This study selects data from 30 provincial-level units on the mainland of China, covering the period from 2008 to 2022. Hong Kong, Macau, Taiwan, and Tibet are excluded from the research scope due to significant data gaps. Based on relevant literature (20, 40), the input variables are categorized into 16 variables in four major factors, namely, demographic, economic, land and social factors, and the data sources are all from the National Bureau of Statistics as well as provincial statistical yearbooks, while the output variable is the incidence rate of class A and B infectious diseases per 100,000 population, and the data source is the China Health Statistical Yearbook. The Total Dependency Ratio is not published for 2020. According to its definition, this study imputes the 2020 value based on census data, using the ratio of the non-working-age population (0–14 and 65+) to the working-age population (15–64). See Table 1 for descriptive statistics on the classification of variable categories.

The Random Forest algorithm, proposed by Breiman (30), is an ensemble learning-based method that extends the Bagging algorithm. It trains a decision tree on each subset and randomly selects a subset of features for splitting at each node, increasing feature randomness and reducing correlations between decision trees. The prediction results are obtained by averaging or voting the predictions of all trees. This process helps to reduce variance and errors, preventing both overfitting and underfitting. Each feature is ranked according to its assigned weight, referred to as feature importance (31). Through multiple random sampling and feature selection, Random Forest effectively handles high-dimensional data and uses tree-based ensemble methods to capture complex non-linear relationships between variables. Random forests evaluate feature contributions by calculating the Gini index after each feature split. Feature importance is defined as the total reduction in Gini impurity achieved by all splits involving that feature across the ensemble of trees. A higher value indicates a greater contribution of the feature to the classification task.

Gradient Boosting Decision Tree (GBDT) is a Boosting algorithm with CART regression trees as the base learners, primarily aimed at optimizing general loss functions. The core idea is to fit the residuals of the previous round’s base learner using the negative gradient of the loss function, progressively reducing the residuals with each round. This approach ensures that the output of each round’s base learner approaches the true values. By fitting along the negative gradient direction, the method guarantees that the loss function decreases as rapidly as possible in each training round, accelerating convergence to a local or global optimal solution (29). GBDT is a powerful ensemble learning method capable of handling both continuous and discrete values. By iteratively optimizing the loss function, it captures complex nonlinear relationships, and the stage-by-stage residual fitting process ensures robustness even in the presence of noise, with regularization parameters effectively preventing overfitting. The calculation of feature importance in GBDT essentially quantifies “the aggregate contribution of a feature to the reduction of the loss function across all trees.” The more frequently a feature is utilized for node splitting, and the more significantly each split reduces the loss function, the higher its importance weight.

XGBoost is an efficient implementation of GBDT, which first uses the training set and sample true values to train a tree. This tree is then used to predict the training set, and the residuals (i.e., the differences between the predicted and true values) are calculated. The residuals are then used as the target for training the second tree. After the second tree is trained, the residuals are recalculated for each sample, and a third tree is trained, and so on (32). XGBoost utilizes a second-order Taylor expansion to more accurately approximate the optimal solution of the loss function, thereby improving prediction accuracy. During decision tree construction, it uses an “approximate greedy algorithm” to reduce computational complexity. The method also supports custom loss functions and evaluation metrics, allowing model parameters to be adjusted according to the research problem. Additionally, it can handle missing values automatically and supports parallel computation. XGBoost quantifies feature importance by measuring the reduction in the objective function (e.g., loss function) achieved at each split, referred to as the “gain.” After each split, the gain score of feature A accumulates the reduction in loss attributable to that split. The final importance score for feature A is obtained by summing its total gains across all trees. A higher gain indicates that the feature plays a more significant role in distinguishing samples.

The Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), and Mean Absolute Error (MAE) are the main metrics for evaluating the model’s prediction performance. The coefficient of determination (R²) is used to assess the goodness of fit of the model, and a higher R² value indicates a superior goodness of fit. MAE represents the average absolute error between predicted and actual values, MAPE indicates the degree of data dispersion, and RMSE measures both the error between actual and predicted values and the dispersion between the two errors. The smaller these metrics, the better the model’s performance. However, all three metrics are influenced by sample values and predicted values, and there is no predefined standard. Given that K-fold cross-validation is appropriate for small datasets and facilitates effective use of all data for training and evaluation, and to ensure that each fold contains an adequate number of samples during cross-validation while preventing sample structure imbalance, this study adopts three-fold and four-fold cross-validation. The evaluation is conducted using holdout samples under various training/testing splits (70, 80, and 90%). For Boosting-based models (GBDT, XGBoost), their mechanism of sequentially fitting residuals tends to drive training error toward 0 and training R² toward 1 when the number of iterations increases. Since RMSE has a broader evaluation range, this study chooses R² and RMSE to evaluate prediction performance.

In this study, SPSSPRO is used to execute all algorithms, employing machine learning regression modules such as Random Forest, GBDT, and XGBoost. Table 1 shows the descriptive statistical characteristics of variables.

Initial implementation allocates 70% of samples to the training set and 30% to the test set, with three-fold cross-validation. As Table 2 shows, the results indicate a training set RMSE of 27.454 and a test set RMSE of 48.524, corresponding to coefficients of determination (R²) of 0.93 and 0.696, respectively.

Subsequent trials adjusted the training set proportions to 80 and 90%. While test set RMSE decreased with larger training sets, the 80% training allocation exhibits lower training set RMSE (23.117) compared to the 90% partition (25.832). However, the 90% training configuration demonstrates superior generalization performance, achieving a higher R². Comparative analysis of these metrics suggests the 90% training ratio optimizes predictive accuracy and model fit.

To mitigate potential overfitting risks inherent in high training set allocations (90%), this study adopts complementary methodologies to enhance model generalizability and investigates the model’s adaptability to varying data volumes. Subsequent sections detail these refinements and their validation outcomes.

Table 3 presents the performance metrics of three algorithms under varying training ratios and cross-validation folds.

In Scenario 1, the GBDT algorithm achieved the lowest RMSE values (training set: 0.124; test set: 47.721), significantly outperforming the other two algorithms. Correspondingly, its coefficients of determination (R²) for both training and test sets surpasses those of alternative methods. Similar trends are observed in Scenario 2. However, in Scenario 3, GBDT and XGBoost demonstrated comparable performance without statistically significant differences in metrics. In Scenario 4, GBDT exhibits a lower test set RMSE than XGBoost, while achieving the highest R² value.

Six optimal configurations are subsequently identified: GBDT (Scenario 1), GBDT (Scenario 2), GBDT/XGBoost (Scenario 3), and GBDT/XGBoost (Scenario 4). Comparative analysis reveals that GBDT (Scenario 3) delivers the lowest RMSE values across both training and test sets. Although its R² (0.783) marginally trails that of GBDT (Scenario 4, R² = 0.812), the negligible performance gap (<3%) prioritizes RMSE minimization for model selection. Consequently, GBDT (Scenario 3) is deemed the optimal predictor, followed by GBDT (Scenario 4) and XGBoost (Scenario 3). The boosting models achieve near-perfect fits on the training data (R² = 1) while exhibiting a generalization gap on the test sets—a common pattern under noisy observational data. The selected specification best balances bias and variance, yielding the lowest test RMSE. Subsequent analyses focuses on these three configurations.

Table 4 summarizes variable importance rankings across three selected configurations. Social factors consistently dominated all models, accounting for over 45% of total feature weights, with GBDT (Scenario 4) reaching 0.477.

Discrepancies emerged among secondary factors: GBDT (Scenario 3) prioritized demographic factors (28.1%), followed by economic (15.6%) and land-related (11.3%) elements, whereas GBDT (Scenario 4) and XGBoost (Scenario 3) assigned higher weights to land factors (22.4 and 19.8%, respectively), relegating demographic and economic factors to tertiary positions. A composite ranking derived from mean importance scores established the hierarchy as: social factors > land factors > demographic factors > economic factors. Quantitative weight calculations are shown in Table 5 and Figure 2.

