Category: 3. Business

Introduction

Health resources serve as the foundation for the advancement of health services, and their effective allocation not only impacts the health status of residents but also significantly contributes to the sustainable progress of medical and health services.¹ Efficiency is regarded as a primary objective in public health management and is one of the fundamental principles promoted by the World Health Organization. Currently, many developing countries face challenges related to low efficiency in resource management,² making the rational allocation of health resources an increasingly relevant global issue.³ The “Healthy China 2030” Outline Plan suggests that by 2030, China’s healthcare system will be more comprehensive, the development of the healthcare sector will be more coordinated, healthy lifestyles will be widely adopted, and the quality and security of healthcare services will continue to improve. Improving health resource efficiency is one of the core measures to achieve these goals, as it enhances the accessibility, equity, and sustainability of health services. This study directly responds to the policy call by providing empirical evidence on efficiency gaps and identifying areas for optimization, particularly in underdeveloped ethnic minority regions such as Guangxi. The rational allocation and effective utilization of health resources are crucial prerequisites for enhancing public health and promoting health equity. Improving the efficiency of health resource allocation not only contributes to the accessibility and quality of healthcare services but also plays an important role in improving the standard of living and promoting economic growth. Since the implementation of the new medical reform in 2009, China has seen a general increase in health resource inputs. Between 2010 and 2022, there was a significant rise in the number of health institutions, hospital beds, and health technical personnel, with increases of 10.24%, 103.65%, and 98.37%, respectively. However, despite the growing quantity of health resources in China, the efficiency of their allocation remains low. In particular, in the western minority region of Guangxi, recent studies have shown that the efficiency of health resource allocation is significantly lower than that of other western provinces, such as Sichuan and the Ningxia Hui Autonomous Region.⁴

The Guangxi Zhuang Autonomous Region (20°54′-26°20′ N, 104°26′-112°04′ E) is located in the western part of China. It is one of the five major ethnic minority autonomous regions in China and the only coastal minority autonomous region in the western part of the country. As an important hub of the Maritime Silk Road, it holds a unique position in the strategy of Western China’s development and the nation’s overall opening-up. According to the Seventh National Census, Guangxi has a population of 50.13 million, making it the most populous ethnic minority province in China. The region’s terrain is complex, with mountainous areas accounting for 70.8% of its landscape, posing significant transportation challenges. It is also generally regarded as an economically underdeveloped region,⁵ with a GDP of 2.63 trillion yuan in 2022, ranking 19th nationwide. Its per capita GDP is only 52,164 yuan, ranking third from the bottom in the country. In recent years, Guangxi has been focused on building health institutions to meet the health needs of its residents. While there has been significant change in the total amount of health resources in Guangxi, little is known about the efficiency of health resources within the region and its variations.

To gain a more comprehensive understanding of, and enhance, the efficiency of health resource allocation in the region, it is essential to employ rigorous quantitative methods. At the same time, recent studies have highlighted the importance of integrating efficiency evaluation with modern information technologies — such as smart healthcare systems, the Internet of Things, and artificial intelligence — to strengthen the resilience and adaptability of healthcare systems.^6–10 DEA provides a valuable approach for assessing the efficiency of health resource allocation in this context. It is particularly suitable for underdeveloped or ethnically diverse regions such as Guangxi, as it can accommodate multiple inputs and outputs without requiring a predefined production function. This flexibility is especially beneficial in heterogeneous settings where healthcare systems face diverse constraints in terms of geography, infrastructure, and population needs. In recent years, several studies have measured and analyzed healthcare efficiency from different perspectives. First, studies at the national level primarily focus on evaluating the efficiency of different countries. For example, Aydin et al¹¹ assessed the efficiency of healthcare services in the Organization for Economic Co-operation and Development (OECD) countries, while Top et al¹² measured the efficiency of healthcare systems in 36 African nations. Second, some scholars have investigated efficiency at the regional level within a single country. For instance, Mazon evaluated the technical efficiency of municipalities in Santa Catarina, Brazil, in terms of public health expenditure and its relationship with health management.¹³ Similarly, Ngobeni assessed and compared the technical efficiency of healthcare delivery across the nine provinces of South Africa.¹⁴ Additionally, several studies have examined the efficiency of different types of hospitals. For example, researchers have compared the efficiency of teaching and non-teaching hospitals in the United States,¹⁵ while in Iran’s southwest region, studies have analyzed the efficiency of general hospitals, specialized hospitals, and multi-specialty hospitals.¹⁶ Similar analyses have been conducted in Saudi Arabia, comparing general hospitals, specialized hospitals, and primary healthcare centers.¹⁷ In the case of China, Gong et al evaluated the overall and two-stage efficiency of provincial healthcare systems,¹⁸ while Du examined the relationship between healthcare quality and efficiency across the national, eastern, central, and western regions.¹⁹ Jing et al²⁰ analyzed efficiency differences between public and private hospitals in Beijing. Following China’s healthcare system reforms, government agencies and scholars have paid significant attention to efficiency at the national level,¹⁸ in developed regions,²¹ and within primary healthcare services.²² However, relatively less focus has been given to the efficiency of healthcare services in underdeveloped ethnic minority regions.

