Research on the Impact of PM_2.5 Pollution and Climate Change on Respiratory Diseases in Chinese Children Based on XGBoost-SHAP

Abstract

Children are among the most sensitive groups to air pollution. This study focuses on Chinese children aged 0–16 years, integrating six waves of tracking data from the China Family Panel Studies (CFPS, 2012–2022), the ChinaHighAirPollutants (CHAP) dataset, and MOD11A1 land surface temperature (LST) data, covering 20,241 samples across 25 provinces. Using the eXtreme Gradient Boosting–SHapley Additive exPlanations (XGBoost-SHAP) framework, we quantified the relative contributions of fine particulate matter (PM_2.5), ozone (O₃), nitrogen dioxide (NO₂), sulfur dioxide (SO₂), and climate factors to children’s respiratory disease risk. The overall area under curve (AUC) was 0.6765, with urban and rural sub-models achieving 0.6576 and 0.6864, respectively. SHAP analysis revealed that the temporal variable ranked first, reflecting population-level improvements from 2012 to 2022; age ranked second, with a 70.1% prevalence in the 0–6 age group. Rural PM_2.5 contribution was approximately 1.68 times that of urban areas; the O₃ effect showed opposite directions between urban (risk) and rural (protective association) settings; solid fuel contribution in rural areas was approximately 2.25 times the urban level. Regional clustering analysis identified differentiated environmental drivers across five geographic types. These findings provide a quantitative basis for differentiated regional prevention strategies.

1. Introduction

Air pollution is one of the leading environmental health risks in the world at present, and the World Health Organization (WHO) has listed air pollution as the top environmental health risk [1]. The report shows that 90 percent of the population is breathing polluted air [2] every day. Particulate matter in the air can enter the lungs and into the bloodstream, which in turn affects the cardiovascular and respiratory systems. Fine particulate matter PM_2.5 (aerodynamic diameter ≤ 2.5 μm), due to its extremely small size, can reach deep into the alveoli and penetrate the blood-air barrier into the systemic circulation, inducing systemic oxidative stress and inflammatory responses, thereby triggering or exacerbating respiratory diseases (asthma, chronic obstructive pulmonary disease, pneumonia) and cardiovascular and cerebrovascular diseases. It has been classified as a Group 1 definitive carcinogen by the International Agency for Research on Cancer (IARC) [3]. Studies have shown that air pollutant levels have a significant impact on human health. In the short term, this is reflected in a rapid increase in the incidence of respiratory diseases, especially among susceptible populations such as the elderly, children, and pregnant women. In the long term, particulate matter can enter the lungs and bloodstream [4,5,6], thereby affecting the cardiovascular and respiratory systems [1,7,8]. According to the Global State of Air report released by the Institute for Health Impacts (HEI) in June 2024 based on GBD 2021 data, 8.1 million people died from air pollution worldwide in 2021, making air pollution the second leading risk factor for death after hypertension. Among them, 58 percent died from PM_2.5 in the environment, 38 percent from household air pollution, and 6 percent from ozone [9]. In addition, for the first time, the report clearly pointed out that more than 700,000 children under the age of five died from air pollution worldwide in 2021, and air pollution has become the second leading risk factor for death in this age group after malnutrition [9].

Children are among the most sensitive exposure groups to air pollution, and their vulnerability stems from multiple physiological factors: breathing approximately twice as much air per unit of body weight as adults, having smaller airway diameters, having underdeveloped mucosal barrier function, possessing an immune system that is still maturing, and spending significantly more time outdoors than adults [10,11,12]. Epidemiological evidence suggests that long-term exposure to PM_2.5 is closely associated with impaired lung function growth in children, an increased incidence of acute respiratory tract infections, and an elevated risk of new-onset asthma [13,14]. Wu et al.’s study at 3106 pediatric asthma clinics in Xiamen found that exposure to PM_2.5 within two weeks significantly increased the risk of acute asthma attacks (OR = 1.049, p < 0.001) [15]. Liu et al.’s analysis of more than 10,000 children under 5 years old hospitalized for pneumonia and asthma in Ningbo from 2013 to 2017 showed that elevated concentrations of PM_2.5, PM₁₀, SO₂ and NO₂ were significantly positively correlated with hospitalization risk, and the 1–5 years old group and girls were more likely to be affected [16]. In recent years, the threat of ozone (O₃) pollution to children has drawn increasing attention. Contrary to the trend of PM_2.5 reduction, ground-level O₃ concentrations in China have continued to rise since 2013, with an average annual growth rate of 2.06 μg/m³ between 2021 and 2024 [17]. The photochemical formation of O₃ intensifies under high temperature and strong radiation conditions, which can induce airway hyperresponsiveness, aggravate asthma and bronchitis, and cause long-term irreversible damage to lung function in children during development. Nitrogen oxides (NO_X) and sulfur oxides (SO_X) damage the respiratory mucosa and impair cardiovascular function through chronic cumulative exposure. The cumulative effect of exposure to multiple pollutants makes the health risk in children far greater than predicted by the assessment of a single pollutant.

