intervals may widen considerably in later years because of increasing uncertainty. To improve future burden estimations, these projections should be considered hypothesis-generating scenarios rather than conclusive forecasts. Moreover, it is necessary to regularly update the burden estimates with new GBD data [23, 29].

GBD modeling assumptions: All findings are contingent on the modeling assumptions incorporated in the GBD 2021 framework. These assumptions include the theoretical minimum risk exposure level for radon, the exposure–response function primarily derived from European and North American studies, and statistical models used to impute missing data. The anticipated burden's level and distribution may change if any of these assumptions are altered. Although they are beyond the purview of this study, sensitivity analyses investigating alternative modeling assumptions may facilitate quantifying the robustness of GBD findings.

With thorough temporal trend analysis, age-period-cohort modeling, and long-term predictions across 204 countries, this study provides the most comprehensive and up-to-date estimate of radon exposure-attributable lung cancer burden. These findings can advance global surveillance, priority-setting, and the targeting of evidence-based radon mitigation efforts toward susceptible populations and regions.

Conclusions

This study offers a comprehensive, up-to-date account of the global, regional, and national trends in lung cancer burden caused by residential radon exposure. As ecological research based on modeled estimates, it is central to burden quantification, surveillance, and priority-setting. However, it does not yield new causative evidence. Our findings have several implications for international radon mitigation policies. First, radon action plans, which are focused on high-income countries, should be expanded to lower-resource regions owing to the growing burden in low- and middle-SDI regions (5.51% increase in DALYs). International health organizations and development partners should prioritize the establishment of radon monitoring networks and mitigation initiatives in these regions. Second, the continued high burden among men and older adults indicates that mitigation initiatives should focus on these demographic groups, particularly in regions with high geological radon potential, such as Eastern Europe and Central Asia. Because radon accumulation is more likely to occur in older housing stock, screening programs and remediation strategies should prioritize it and account for cumulative exposure histories. Third, despite a stable female ASMR, the projected increase in DALYs in women through 2050 necessitates sex-sensitive prevention messaging and enhanced surveillance to ensure the early detection of emerging trends. The ongoing burden among men, older adults, and populations in high-SDI regions, as well as emerging trends in low-SDI regions and among women, prioritize future research and potential interventions. These findings should inform the focus of evidence-based radon mitigation strategies on the most vulnerable populations and regions. Individual-level research has demonstrated the efficacy of these strategies. To reduce the reliance on model-based estimates and strengthen the empirical basis for future assessments, it is imperative to continue funding primary data collection, particularly radon surveillance and cancer registration in underrepresented regions.

Abbreviations

AAPC

Average Annual Percent Change

APC

Age-Period-Cohort

ASDR

Age-Standardized DALY Rate

ASMR

Age-Standardized Mortality Rate

BAPC

Bayesian Age-Period-Cohort

DALYs

Disability-Adjusted Life Years

EAPC

Estimated Annual Percent Change

GBD

Global Burden of Disease

GHDX

Global Health Data Exchange

HAQ

Healthcare Access and Quality Index

IARC

International Agency for Research on Cancer

IE

Intrinsic Estimator

PC

Penalized Complexity

RW2

Second-Order Random Walk

SDI

Socio-demographic Index

SHAP

SHapley Additive exPlanations

UI

Uncertainty Interval

YLD

Years Lived with Disability

YLL

Years of Life Lost

Authors’ contributions

L Zhao, D Liu, Y Zhou, X Zhang, Y Jia, J Li, Y Liu, G Lv, Y Zhang, and M Hao designed the study. L Zhao, L Jia, J Li, X Lu, H Zhang, Y Sun, M Luo, Y Zhang, Y Qiu, R Aimaiti, and Y Wang analyzed the data and performed the statistical analyses. L Zhao, W Shen, K Cui, L Fang, Y Guo, H Cui, Y Kong, J Liu, L Chen, and C Wu drafted the initial manuscript. All authors have reviewed and revised the manuscript for critical content. All authors have approved the final version of the manuscript.

Funding

The Science and Technology Program of XPCC (grant number 2023ZD019) provided funding for this research.

Data availability

The Global Health Data Exchange query tool (http://ghdx.healthdata.org/gbd-results-tool) provided online access to the data resources for the GBD 2021 study.

Declarations

Ethics approval and consent to participate

This study is based on the Global Burden of Disease (GBD) data. The University of Washington Institutional Review Board authorized a waiver of informed consent for the GBD study, which uses deidentified data. Because this research used this deidentified, publicly accessible dataset, no further ethics approval was required for the specific analysis. This approach complies with ethical requirements for secondary data analysis. Clinical trial number: not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.