The selection of appropriate input and output indicators is crucial for meaningful analysis.²³ Most researchers choose input indicators for hospital performance evaluation based on labor and capital inputs.^{18,19,24–27} The number of health institutions and hospital beds is commonly used to represent capital input.^18,19,25 For labor-related variables, health workers,^18,26 health technical workers,^19,25 physicians,^12,24 and nurses^12,24,27 are frequently considered as types of labor input. Regarding output indicators, they are generally categorized into expected and undesirable outputs. Due to data availability constraints, most studies focus on expected indicators in their analyses. Outpatient or emergency visits,^{18,19,25–27} inpatient admissions or discharges,^18,19,25,27 bed turnover rate,^23,28 and medical revenue^25,29 are commonly regarded as expected outputs. In contrast, infection incidence, the number of infectious disease cases, and patient or population mortality rates^24,30 are considered undesirable outputs.

Overall, researchers have provided valuable insights into the selection of input and output indicators. However, there are still several shortcomings. First, the frequent combination of absolute and relative figures in their analyses is prevalent. For instance, numerous studies simultaneously use the bed turnover rate and the number of visits as input indicators, which is not justifiable. Second, many researchers, while focusing on the input and output efficiency of health resources, have failed to consider the influencing factors. Consequently, we have made improvements based on these studies. In the context of minority regions, this study employs the BCC-DEA model and the Malmquist index model to conduct both static and dynamic analyses of the efficiency of health resource allocation in Guangxi from 2010 to 2022. By combining the Tobit regression model, we explore the factors influencing its productivity, analyze the problems and their causes, and offer rational suggestions to provide a reference for the scientific formulation of health development plans.

Materials and Methods

Data Envelopment Analysis

This study employs the Data Envelopment Analysis (DEA) approach to evaluate the efficiency of resource allocation. DEA is a technique introduced by researchers like Charnes and Cooper in 1978, designed to assess the relative efficiency of Decision Making Units (DMUs) that utilize multiple inputs and outputs. Subsequently, after years of adjustments and improvements, various model types have been developed, including the CCR model and the BCC model. The difference between the two lies in the assumption that the CCR model assumes constant returns to scale (CRS), while the BCC model is suitable for variable returns to scale (VRS) situations.³¹ This study employs the BCC-DEA model under variable returns to scale (VRS) and adopts an input-oriented approach. The rationale for this choice is as follows:

In this study, we adopt an input-oriented BCC-DEA model to evaluate the efficiency of health resource allocation. This choice is particularly relevant for underdeveloped or ethnically diverse regions such as Guangxi, where resource scarcity and input constraints are more critical than output expandability. First, in many healthcare systems in economically disadvantaged regions, inputs such as medical personnel, equipment, and infrastructure are limited, making input reduction and resource optimization key priorities for local governments. Second, output levels (eg, outpatient visits, bed utilization) are often influenced by external factors—such as population health needs or migration—that are harder to control in the short term. Thus, focusing on minimizing input waste provides a more stable and actionable efficiency assessment framework.

Additionally, we apply the BCC model under variable returns to scale (VRS), which is more suitable for capturing efficiency in regions with diverse healthcare system sizes and institutional capacities. Unlike the CCR model, which assumes constant returns to scale, the BCC model accommodates scale inefficiencies that are common in fragmented or unevenly developed healthcare settings. This flexibility enables us to distinguish between pure technical efficiency and scale efficiency, thereby offering more nuanced policy insights for both input management and system scaling.

The BCC-DEA model allows for the decomposition of the overall efficiency score of decision-making units into scale efficiency (SE) and pure technical efficiency (PTE) when variable returns to scale are present. The relationship among these components can be expressed as: Overall Efficiency = Pure Technical Efficiency × Scale Efficiency. In this research process, various cities of Guangxi are selected as actual decision-making units (DMUs). Assuming there are n decision-making units, where j =1, 2, …, n, each decision-making unit has a types of input and b types of output. That is, the input quantity X_j = (X_1j, X_2j, …, X_aj)^T and the output quantity Y_j = (Y_1j, Y_2j, …, Y_bj)^T, where X_ij represents the input quantity of the j-th decision-making unit for the i-th input indicator, and Y_rj represents the output quantity of the j-th decision-making unit for the r-th output indicator. The formulas are given as follows:

In the above expression, θ represents the efficiency value of DMUj, λj denotes the weight of the decision-making unit, Xj represents the input quantity of the j-th decision-making unit, Yj represents the output quantity of the j-th decision-making unit, n is the number of decision-making units, S ⁺ and S ⁻ represent the slack variables for inputs and outputs, respectively. Therefore, when evaluating the effectiveness of decision-making unit DMUj, corresponding results can be obtained based on the optimal solution values of the model.

(1) If θ=1, and S ⁺ and S ⁻ are both 0, it indicates that DMUj is in a DEA efficient state;

(2) If θ=1, and S ⁺ and S ⁻ are not both 0, it indicates that DMUj is in a DEA weakly efficient state;

(3) If θ<1, and S ⁺ and S ⁻ are not both 0, it indicates that DMUj is in a non-DEA efficient state.