The rise in machine learning methods offers new approaches to environmental health research. Extreme gradient boosting (XGBoost) is gradually replacing traditional statistical models in the prediction of pollution health effects [18,19,20,21], thanks to its natural adaptability to nonlinear effects and high-order feature interactions, its insensitivity to variable dimensions, and its built-in missing value handling mechanism. Machine learning prediction models developed by Rajesh M et al. are capable of reliably predicting air pollutants such as PM_2.5, PM₁₀, and CO [22]; Mishra used machine learning to correlate population density, road proximity, and meteorological data with pollution readings to identify hotspots of PM₁₀ [23]. However, black-box models with excellent predictive performance are difficult to apply directly to policy interpretation, and interpretability becomes a core bottleneck for machine learning applications in environmental health [24]. SHapley Additive exPlanations (SHAP), based on the Shapley value in cooperative game theory, decomposes each prediction into a linear sum of the marginal contributions of each feature, satisfying theoretical axioms such as efficiency, symmetry, dummy, and additivity, and is currently the most theoretically guaranteed method for model interpretability [25]. Zhang et al. used SHAP to provide globally consistent feature importance ranking and numerical contribution analysis and combined it with the LIME method to improve the credibility of the medical environment comfort prediction model [26]. Rajesh M et al. also employed SHAP analysis to identify the environmental and demographic variables contributing most to each prediction, thereby achieving model interpretability [22].

Although XGBoost-SHAP has been applied in the fields of medicine and air quality prediction, there is still a lack of research that integrates multi-source heterogeneous environmental data and conducts a national systematic analysis of respiratory diseases in Chinese children [27,28]. A review of the existing literature reveals four core limitations that restrict a deeper understanding of the problem. First, the vast majority of studies focus on the main effects of a single pollutant or a single meteorological factor and fail to capture the synergy and interaction patterns of multiple environmental factors. Second, nationwide, child-specific, age-differentiated, large-scale multi-center studies are still very limited [29,30,31]; existing studies are mostly single-center or regionally limited time series analyses, with insufficient external validity [32]. Third, studies that accurately identify the net effects of various environmental factors after controlling for time trends (i.e., macro background variables such as air governance policies and economic and social progress) are still lacking [33,34]. Fourth, there is a lack of research on the heterogeneous mechanisms of environmental health in urban and rural areas, especially on the possibility that the same pollutant may have completely opposite effects between urban and rural areas [35,36].

This study focused on Chinese children aged 0–16 years, based on six national tracking periods of the China Household Tracking Survey (CFPS) from 2012 to 2022, and integrated multi-source environmental remote sensing data such as ChinaHighAirPollutants (CHAP) dataset and MOD11A1 land surface temperature (LST), covering a total of 20,241 sample observations in 25 provincial administrative units across the country. Based on the core methodological framework of XGBoost-SHAP interpretable machine learning, the following research objectives were achieved: (1) to quantify the relative contributions and nonlinear effects of multi-dimensional environmental and individual factors on the risk of respiratory diseases in children; (2) to accurately identify the net health effects of pollutants such as PM_2.5 by statistically stripping away the confounding effects of macro-temporal trends such as air governance policies through the inclusion of relative year variables; (3) to systematically analyze the magnitude and direction differences in SHAP contributions of various environmental factors under urban-rural stratification conditions to reveal the urban-rural heterogeneity mechanism; (4) to provide a multi-scale interpretability perspective from group patterns to individual mechanisms through the analysis of the SHAP waterfall diagram at the individual level; (5) to establish an evidence-based basis for formulating differentiated and precise policies for children’s environmental health.