China spans a vast territory with significant geographical diversity. There are notable differences across provinces in terms of geographical location, resource endowment, industrial structure, and degree of openness. Additionally, the severe imbalance in regional development leads to regional variations in the impact of rapid urbanization on public health. Therefore, in this study, the 30 provincial-level units of Mainland China are divided into three regions, as outlined in Table 6.

Given that the GBDT algorithm outperforms other algorithms, this section continues to use the GBDT algorithm to discuss the situation in different regions. The training set accounts for 70%, and a four-fold cross-validation method is similarly employed. The predictive performance is shown in Table 7.

Table 7 shows that there is no significant difference in the RMSE of the training sets across the three regions. In the test set, the western region has the highest RMSE, followed by the eastern region, while the central region exhibits the lowest RMSE. The R² values of the training sets for all three regions are satisfactory. However, in the test set, the R² for the central region is less, which may be attributed to the relatively smaller sample size in the central region.

Table 8 presents the weights of various factors across the three major regions. In the eastern region, the most significant factor is the social factor, with a weight of 60%, followed by land factors, which account for nearly a quarter, and the lowest is the economic factor, with a weight just slightly above 5%. In the central region, economic factors dominate, with a weight exceeding 60%, followed by land factors, which account for 17.2%, while population factors have the lowest weight, under 10%. In the western region, land factors take the lead, with a weight greater than half, followed by economic factors, which account for more than one-fifth, and population factors also represent a relatively small proportion. Table 9 and Figure 3 show the weight calculation results for the variables in the three regions, which also exhibit considerable differences.

Social factors, with a weight of 0.467, rank first among all influencing factors, indicating their predominant influence on public health throughout urbanization. Social factors include healthcare and sanitation expenditures, public health facilities, and the accessibility of urban public transportation, reflecting the crucial role of infrastructure in the rapid urbanization process. Fitch (33) emphasized that well-developed infrastructure is essential for preventing and controlling epidemics, as it directly affects residents’ access to healthcare resources and their capacity to cope with diseases. Indicators such as the proportion of healthcare spending in general fiscal expenditure, the number of urban practicing assistant physicians per 10,000 people, and the number of hospital beds per 10,000 people reflect the distribution of public health services during urbanization. If public health service provision can keep pace with urbanization, it can effectively alleviate the supply–demand imbalance in healthcare and control the scope of infectious diseases. The medical resource shortages during the COVID-19 pandemic highlighted that higher healthcare investment, more specialized physicians, and broader disease prevention programs are crucial for controlling the spread of epidemics. Indicators such as the number of public transportation vehicles per 10,000 people and urban water supply coverage reflect the convenience of urban living conditions. A well-developed public transportation system can reduce the time cost for residents to access healthcare services, enhance the accessibility of medical services, and decrease the use of private cars, thereby reducing traffic congestion and emissions, and improving air quality. However, public transportation vehicles are crowded public spaces and enclosed environments, which inevitably become breeding grounds for the transmission of infectious diseases. They can also facilitate the spread of viruses from one area to another, contributing to the dissemination of diseases (34). Safe drinking water supply is fundamental for preventing diseases such as enteric infections (35). The 1854 cholera outbreak on Broad Street in London, during the first industrial revolution, was primarily caused by contaminated drinking water. Therefore, during urbanization, the government must prioritize the construction and improvement of the social service system. By increasing healthcare investment, optimizing public transportation infrastructure, and improving water supply coverage, the urban response capacity to public health events can be enhanced, ensuring that prevention and control measures are implemented quickly and effectively, thus controlling the spread and speed of epidemics.