To examine the robustness of the efficiency estimates, we further re-estimated the DEA scores using an output-oriented BCC model and a constant-returns-to-scale (CCR) model. The results were compared to those from the original input-oriented BCC model to assess the consistency across model specifications.

Bootstrap and Statistical Inference

To assess the statistical precision of the DEA efficiency estimates, we employed a bias-corrected and accelerated (BCa) bootstrap approach with 2,000 resamples. This method corrects for bias and skewness in the sampling distribution, providing confidence intervals around the efficiency scores of each DMU.

Additionally, we applied the Kruskal–Wallis rank sum test to evaluate whether efficiency scores differed significantly among the 14 cities. This non-parametric test is suitable for comparing multiple independent samples when the normality assumption is not guaranteed.

Serial Correlation Test in Panel DEA

To assess whether time-dependent structures in the data could bias the estimation of the Malmquist productivity index, we performed a serial correlation test following the bootstrap-based approach proposed by Simar and Wilson (2007). This test examines whether DEA efficiency scores in period t are statistically dependent on those in period t–1. The test was conducted using 2,000 bootstrap replications.

Malmquist Index Model

Traditional DEA models primarily facilitate static assessments of resource allocation efficiency using cross-sectional data. However, this study spans a longer time frame, necessitating consideration of temporal variations. The Malmquist index provides a means to evaluate changes in efficiency over a specified period, reflecting the performance of decision-making units.³² Consequently, we implemented the Malmquist Productivity Index (MPI) method to analyze panel data and illustrate dynamic shifts in efficiency. The MPI, also known as Total Factor Productivity Changes (TFPCH), is derived from the distance function and can be represented by the following mathematical equations:

To thoroughly understand the technical level during both periods, we took into account the geometric mean:

The productivity function can be categorized into input-oriented efficiency change (EFFCH) and technical change (TECHCH). Additionally, efficiency change can be further broken down into scale efficiency change (SECH) and pure efficiency change(PECH).

Tobit Regression

Given that the overall efficiency value is a restricted dependent variable and is categorized into different stages, we employed the Tobit regression model to mitigate biases associated with the least squares regression method. Introduced in 1958, Tobit regression utilizes the maximum likelihood approach to randomly select n groups of samples for estimating maximum probabilities.³³ In our analysis, we treated the overall efficiency value from the DEA model as the dependent variable, while the influencing factors were designated as independent variables. The Tobit models are as follows:

where Yi* represents the dependent variable, Xi denotes the explanatory variable, and βi is the coefficient associated with the explanatory variable, where, i = 1, 2, …, n.

To ensure model validity, we reported diagnostic statistics including the log-likelihood, pseudo-R², and conducted a likelihood-ratio test for heteroskedasticity. The left-truncation point was set at 0 to reflect the lower bound of DEA efficiency scores. However, to address the potential limitations of the Tobit model—particularly issues related to endogeneity and bias from censoring at the boundaries—we further adopted a two-stage bootstrap truncated regression approach, following Simar and Wilson (2007). This method corrects for bias and provides consistent estimators of the relationship between efficiency scores and their determinants. The first stage involves estimating DEA efficiency scores, while the second stage applies truncated regression with bias-corrected bootstrap confidence intervals based on 2,000 replications. This approach allows for more robust inference and minimizes the risk of misleading conclusions due to serial dependence or data truncation.

Data Resources and Regional Division

Demographic, economic, and geographic data were gathered from the Guangxi Statistical Yearbook for the years 2011 to 2023. Information regarding health resources was extracted from the Guangxi Health Statistics Yearbook covering the same period. It is important to mention that the data presented in each year’s edition of the yearbook reflects figures from the preceding year. Guangxi province consists of 14 cities, According to the geographical position and the level of the GDP per capita, all the 14 cities were divided into three groups: Northern, Middle and Southern regions. The Northern region included Liuzhou, Guilin, Hechi, Baise. The Middle region included Hezhou, Wuzhou, Laibin, Guigang, Yulin. The Southern region included Nanning, Chongzuo, Qinzhou, Beihai, Fangchenggang.

Data analysis: Statistical analyses were conducted using STATA 18 and SPSS 24 software. All tests were performed at a 5% significance level (p < 0.05).

As the study relied on secondary data, there was no need for direct involvement from patients or the public.

Input and Output Variables

The efficiency indicators were chosen based on the literature previously reviewed^{18,19,25–27} and to ensure the credibility of the results (the total of input and output indicators should not exceed the total value of the DMU).¹¹ While the bed turnover rate is commonly used in China, some studies have noted the inappropriate mixing of absolute and relative figures,²³ leading this research to exclude that indicator. Additionally, since DEA focuses on efficiency, this study does not incorporate quality metrics such as mortality and cure rates. For input indicators, we selected health technical personnel, the number of medical institutions, and actual beds to represent human resources and capital. As for output indicators, we chose the total number of visits and discharges to capture both outpatient and inpatient services (Table 1).