2. Data and Methods

2.1. Research Data

The data used in this study were sourced from multi-source databases such as the China National Environmental Monitoring Centre, the China Meteorological Administration, and the CFPS. The dataset encompassed six survey waves: 2012, 2014, 2016, 2018, 2020, and 2022, spanning 25 provinces and comprising 34 original variables. The original data covered five dimensions: (1) air pollutant concentrations (PM_2.5, NO₂, SO₂, CO, ozone); (2) meteorological conditions (monthly average temperature, monthly average wind speed, monthly average air pressure, etc.); (3) socio-economic indicators (per capita GDP, population density, carbon emissions, etc.); (4) child respiratory health indicators (prevalence of respiratory diseases); (5) geographical and environmental characteristics (normalized difference vegetation index (NDVI), nighttime light intensity from the Defense Meteorological Satellite Program-Operational Linescan System (DMSP-OLS), traffic network density, railway density, etc.).

After systematic data cleaning—including removal of duplicate records, exclusion of samples with more than 30% missing values across key variables, and multiple imputation (using the MICE algorithm with 5 iterations) for remaining missing environmental data—this study ultimately retained a valid sample of 20,241 children aged 0–16 years (see

Table 1. Descriptive statistics of Basic characteristics of the sample.

Variables	Variable Description	Number of Disease Cases/Total (%)
Total	Total sample size	12,492/20,241 (61.7)
Demographic characteristics
Age (years)	Continuous variable (mean ± standard deviation)	-
0–6 years old	Pre-school age	5192/7406 (70.1)
6–12 years old	School age	5267/8885 (59.3)
12–16 years old	Adolescence	2031/3950 (51.4)
Gender (1)	Binary classification	-
Residence (urban = 1)	Binary classification	5785/9214 (62.8)
Residence (rural = 0)	Binary classification	6707/11,027 (60.8)
Parental behavioral factors
Parents smoke (is = 1)	Smoking in the past month	-
Cooking fuel (solid = 1)	Use solid fuel	-

). Continuous environmental variables were standardized using z-score normalization prior to model training. The average age of the sample was 8.35 years (standard deviation = 3.95), with 53.0% being female, 45.5% being urban, and 54.5% being rural. In terms of family behavioral characteristics, 61.0% of families had parents who smoked and 30.0% used solid fuel for cooking, the latter being a major source of indoor air pollution and more concentrated in rural areas.

2.2. Research Methods

2.2.1. XGBoost Model Construction

XGBoost builds weak classifiers (decision trees) in a sequential iterative manner based on gradient boosting, with each round taking the residual gradient of the previous round as the fitting target and imposing L1/L2 regularization constraints on the complexity of the tree structure, thereby effectively suppressing overfitting while achieving high-precision modeling of nonlinear relationships [37,38]. Compared with traditional statistical models, the XGBoost model naturally adapts to nonlinear effects and high-order interactions between features, is insensitive to multicollinearity, and does not require a preset function form of linear addition, making it particularly well-suited for modeling the multi-source heterogeneous and high-dimensional, strongly correlated environmental health data in this study.

Hyperparameter optimization was carried out using the Optuna Bayesian framework (TPE sampler, 50 trials), and the search space included the following parameters: number of iteration rounds (200–800), maximum tree depth (3–7), learning rate (0.01–0.15, logarithmic scale), row sampling ratio (0.6–1.0), column sampling ratio (0.5–1.0), minimum leaf node weight (1–15), and L1/L2 regularization coefficient. The mean AUC from 5-fold stratified cross-validation was used as the optimization objective to ensure that the class distribution in each fold was consistent with the full sample. The random seed was uniformly set to 42 to ensure reproducibility of the results. The full sample model and the urban-rural stratified sub-model were independently hyperparameterized to fully accommodate the differences in feature distribution among different data subsets.