Land factors rank second in significance, with an average weight of 0.215, underscoring their essential role in shaping infectious disease patterns. Rapid urban growth intensifies land conversion and population agglomeration, rendering prudent spatial planning indispensable. Poorly designed land-use systems—particularly those characterized by unbalanced development or inadequate green infrastructure—tend to give rise to overcrowded settlements and unsanitary conditions conducive to disease transmission (36). The green coverage rate in built-up areas occupies the largest proportion among land-related indicators, suggesting that a higher green coverage rate not only improves air quality but also absorbs airborne pathogens, thereby reducing the likelihood of residents contracting diseases (8, 23, 37). Per capita road area reflects the smoothness of traffic flow, and smooth transportation is beneficial for the rapid passage of emergency medical vehicles, significantly shortening emergency response times and improving the efficiency of medical rescue efforts. However, it can also serve as an important medium for the spread of viruses, providing a rapid channel for viral transmission. This contradiction stems from the dual nature of the transportation system in terms of spatial mobility and social connectivity. Moreover, indicators such as the proportion of built-up areas relative to total land and per capita urban road area generally remain low in direct correlations with infection incidence, likely because these parameters influence health through longer, indirect feedback loops rather than immediate effects. Consequently, sustainable urban planning should conform to ecological principles by balancing expansion with environmental preservation. Expanding green zones, regulating urban density, and designing efficient road networks collectively enhance urban livability while constraining the spatial scope of disease diffusion. For most developing countries, particular attention should be paid to the sustainable development of suburban areas, and systemic measures should be implemented to eliminate potential sources of infection.

The average weight of population factors is 0.17, exert measurable yet less direct effects compared to social and land determinants. Their influence derives primarily from the contingent risks embedded in population distribution and demographic composition. The total dependency ratio, an indicator of age structure, illustrates the socioeconomic burdens related to non-working populations. Elevated dependency ratios correspond to a higher proportion of older adults and juvenile residents—groups inherently more susceptible to infectious diseases due to weaker immune capacity. A disproportionate presence of such vulnerable demographics intensifies the likelihood of disease spread. The proportion of urban population is an important indicator of urbanization. In the short term, the large-scale concentration of population in cities increases the frequency of person-to-person contact, accelerating the spread of infectious diseases and expanding their transmission range. If health infrastructure lags behind the speed of urbanization, it also adds pressure to the operation of existing urban infrastructure, worsening living conditions. For example, densely populated areas such as urban villages, where sanitation facilities are relatively underdeveloped, can easily become sources of infection (38). However, population concentration can also facilitate the centralized provision of public health services, improving the efficiency of medical resource utilization. Illiteracy rates reflect the quality of the population. Residents with higher levels of education are generally better able to access and absorb knowledge related to infectious disease prevention and control, applying it in their daily lives, and are more likely to proactively share such knowledge. Therefore, governments should pay close attention to changes in population structure during the urbanization process, strengthen the management and health education of the mobile population, widely disseminate public health knowledge, and effectively protect the rights of vulnerable groups while ensuring the implementation of basic healthcare protection.

The average weight of economic factors is 0.148, possess the least direct impact on infectious disease prevalence, yet they underpin the broader environment that determines public health capacity. Economic growth elevates material living conditions and provides fiscal resources for healthcare expansion; however, without deliberate prioritization of public health infrastructure, infection rates may persist despite increased wealth. From the perspective of different indicators, the ratio of the value added by the tertiary industry to that of the secondary industry is an important measure of the optimization and upgrading of a country’s or region’s industrial structure. An increase in the share of the tertiary industry typically indicates a shift toward a service-oriented economy, accompanied by an improved production environment and a relative reduction in occupational hazards. For example, compared to traditional manufacturing, high-tech industries demonstrate better resource utilization efficiency and environmental friendliness. Certain sectors of the tertiary industry, closely related to residents’ health—such as pharmaceutical research and development, healthcare services, and professional technical services—directly reinforce societal health by promoting longevity and expanding access to scientific advancements. These sectors generally have high knowledge intensity and technical added value, effectively promoting high-quality employment, facilitating the transformation and application of scientific and technological achievements, and driving the continuous optimization and upgrading of the economic structure. This, in turn, provides solid support for the realization of the “Healthy China” strategic goals. The ratio of total imports and exports to GDP reflects the degree of openness. In the context of globalization, urbanization is closely linked to international trade. A higher share of imports and exports indicates frequent exchanges between cities and the outside world, which could bring advanced medical technologies but also increases the risk of cross-border transmission of infectious diseases. From July to August 2020, eight cases of COVID-19 were detected on the packaging of imported cold-chain products in Liaoning, Fujian, Jiangxi, Chongqing, Yunnan, Shandong, and Anhui, highlighting that infectious viruses can be transmitted to cities through international business travel and goods transportation (39). The per capita disposable income of urban residents influences their capacity for health-related consumption. As income rises, residents are more willing and able to undergo regular check-ups, vaccinations, and early treatment, enabling “early detection, early diagnosis, and early treatment,” and affording better medical services and medications. The proportion of investment in industrial pollution control relative to GDP reflects the extent of industrial pollution management. Insufficient investment in this area could result in industrial pollution damaging residents’ respiratory and digestive systems, thereby increasing the incidence of infectious diseases. For example, six out of the eight major pollution incidents that occurred during the golden age of capitalism between the 1950s and 1960s were linked to industrial pollution. Although the weight of economic factors is low, they still form the material foundation for the establishment of public health and medical systems. Therefore, during the urbanization process, governments need to focus on coordinating economic development and environmental protection. By increasing investments in industrial pollution control, promoting industrial structure upgrading, and raising residents’ income levels, a solid economic foundation for public health can be created. For instance, the government could encourage businesses to adopt environmentally friendly production technologies and develop green industries to reduce the impact of industrial pollution on residents’ health.