Table 1 Summary Statistics of the Variables

To ensure that the selected input and output variables are theoretically sound and statistically valid, we conducted a Pearson correlation analysis and a principal component analysis (PCA). Table S1 presents Pearson correlation coefficients among the five variables, all of which are statistically significant at the 1% level. Furthermore, the PCA was conducted to explore the internal structure of the variables and avoid redundancy. The Kaiser-Meyer-Olkin (KMO) measure was 0.826, and Bartlett’s Test of Sphericity was significant (χ² = 1412.20, p < 0.001), indicating sampling adequacy (Table S2). As shown in Table S3, the first principal component explained 86.15% of the total variance, with all five variables loading highly on this component (Table S4). These results confirm that the variables share common structure and are appropriate for inclusion in a unified DEA model.

Explanatory Variables Affecting Efficiency

The literature indicates that the factors influencing the efficiency of health resource allocation can be categorized into four main aspects: economic, demographic, social, and policy-related factors.^3,34–37 Economic factors primarily include regional economic development levels, per capita disposable income, the degree of healthcare marketization, etc. Demographic factors mainly encompass population size and structure, urban-rural distribution, the scale of the floating population, etc. Social factors involve education levels, healthcare service demand, medical technology levels, etc. Policy-related factors include government health policies, health insurance systems, the proportion of health expenditure, etc.

After reviewing the literature, we identified candidate variables based on data availability. Subsequently, two rounds of expert consultations were conducted to discuss which explanatory variables should be selected from the candidate variables. Economic factors were represented by per capita GDP (measured as real per capita GDP to reflect the level of economic development, calculated as the ratio of the actual GDP of each city to its population, adjusted using the base-year CPI) and residents’ income levels (measured as the annual per capita disposable income in different regions). Second, population factors were captured by population density (measured as the number of people per square kilometer of land area) and urbanization rate (measured as the percentage of the urban population relative to the total population). Third, social factors were reflected in education level (measured as the proportion of primary and secondary school graduates in the total population) and the proportion of health technical workers (measured as the ratio of health technical workers to total health personnel). Finally, policy factors were represented by the number of people enrolled in basic medical insurance and the proportion of government health expenditures in total fiscal spending. Statistical descriptions of these variables are presented in Table 1.

Results

Descriptive Analysis

Table 2 shows that from 2010 to 2018, the total number of visits and beds in Guangxi Province exhibited an upward trend year by year. The number of institutions increased annually from 2010 to 2017, dropped sharply from 2018 to 2020, and then increased slowly from 2021 to 2022. The number of health technical personnel increased year by year from 2013 to 2022, with a notable decrease from 2011 to 2012. The number of discharges from hospitals increased annually from 2010 to 2022, but deviated from this trend in 2021.

Table 2 Descriptive Statistics of Inputs and Outputs

The Efficiency of Health Resource Allocation in Guangxi

This study employs the BCC-DEA model to measure the efficiency of health resource allocation in 14 cities in Guangxi from 2010 to 2022 (Table 3). The average overall efficiency score of the sample is 0.675, which does not meet DEA effectiveness, indicating that these regions still have varying degrees of room for improvement. From a temporal perspective, the overall efficiency score shows an upward trend from 2010 to 2022, increasing from 0.558 in 2010 to 0.879 in 2022. From a regional perspective, the average overall efficiency scores across different cities range from 0.530 to 0.789, with Qinzhou having the highest score (0.789). The cities that need the most improvement in resource allocation efficiency are Wuzhou (0.530), Yulin (0.590), and Hezhou (0.604).

Table 3 Efficiency Values of the 14 Regions in Guangxi in 2010–2022

Overall efficiency can be decomposed into technical efficiency and scale efficiency. From the perspective of the pure technical efficiency index, the average pure technical efficiency of health resources across Guangxi’s cities was 0.646 in 2010 and increased to 0.912 in 2022, reflecting a rise of 0.265 over time. From a regional perspective, the number of regions achieving DEA effectiveness in pure technical efficiency is lower than that in scale efficiency. Regarding the scale efficiency index, the average scale efficiency of health resources across Guangxi’s cities was 0.879 in 2010 and increased to 0.962 in 2022, consistently exceeding the average pure technical efficiency of health resources. From a regional perspective, the scale efficiency values of all cities remain below 1 (Tables S5 and S6).

Table S7, most cities operated under increasing returns to scale (IRS) throughout the study period, while very few reached constant returns to scale (CRS). This pattern indicates that the majority of decision-making units had not yet attained their optimal operational scale. To test the robustness of the efficiency scores, we estimated alternative models using output-oriented BCC and input-oriented CCR assumptions. The results remained identical across all models (Table S8), indicating strong consistency and confirming the reliability of the original input-oriented findings.

Bootstrap Confidence Intervals and Inter-City Comparisons

Table S9 presents the bias-corrected and accelerated (BCa) bootstrap confidence intervals for the DEA efficiency scores of each city. The results show moderate variation in the lower and upper bounds across cities, suggesting differences in score reliability. To further test the significance of inter-city differences, we conducted a Kruskal–Wallis rank sum test. The test yielded a significant result (χ² = 56.521, df = 13, p < 0.001), indicating that efficiency levels vary significantly among the 14 cities (Table S10).