Model performance was comprehensively evaluated across three dimensions: AUC (reflecting the overall discriminative ability of the model), F1 score (a harmonic mean metric that measures precision and recall), and Brier score (measuring the calibration quality of probabilistic predictions, with lower values indicating better calibration results).

2.2.2. SHAP Explainability Analysis Framework

In this study, the TreeExplainer algorithm was used to calculate SHAP values. The algorithm derived an exact polynomial-time solution for the ensemble tree model (with a time complexity of O(TLD²) [39,40], where T is the number of trees, L is the depth of the trees, and D is the maximum number of leaf nodes), and performed full SHAP matrix computations on all 20,241 samples. The SHAP framework is based on the additive decomposition principle of Shapley values. For the prediction of sample i, the model output can be decomposed as:

$$\mathrm{f}({\mathrm{x}}_{\mathrm{i}})=\mathrm{E}[\mathrm{f}(\mathrm{x})]+\sum _{\mathrm{j}}{\mathsf{\Phi }}_{\mathrm{i}\mathrm{j}}$$

where ${\mathsf{\Phi }}_{\mathrm{i}\mathrm{j}}$ is the SHAP value of feature j for sample i, and E[f(x)] is the mean prediction across the full sample (i.e., the baseline value). The positive or negative SHAP value indicates the direction of contribution, the absolute value reflects the magnitude of contribution, and the global feature importance is measured by the order of each feature’s SHAP value.

Interpretability analysis was conducted at three levels: (1) full sample SHAP feature importance and direction analysis, revealing the net effect ranking of each factor at the overall sample level; (2) urban-rural stratified SHAP contrast analysis, systematically comparing the differences in contribution magnitudes and directions of each feature between the urban subsample (n = 9210) and the rural subsample (n = 11,031) to reveal the mechanism of heterogeneity between urban and rural areas; (3) individual-level SHAP waterfall graph analysis, which involves feature-by-feature decomposition of four typical types of samples (high-risk cities, high-risk rural areas, low-risk cities, and low-risk rural areas) to examine the specific manifestations of population patterns at the individual level. All analyses were conducted in Python 3.12.1 (key packages: NumPy 2.3.1, Pandas 2.3.0, scikit-learn 1.6.1, XGBoost, SHAP, GeoPandas 0.12.2). Statistical tests were two-sided, with a significance level of α = 0.05.

3. Results and Discussion

3.1. Model Performance Evaluation

As shown in

Table 2. Performance Comparison of eXtreme Gradient Boosting (XGBoost) Models.

Model	Area Under Curve (AUC) Mean	Standard Deviation of AUC	F1 Mean	Standard Deviation of F1	Brier Mean	Brier Standard Deviation
XGBoost overall	0.6765	0.0110	0.7541	0.0056	0.2153	0.0022
XGBoost Urban	0.6576	0.0132	0.7553	0.0056	0.2177	0.0033
XGBoost Rural	0.6864	0.0063	0.7508	0.0051	0.2142	0.0019

, the 5-fold cross-validation AUC of the overall XGBoost model was 0.6765, the AUC of the urban sub-model was 0.6576, and the F1 score reached 0.7553. The rural sub-model had an AUC of 0.6864 and an F1 score of 0.7508.

The AUC of each model was in the medium range of 0.63 to 0.69, and the overall discriminative ability was somewhat limited. This moderate AUC is primarily constrained by the data structure itself: this study matches individual samples with provincial-level average annual environmental indicators, a design that makes it difficult to capture the actual exposure levels of individuals in microenvironments such as schools and residences; at the same time, the absence of key variables such as genetic factors, household hygiene, and medical treatment behavior objectively compressed the upper limit of the model’s predictive accuracy. It should be noted, however, that the core objective of this study is mechanism analysis rather than precise prediction, and a moderate level of AUC does not affect the scientific significance of feature attribution analysis [41].

3.2. Results Analysis and Discussion

3.2.1. Full Sample SHAP Analysis

Figure 1 shows the average SHAP values of each feature of the overall model. As shown in the figure, relative years are at the top with a mean SHAP of 0.3120, higher than age (0.2689). This result does not imply that the time trend is biologically more important than age, but rather reflects significant cross-year heterogeneity in the data: the survey years from 2012 to 2022 encode multiple historical contexts such as policy evolution, economic and social transformation, and the effectiveness of air governance, all of which work together on children’s respiratory health, and their combined effect exceeds the variation attributable to age alone within the sample.