In the eastern region, social factors dominate, signifying that public health outcomes rely heavily on efficient allocation of public service resources. This aligns with the characteristics of the mature urbanization stage in the east, where high population concentration forces the optimization of medical resources and infrastructure. This is reflected in the proportion of health expenditure in total fiscal expenditure and the number of healthcare institution beds per 10,000 population, both around 0.15, while urban water supply coverage rate reaches 0.262. Equalization of public services can effectively break the transmission chain of infectious diseases.

In the central region, economic factors exert greater weight, exposing structural tensions characteristic of mid-industrialization phases. The weight of industrial structure transformation indicators (ratio of tertiary industry value-added to secondary industry value-added) and openness indicators (ratio of foreign trade dependence) approaches 30%, driving the risk of infectious diseases. This corresponds to the “Environmental Kuznets Curve” theory—during industrial transfer in the central region, lagging environmental regulations increase occupational exposure risks, making it more likely for infectious disease outbreaks.

In contrast, the western region is defined primarily by land-based determinants. The ecological environment in the west is relatively fragile, and excessive urban expansion may damage the original vegetation and soil structure, leading to desertification, soil erosion, and more dust storms. Dust carries bacteria, viruses, and other pathogens, which can increase the incidence of respiratory infectious diseases. Some western regions are natural plague foci, where wild animals serve as hosts or vectors for infectious diseases. When their habitats are disturbed, increased contact with humans can raise the risk of zoonotic disease transmission (38). The eastern region should provide high-quality medical and health resources and improve the public health security system. The central region should alleviate the contradiction between the economy and the environment during the process of industrialization and pursue green and sustainable development. The western region should attach importance to the protection of natural ecology and curb the spread of zoonotic diseases.

We explicitly acknowledge slight overfitting signals in the training phase and underfitting on the test sets. The gap mainly reflects (i) small-sample, high-heterogeneity provincial panels, (ii) unobserved spatial spillovers and interactions not modeled her, and (iii) structural breaks during rapid urbanization and the COVID-19 period. Our design mitigates—but cannot eliminate—these risks via cross-validation, repeated shuffling, and prioritizing out-of-sample metrics; thus, results should be interpreted with caution.

This study has several limitations. Due to methodological constraints, we did not account for spatial spillover effects among provincial units or interactions between different factors. Future research could be further advanced in multiple dimensions. On the one hand, it is necessary to incorporate the spatial spillover effects among provincial administrative units and the interaction effects of different factors. By constructing a more comprehensive model, this study can accurately capture inter-regional mutual influences and the combined effects of factors, thereby providing precise policy recommendations for regional coordinated development. On the other hand, further investigate the nonlinear relationships between specific variables and public health outcomes, and conduct long-term follow-up studies to observe the changing trends of impacts and long-term cumulative effects. Meanwhile, expand the research scope and sample size, integrate multidisciplinary approaches, and comprehensively apply theories and technologies from different disciplines to conduct in-depth research from multiple perspectives, so as to obtain more comprehensive and in-depth findings.