The Input-Output Optimization Plan for the Allocation of Regional Health Resource

In the BCC model, we also assessed the projected values for the input and output indicators of the inefficient regions (Table 4). To enhance the efficiency of health resource allocation, these regions should either reduce their inputs or boost their outputs. Since the study covers a long time span, the latest data from 2022 is used as an example. For instance, the input-output projection analysis for Nanning reveals that to achieve a more optimal allocation of resources, the region could decrease its average health personnel count by 7,213.64 and the average number of beds by 1,068.31, while maintaining current output levels. Alternatively, Nanning could aim to increase outpatient and emergency visits by 1,880.72 without altering its present input levels.

Table 4 Variation of Inputs and Outputs Needed to Be Adjusted in 2022

Serial Correlation Test for Panel DEA

Prior to Malmquist index estimation, we conducted a serial correlation test to assess whether time-dependent bias might affect the efficiency scores. Specifically, the Simar–Wilson (2007) bootstrap-based approach was applied using 2,000 replications to test the lag-1 correlation of DEA scores. As shown in Table S11, the test revealed no statistically significant serial correlation (mean lag-1 correlation = 0.256, p = 0.1284), suggesting that the use of the conventional Malmquist index is appropriate.

The Productivity of Health Resource Allocation in Guangxi

The MPI of annual means was utilized to analyze productivity changes from 2010 to 2022, as presented in Table 5. The geometric mean of TFPCH stood at 1.046, signifying a 4.6% increase in the productivity of health resource allocation in Guangxi Province over the period from 2010 to 2022. Further delving into the reasons behind this increase, it was primarily attributed to a 4.2% rise in TECHCH. As for EFFCH, it exhibited a modest 0.4% growth, which was solely due to a corresponding 0.4% increase in PECH (Pure Efficiency Change, referring to changes in the efficiency of resource utilization without changes in technology). The year with the highest TFPCH was 2018–2019 (1.137), and the year with the lowest TFPCH was 2016–2017 (0.918).

Table 5 Malmquist Index Analysis of Regional Health Resource Efficiency

Similarly, as shown in Figure 1, the TFPCH rate exhibits two peaks and one trough. The peaks occurred during 2018–2019 and 2021–2022, while the trough was observed during 2016–2017. The TECHCH index exhibited substantial variation, especially declines in 2015–2017 and 2020–2021, suggesting that stagnation or regression in technological advancement had a more pronounced effect on productivity dynamics. Figure 2 depicts the annual average value of the Malmquist index for 14 cities from 2010 to 2022. Among these cities, only Qinzhou has a TFPCH less than 1, while Yulin boasts the highest total factor productivity rate of 1.081. However, the EFFCH value of 5 cities fell below 1, and the PECH and SECH values of 4 cities were also less than 1.

Figure 1 Interannual analysis of regional health resource efficiency.

Figure 2 Analysis on the efficiency of regional health resources at the city level.

Tobit Analysis of Health Resource Allocation

The Tobit regression results indicate that residents’ income levels (β = 0.252, 95% CI = 0.000–0.505) and education levels (β = 5.538, 95% CI = 0.764–10.312) are positively associated with efficiency, both significant at the 5% level. In addition, population density (β = –0.0005, 95% CI = –0.0007 – –0.0002) and the proportion of government health expenditure (β = 0.011, 95% CI = 0.0039–0.0173) show statistically significant associations at the 1% level, with population density negatively and health expenditure positively influencing efficiency (Table 6).

Table 6 Tobit Regression Analysis of the Allocation Efficiency of Health Resources (N=112)

However, to address potential biases arising from the bounded nature of DEA efficiency scores and the possible endogeneity of explanatory variables, a two-stage bootstrap truncated regression was also conducted, as shown in Tables S12 and S13. Results from this more robust method confirmed the significance of residents’ income level, population density, and the proportion of government health expenditure, while the previously significant effect of education level was no longer statistically supported. This discrepancy highlights the importance of using bias-corrected methods to ensure the reliability of inferences drawn from efficiency analyses. All other variables were not significantly associated with the efficiency of health resource allocation in both the Tobit and the two-stage bootstrap truncated regression models (p > 0.05).

Discussion

In general, due to the increase of population and the increase of demand for health services, the allocation of health resources in Guangxi is also increasing year by year. In response to the WHO’s requirement that everyone have access to basic health services, China has carried out a number of comprehensive healthcare reforms since 2009, including enhancing the efficiency of health resource allocation to increase people’s access to health services. This study calculated the effective allocation of health resources in Guangxi minority provinces from 2010 to 2022, which can explain the effect of the health system reform started in 2009 to a certain extent, and also provide some suggestions for future health system reform. Furthermore, Guangxi is located in the west, which is the province with the largest population of the five major ethnic minority areas in China, so it is representative to choose Guangxi as the sample for analysis.