The mean SHAP of age is positive. The prevalence rate of 70.1% among children aged 0–6, compared with only 51.4% among those aged 12–16, indicates that the respiratory system of young children is more physiologically vulnerable owing to its developmental immaturity, a pattern that remains consistent across different environmental contexts and survey years.

3.2.2. Urban-Rural Stratified Analysis of SHAP

As shown in

Table 3. Comparison of SHapley Additive exPlanations (SHAP) Feature Importance in Urban-Rural stratified Models.

Numbers	Urban Sub-Model			Rural Sub-Model			Main Differences
	Features	|SHAP|	Position	Feature	|SHAP|	Position
1	Relative years	-	1	Relative Years	-	1	1 on both sides, showing the universality of the time trend
2	Age	0.314	2	Age	0.271	2	The urban age effect is slightly stronger
3	O3	0.054	5	O3	high	3	Rural O3 ranks higher (protection direction)
4	Average monthly temperature	0.078	4	Average monthly temperature	0.083	4	The temperature effect is similar between urban and rural areas
5	Relative humidity	0.056	8	PM2.5	0.085	5	PM2.5 Rural > Urban (1.68 times)
6	Normalized Difference Vegetation Index (NDVI)	0.058	3	SO2	0.073	6	SO2 sources in rural areas are associated with coal burning
7	PM2.5	0.051	7	Solid fuel	0.049	7	Rural indoor fuel exposure is prominent
8	Land Surface Temperature (LST)	0.052	6	LST	0.082	8	Heat exposure is stronger in rural areas
9	NO2	0.049	9	Relative humidity	0.078	9
10	Carbon emissions	0.045	11	NO2	0.085	10
11	SO2	0.044	10	Rainfall	0.092	11	The protective effect of rainfall is more pronounced in rural areas
12	Parents smoke	0.044	13	Population density	0.024	12
13	Defense Meteorological Satellite Program–Operational Linescan System (DMSP-OLS)	0.038	15	NDVI	0.056	13
14	Railway density	0.034	14	Per capita GDP	0.058	14
15	Gender	0.033	16	Air pressure	0.023	15

, the macro-time trends from 2012 to 2022 had universal and systematic effects on the reduction in the risk of respiratory diseases among children, both in urban and rural areas, and these effects were of a magnitude greater than that of any single environmental or individual variable. Figure 2 presents this information more intuitively.

The mean SHAP of PM_2.5 in the rural sub-model (0.0854) was significantly higher than that in the urban sub-model (0.0507), approximately 1.68 times that of the latter, indicating differentiated exposure-response mechanisms between urban and rural areas. Children in rural areas spend more time outdoors, have lower air tightness in their rooms, and generally lack the air purification equipment commonly found in urban areas. These differences in exposure have direct policy implications: the same reduction in ambient PM_2.5 would yield greater marginal health benefits in rural than in urban areas, suggesting that rural areas should be prioritized in PM_2.5 control strategies.

O₃ ranked high in both sub-models, but the effects are in opposite directions: in the urban sub-model, O₃ ranked approximately fifth with a positive mean SHAP (+0.016), indicating a positive correlation between increased O₃ concentration and disease risk in urban areas; in the rural sub-model, O₃ ranked approximately third with a negative mean SHAP (−0.012). This reversal can be explained from the perspective of atmospheric chemistry: in urban areas, NOx emissions are high, and there is a significant “titration effect” on O₃. Once NOx concentration drops (such as at night or after policy control), O₃ concentration rebounds, thereby generating oxidative stress risk; in rural areas, where NOx sources are scarce and O₃ background concentrations are relatively high, high O₃ is often accompanied by abundant solar radiation and favorable meteorological conditions (dryness, high pressure, weak wind), which have a certain inhibitory effect on the spread of some respiratory pathogens (such as influenza virus, respiratory syncytial virus), thus forming a conditional protective effect. This finding has methodological implications: the SHAP direction of the same pollutant may be opposite in different urban and rural contexts, and if only the overall analysis is conducted, the two mechanisms will cancel each other out, resulting in the masking of the true effect. However, it should be noted that these directional interpretations are based on statistical associations within the model framework and may be influenced by unmeasured confounders. The proposed atmospheric chemistry mechanisms (NOx titration effect, pathogen suppression) require further validation through controlled epidemiological studies and atmospheric modeling analyses.