A well-functioning health service system is often characterized by its high efficiency. Based on the Malmquist index, the average total factor productivity change (TFPCH) for health resource allocation efficiency in Guangxi from 2010 to 2022 exceeded 1, suggesting a positive trend in productivity. The TFPCH can be decomposed into two main components: technical efficiency change (EFFCH) and technological change (TECHCH). Furthermore, EFFCH can be further divided into pure technical efficiency change (PECH) and scale efficiency change (SECH). As illustrated in Figure 1, the increase in TFPCH observed in this study is attributed to improvements in both EFFCH and TECHCH, as both indicators remained above 1 during the study period. This suggests progress not only in the management and organization of health resources but also in the technological capabilities of the healthcare system in Guangxi. The findings highlight the urgency of enhancing technological advancement and innovation within the health sector.³⁸ Research has shown that such improvements may be linked to factors such as the education level of healthcare professionals and the utilization of medical services.³⁹ Additionally, Our study shows that the growth in EFFCH is primarily driven by gains in PECH. In contrast, SECH emerges as a limiting factor, indicating that the scale of operation in some regions may not be optimal. Therefore, to further improve the efficiency of health resource allocation and support the sustainable development of health services, Guangxi and its municipalities should prioritize technological innovation and optimize scale efficiency. Strengthening the service capacity of primary healthcare institutions, fostering collaboration across different levels of medical facilities, and improving training and incentive mechanisms for healthcare personnel are essential steps in this direction.⁴⁰

The DEA efficiency analysis results indicate that the average efficiency score of health resource allocation in Guangxi from 2010 to 2022 was 0.675. This score exceeds that of Bratislava (0.564), Saudi Arabia (0.624), and Turkmenistan (0.639),⁴¹ as well as Australia (0.588) and Denmark (0.629).⁴² However, it remains lower than that of Spain (1.016), Taiwan, China (0.973), and Changsha, China (1.000).⁴³ Furthermore, when benchmarked against domestic minority or less-developed provinces using similar DEA approaches, Guangxi’s average efficiency score (0.675) is comparable to Inner Mongolia (0.690)⁴ and Qinghai (0.708),⁴⁴ but higher than Xinjiang (0.590).⁴ These comparisons suggest that Guangxi’s performance is not uniquely low, but rather reflects a broader pattern of resource allocation challenges faced by western and ethnic minority regions in China. This pattern is also consistent with findings from rural regions in other middle-income countries, such as Brazil, South Africa, and Indonesia, where geographic disparities and infrastructure constraints have similarly hindered efficient resource allocation.^45–47 These parallels further reinforce the relevance of Guangxi’s experience to global health system reform discussions in resource-constrained settings. Notably, in 2022, only 28.57% of regions in Guangxi achieved high efficiency in health resource allocation, while more than 70% were found to be relatively inefficient. This underlines the structural and regional disparities that continue to hinder the optimal utilization and spatial distribution of health resources in the region. Improvements are needed in the efficiency of resource utilization and the rational planning of regional health resource distribution.⁴⁸ In remote areas, residents may have limited access to high-quality healthcare services, leading to deteriorating health conditions and increased medical demand. At the same time, some resources may be underutilized, resulting in wasted or misallocated medical resources. From the decomposition of overall efficiency, it is evident that pure technical efficiency has not reached an optimal level. This may be due to the misallocation of health human resources, preventing the full utilization of technical capabilities, as well as the irrational distribution of medical equipment, resulting in low utilization rates.⁴⁹ Scale efficiency in health resource allocation reflects whether resource supply is at an optimal scale. Our study indicates that scale efficiency is also below 1, which could be attributed to several factors. First, in some regions, the size of medical institutions may be either too large or too small, leading to a mismatch between the supply of health resources and actual demand. Second, there is an excessive reliance on large hospitals, while the service capacity of primary healthcare institutions remains insufficiently developed. This imbalance results in a significant patient influx into large hospitals, while primary institutions remain underutilized. Finally, disparities in economic development across Guangxi contribute to inefficiencies in resource allocation. High-quality medical resources are often concentrated in economically developed cities, whereas underdeveloped regions suffer from inadequate healthcare infrastructure and service capacity, ultimately reducing the overall efficiency of health resource allocation. These findings are consistent with the conclusions of Zheng D’s study.⁵⁰ Although the scale efficiency scores were, on average, higher than the pure technical efficiency scores, the dominance of IRS among DMUs suggests that scale inefficiency remains a persistent and structural issue in the regional healthcare system. The analysis of predicted input-output values indicates that relatively low efficiency in health resource allocation is primarily observed in economically disadvantaged areas of Guangxi, such as Beihai and Chongzuo, which may be related to local economic development. The analysis of predicted output indicators reveals significant potential for improvement in both the total number of visits and the number of discharges, with the latter showing particularly substantial room for enhancement. Before the government decides to reduce health resources or expand health services to improve allocation efficiency, further investigation into the factors influencing health resource allocation efficiency is crucial.⁴³

Studies have suggested that health resources in economically disadvantaged areas may be better utilized, potentially improving efficiency.⁴¹ The Tobit regression results of this study indicate that per capita GDP does not have a significant impact on health resource allocation efficiency. This finding is consistent with the study by Zhong K et al, who attributed it to the guiding role of government policies and the implementation of healthcare reform, which have helped narrow the regional disparities in health resource allocation.³⁶ Therefore, the impact of regional economic levels on resource allocation efficiency appears to be relatively limited. However, our results show that per capita disposable income has a positive effect on resource allocation efficiency, which aligns with the findings of Gong J.³ Residents’ purchasing power directly affects healthcare demand and utilization. Individuals with higher disposable income are more likely to seek regular medical care and health check-ups, thereby increasing the utilization of health resources. Additionally, they tend to have greater health awareness and are more willing to invest in their well-being. However, our study indicates that the urbanization rate does not have a significant impact on the efficiency of health resource allocation, possibly due to mismatches between the supply and demand of medical resources and insufficient policy support. This finding is consistent with the study by Liang B.³⁴ In contrast, population density exhibits a negative effect on resource allocation efficiency. Several factors may contribute to this result. First, excessive healthcare demand in densely populated areas can lead to medical facilities operating beyond capacity, prolonged patient waiting times, and overworked healthcare professionals, ultimately reducing resource utilization efficiency.⁵¹ Second, in high-density regions, health resources are often concentrated in major cities or large hospitals, while primary healthcare institutions remain underutilized, leading to inefficiencies in resource allocation.⁵² Additionally, some densely populated areas may suffer from inadequate planning of medical institutions, resulting in redundant construction or suboptimal distribution of resources, further weakening overall allocation efficiency. This finding is supported by the research of El Husseiny⁵³ and Ahmed et al.⁴¹

Moreover, it is noteworthy that while previous studies have generally suggested that health technical personnel influence the efficiency of health resource allocation,⁵⁴ our study reached the opposite conclusion. Several factors may explain this discrepancy. First, in some impoverished areas, an increase in medical personnel does not necessarily lead to improved resource utilization efficiency, particularly in cases where infrastructure is underdeveloped, medical equipment is insufficient, or the hierarchical medical system is not well-established.⁵⁵ Second, if a significant portion of health personnel are assigned to administrative roles or inefficient departments, a higher proportion of medical staff may not result in a substantial efficiency improvement.⁵⁶ Similarly, our study found that the number of people enrolled in medical insurance does not have a significant impact on the efficiency of health resource allocation. Possible explanations include the fact that despite high insurance coverage, inefficiencies such as excessive medical treatment, redundant examinations, overprescription of medications, and the misuse of medical resources may still exist.⁵⁷ Additionally, in China, many insured residents tend to seek treatment at tertiary hospitals rather than primary healthcare institutions, leading to underutilization of community-level medical resources while overburdening large hospitals. As a result, the overall efficiency of health resource allocation remains unimproved. While education level showed a positive coefficient in the baseline model, its effect did not remain significant in robustness checks and should therefore be interpreted with caution. This suggests that the relationship between education and resource efficiency may be more complex or mediated by other factors not captured in the current model. By contrast, the proportion of government health expenditure was found to have a consistently significant and positive impact on efficiency. Increased public spending—particularly when directed toward primary care, public health infrastructure, and preventive services—can improve the spatial distribution of medical resources,⁵⁸ alleviate the financial burden on patients, and reduce delays in care. These improvements help prevent disease progression, promote early intervention, and ultimately enhance the effectiveness and equity of resource allocation across regions. This reinforces the view that improving health resource efficiency requires not only optimizing quantity, but also addressing structural, institutional, and behavioral factors. By applying the two-stage bootstrap truncated regression, we mitigated potential endogeneity stemming from factors such as income and health investment, while controlling for other relevant variables. The regression results (Tables S12 and S13) confirmed that per capita disposable income and public health expenditure are significant positive contributors to healthcare efficiency in Guangxi, whereas population density exerts a significant negative effect. These findings support the robustness of our model and align with the study’s objective to uncover reliable efficiency determinants. Notably, the bias-corrected confidence intervals reinforce the stability of the estimates, enhancing the reliability of policy recommendations derived from our results.

Limitations

There are several limitations in this study. First, the output indicators mainly reflect service volume due to data availability, while quality-related metrics such as patient satisfaction or clinical outcomes were not included. Second, while Tobit regression is appropriate for censored data, it identifies associations rather than causal relationships. Moreover, the regression analysis may be subject to omitted variable bias due to unobserved factors not included in the model, which could influence the estimated effects. Future studies should consider incorporating quality-adjusted indicators, applying causal inference methods, and including a broader set of explanatory variables to enhance robustness.

Conclusion

This study is based on the latest officially published data and examines the allocation of health resources as well as the provision of health services in ethnic minority areas of Guangxi. It conducts both cross-sectional and longitudinal analyses to assess the static and dynamic efficiency of health resource allocation across different regions and explores the factors influencing allocation efficiency. The main findings of this study are as follows: First, the Malmquist index indicates an overall upward trend in the TFPCH in Guangxi from 2010 to 2022, although significant differences exist among cities within the region. Second, the overall efficiency of health resource allocation in Guangxi during 2010–2022 remained relatively low, primarily due to deficiencies in technical efficiency and scale efficiency. Third, factors positively influencing the overall efficiency of health resource allocation include per capita disposable income and the proportion of healthcare expenditure, while population density exhibits a negative impact. The effect of education level was not consistently significant across robustness checks and should be interpreted with caution.