Solid fuel usage ranked seventh in the rural sub-model and only around 18th in the urban sub-model, with rural contributions approximately 2.25 times that of urban ones. This is in line with the structural feature in the dataset that solid fuel usage in rural areas is significantly higher than that in urban areas. Indoor PM_2.5, CO and polycyclic aromatic hydrocarbons produced by burning coal and biomass can act directly on the respiratory mucosa, and the impact is more pronounced in infants and young children whose lung function is not fully developed. Therefore, promoting clean energy substitution in rural areas (replacing coal heating with natural gas and promoting energy-efficient stoves) is both a policy tool for carbon reduction and a direct intervention to protect the respiratory health of rural children.

3.2.3. Hierarchical SHAP Analysis of Clustering Clusters

All five cluster types presented the same top two ranking structure: first relative year, second age. This high consistency across different climate zones, urbanization levels and pollution structures further validates the universality of the time trend effect and the robustness of the biological vulnerability effect at the clustering level. The mean SHAP of age in each cluster was between 0.262 and 0.291, with small differences between groups, reflecting the vulnerability of the respiratory system of young children as a biological universal law independent of the regional environment. The specific results are shown in

Table 4. Top 5 Driving Features for Five Types of clustering.

Serial Numbers	Cluster I High Temperature and High Humidity	Cluster II Megacities	Cluster III High PM2.5	Cluster IV Warm Low Town	Cluster V Low-Temperature Rural Areas	Cross-Cluster Patterns
1	Age	Age	Age	Age	Age	Five clusters consistent: age is the primary risk factor
2	Relative years	Relative Years	Relative Years	Relative Years	Relative Years	Time trends rank second
3	Monthly average temperature	NO2	PM2.5	SO2	Average monthly temperature	The differentiated environmental driving factors of each cluster have emerged
4	O3	PM2.5	SO2	NDVI	LST	Cities/high pollution clusters are dominated by pollutants
5	Rainfall	Carbon emissions	NO2	PM2.5	Rainfall	PM2.5 ranking: 3rd in Cluster III and 5th in Cluster I

.

Cluster I (high temperature and high humidity type, Guangdong, Guangxi, Zhejiang, etc.): the 3rd to 5th places are, respectively, the average monthly temperature, O₃ and rainfall. PM_2.5 ranked lower among pollutants, indicating that the risk of respiratory disease in this cluster is mainly dominated by climatic factors. High temperatures accelerate the proliferation and spread of respiratory pathogens; strong solar radiation and high ozone concentration in South China directly irritate the respiratory tract; the protective effect of rainfall is also most prominent in this cluster.

Cluster II (ultra-high urbanization type, Beijing, Tianjin, Shanghai): NO₂ and PM_2.5 rank 3rd to 4th, and carbon emissions are also among the top 5, presenting a composite pollution structure dominated by traffic sources. The three super-large municipalities have the largest number of motor vehicles in the country, with a high concentration of NOx emissions. The near-surface O₃ driven by their high NOx emissions forms a distinctive photochemical pollution pattern. For children in Cluster II, the control of roadside pollution exposure during morning and evening rush hours and the rational arrangement of outdoor activity periods are key intervention directions distinct from other clusters.

Cluster III (high PM_2.5 type, North China Plain: Henan, Shandong, Hebei): PM_2.5 is the third and SO₂ is the fourth, with similar contributions. This is the only type of the five clusters with PM_2.5 in the top 3, corresponding to the highest annual average PM_2.5 concentration in the country (about 70–85 μg/m³). SO₂ and particulate matter emitted jointly by coal-fired power plants and heavy chemical industries constitute a typical coal smoke-type compound pollution pattern in the North China Plain. Even when time trends are controlled, higher PM_2.5 concentrations in the region still play a significant role in driving the risk of respiratory disease among children.

Cluster IV (Warm and low-urbanized agricultural type, Southwest: Sichuan, Yunnan, Guizhou): SO₂ ranked 3rd, NDVI ranked 4th, PM_2.5 ranked 5th. The outstanding contribution of SO₂ reflects that industrial and mining activities and the use of raw coal in the southwest region remain at a relatively high level; the NDVI ranking in the top 5 indicates that in the complex terrain conditions of the southwest mountains, the regulatory effect of vegetation coverage on microclimate is an important protective factor. Higher vegetation coverage often accompanies better air quality and lower pollution exposure levels, and the service functions of ecosystems in purifying the air and mitigating the heat island are particularly prominent in this cluster. These findings support the inclusion of child health considerations in ecological conservation policies for the southwest region.

Cluster V (low-temperature rural type, Northwest/Northeast: Gansu, Liaoning, Shanxi): monthly average temperature ranked third (with the highest contribution among the five clusters), LST ranked fourth, both pointing to the special health effects of low-temperature exposure in cold regions. In extremely cold environments, children’s respiratory mucosa is more sensitive to cold air stimulation; the demand for heating in winter is high, and the use of solid fuel (coal) is at its peak. The superimposed effect of indoor and outdoor pollution is most prominent in rural areas of the north where the annual temperature difference is the greatest. Rainfall ranked fifth and is negative in direction, indicating that even basic water vapor replenishment can provide some wetting protection for the respiratory tract in the arid and rain-scarce northwest region. For this cluster area, accelerating rural clean heating renovations and reducing indoor soot exposure in winter should be prioritized as intervention directions.

To further illustrate the impact of PM_2.5, this study plotted scatter plots of PM_2.5 concentration and mean SHAP in the five cluster types, as shown in Figure 3. The clusters do not show a simple linear relationship: cluster III with the highest PM_2.5 concentration does not correspond to the highest positive SHAP, once again confirming the nonlinear characteristics of the impact of PM_2.5 on children’s respiratory health.

3.2.4. Analysis of the Typical Case SHAP Waterfall Diagram

Figure 4 supplements the micro perspective of population analysis at the individual level by performing feature-by-feature SHAP decomposition on four typical samples.

High-risk-urban sample (age = 3 years, PM_2.5 = 86.62 μg/m³, NO₂ = 51.67 μg/m³, relative year = 0, that is, 2012): age and PM_2.5 together constitute the most significant positive contribution, and the relative year = 0 (early investigation year) also contributes positively, indicating that the early-survey-year environmental context (a period when air pollution control had not yet been fully implemented) independently contributed to the elevated disease risk of this sample. The lower per capita GDP (61,880 yuan) also played a positive role, in line with the pattern of higher social inequality among low-income groups within cities.

High-risk-rural sample (age = 3 years, PM_2.5 = 86.62 μg/m³, relative year = 0): O₃ (=89.00 μg/m³) replaced per capita GDP as the main positive contributing factor compared with the urban high-risk sample, reflecting the direct risk effect of high concentrations of O₃ at the individual level. This is different from the population pattern where the overall mean O₃ SHAP is negative in rural areas, indicating that even if O₃ has a protective effect at the population level, high concentrations of exposure at the individual level can still translate into a positive risk contribution. The DMSP-OLS nighttime light index and railway density values were extremely low (0.07), and their SHAP contribution directions differed between the urban and rural sub-models, indicating that the scarcity of infrastructure and the lack of medical accessibility formed a superimposed effect in the rural high-risk sample.

Low-risk-urban sample (age = 15 years, relative year = 8, that is, 2020, NO₂ = 43.43 μg/m³, rainfall = 49.28 mm): age (15 years, the respiratory system is fully developed) and relative year (=8, the benefits of air pollution control are shown) together produce the greatest negative contribution; a monthly average temperature of 12.45 °C and a relative humidity of 55.56% also had a negative effect, indicating that a mild and humid climate can reduce the stress response in children’s respiratory tracts.

Low-risk-rural sample (age = 15 years, relative year = 8, rainfall = 2106.45 mm, O₃ = 74.25 μg/m³, urban-rural = 0, that is, rural): rainfall up to 2106 mm (possibly corresponding to the rainy areas in the southwest or southeast coast) produced the greatest negative SHAP contribution, confirming at the individual level the same rainfall protection effect observed at the population level; the negative contribution of O₃ = 74.25 μg/m³ is consistent with the negative mean O₃ SHAP at the rural population level, and the urban-rural identifier (0 = rural) also has a negative contribution, which together constitute the low-risk feature of the sample. In addition, the relatively small contribution of solid fuel use in this sample may be related to the stronger tolerance of older children to indoor pollution than the younger group.

This study has the following limitations at the methodological level that should be considered when interpreting the results. (1) Ecological inference and exposure misclassification: this study matches provincial-level average annual environmental indicators with individual-level health outcomes, which constitutes an ecological study design. This approach introduces considerable exposure misclassification, as provincial averages cannot capture individual-level microenvironmental exposures (e.g., proximity to roads, indoor conditions, school environments). Therefore, the statistical associations identified by SHAP analysis should be interpreted as population-level associations rather than individual-level causal risk drivers. Policy implications derived from these findings require further verification using quasi-experimental epidemiological methods such as regression discontinuity designs. (2) Interpretation of the temporal variable: the dominant contribution of the ‘relative year’ variable likely reflects confounding by multiple macro-level trends, including policy interventions (e.g., the ‘Air Pollution Prevention and Control Action Plan’), socio-economic development, improvements in healthcare access, and changes in diagnostic practices. While this variable captures meaningful temporal variation, attributing its effect solely to governance effectiveness would be an overstatement. A more cautious interpretation is that this variable encodes a composite temporal signal that warrants further decomposition through sensitivity analyses or instrumental variable approaches in future studies. (3) Model performance and SHAP interpretability: the AUC values of the models fall in the moderate range (0.63–0.69), indicating limited discriminative ability. While SHAP-based feature attribution has been shown to remain informative even when model predictive performance is moderate—because SHAP decomposes the model’s learned associations rather than requiring perfect prediction—there is inherent uncertainty in the derived feature importance rankings. The SHAP values should therefore be interpreted as indicators of relative variable importance within the model’s framework, not as precise effect size estimates. (4) Spatial aggregation and temporal autocorrelation: the study does not fully account for spatial clustering effects, regional socio-economic heterogeneity within provinces, and temporal dependence in environmental variables across the six survey waves. These factors may influence the stability of the derived feature importance rankings and introduce additional uncertainty that is not captured by the cross-validation framework. (5) The data only covered 25 provinces and did not include western regions such as Xinjiang, Xizang, and Qinghai, limiting geographical representativeness. (6) The SHAP waterfall plot analysis of typical cases is based on specific samples from the dataset, and their representativeness is affected by sample heterogeneity. These individual-level decompositions serve an illustrative purpose and should not be used as a direct basis for causal inference.

4. Conclusions

Based on the XGBoost-SHAP analysis of 20,241 children aged 0–16 years across 25 Chinese provinces (CFPS 2012–2022), four key findings emerged: (1) The temporal variable ranked first in SHAP importance (mean |SHAP| = 0.3120), reflecting the composite effect of air governance policies and socio-economic progress on children’s respiratory health over the past decade, though this signal likely encodes multiple confounded macro-level trends; (2) age ranked second (mean |SHAP| = 0.2689), confirming that the physiological vulnerability of young children (0–6 years, prevalence 70.1%) is a robust risk factor across all regions and time periods; (3) urban–rural stratification revealed that rural PM_2.5 contribution was approximately 1.68 times the urban level, O₃ showed opposite statistical associations (positive in urban, negative in rural areas), and solid fuel contribution in rural areas was approximately 2.25 times that of urban areas; (4) five regional cluster types exhibited differentiated environmental drivers—temperature and O₃ dominated in the South China cluster, NO₂ in megacities, PM_2.5 in the North China Plain, SO₂ and NDVI in the Southwest, and low-temperature exposure in the Northwest/Northeast. Future studies could employ alternative frameworks such as LightGBM-SHAP, CatBoost, or neural additive models to validate and extend these findings.