This finding has policy implications not only for Guangxi but also for other ethnic minority regions and economically disadvantaged areas in China facing challenges in health resource integration. The empirical results of this study highlight three key policy recommendations. First, the government should further advance comprehensive healthcare reforms to enhance the efficiency of health services across cities. Second, in implementing healthcare reforms, policymakers should pay closer attention to regional economic conditions and health resource levels, tailoring policies to the specific needs of different areas. For instance, in regions with stronger economic development and relatively abundant health resources, efforts should focus on improving internal management efficiency within the healthcare system, optimizing institutional operations, and maximizing resource utilization. DEA results suggest that cities like Nanning could increase outpatient visits by nearly 70% without additional inputs, indicating substantial room for efficiency gains through better internal coordination. Conversely, in regions with weaker economic foundations and limited medical resources, short-term priorities should include expanding the total supply of healthcare resources to enhance service accessibility, particularly in remote and underserved areas. However, DEA results also reveal that some less-developed areas, such as Beihai and Fangchenggang, exhibit structurally inefficient resource allocation. For example, these cities could reduce redundant beds by 15–30% without compromising service delivery. This indicates that even in resource-constrained settings, dual strategies—both increasing investment and optimizing existing resource structures—are necessary to improve system performance. These quantified targets offer actionable guidance for local governments to reallocate resources more effectively. Moreover, the empirical evidence helps to explain how the identified socioeconomic factors influence efficiency. Higher income levels likely enhance health-seeking behavior and awareness, promoting more effective use of services and encouraging system responsiveness. In contrast, high population density may strain infrastructure, leading to congestion and lower service quality. The observed output slack in relatively resource-rich regions like Nanning also points to managerial or institutional bottlenecks, reinforcing the need for internal reform and digital innovations such as smart hospital systems and performance-based management. These mechanisms demonstrate how resource efficiency is not only a function of quantity, but also of how well inputs are organized and managed. Third, maintaining stable economic growth is a fundamental prerequisite for ensuring sustained public health investment. It is also essential to further increase the proportion of healthcare expenditure within total government spending to reduce the financial burden of medical care. Additionally, optimizing the spatial distribution of health resources and strengthening public health education are crucial for enhancing health literacy and promoting overall well-being.

Data Sharing Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Acknowledgments

The authors wish to thank all experts who participated in the study. The authors thank Ajuan Tang, Zhe Sun, and Gai Cao for their contributions to data collection, analysis, and interpretation. Special thanks to Rong Cao for designing the study.

Funding

We did not utilize any financial resources.

Disclosure

The authors declare no conflicts of interest in this work.

References

1. Liu Q, Guo Y. Regional differences of individual and allocation efficiencies of health resources in China. Front Public Health. 2023;11:1306148. doi:10.3389/fpubh.2023.1306148

2. Mohammadpour S, Javan-Noughabi J, Vafaee Najar A, et al. Factors affecting the technical efficiency of rural primary health care centers in hamadan, Iran: data envelopment analysis and tobit regression. Cost Eff Resour Allocation. 2020;18(1):1–8. doi:10.1186/s12962-020-00249-1

3. Gong J, Shi L, Wang X, Sun G. The efficiency of health resource allocation and its influencing factors: evidence from the super efficiency slack based model-tobit model. Intl Health. 2023;15(3):326–334. doi:10.1093/inthealth/ihac054

4. Wang L, Sheng Y, Shao S, Du Z. Analysis of the efficiency of provincial health resource allocation based on network DEA model. Chin J Social Med. 2024;41(06):753–757.

5. Wang Z, He H, Liu X, Wei H, Feng Q, Wei B. Health resource allocation in western China from 2014 to 2018. Arch Public Health. 2023;81(1):30. doi:10.1186/s13690-023-01046-x

6. Amiri Z, Heidari A, Zavvar M, et al. The applications of nature-inspired algorithms in internet of things-based healthcare service: a systematic literature review. Trans Emerg Telecommun Technol. 2024;35(6):e4969. doi:10.1002/ett.4969

7. Zanbouri K, Darbandi M, Nassr M, et al. A GSO-based multi-objective technique for performance optimization of block chain-based industrial Internet of things. Int J Commun Syst. 2024;37(15):e5886. doi:10.1002/dac.5886

8. Amiri Z, Heidari A, Navimipour NJ, et al. The deep learning applications in IoT-based bio-and medical informatics: a systematic literature review. Neural Comput Appl. 2024;36(11):5757–5797. doi:10.1007/s00521-023-09366-3

9. Vakili A, Al Khafaji HMR, Darbandi M, et al. A new service composition method in the cloud-based internet of things environment using a grey wolf optimization algorithm and MapReduce framework. Concurrency Comput. 2024;36(16):e8091. doi:10.1002/cpe.8091

10. Toumaj S, Heidari A, Shahhosseini R, et al. Applications of deep learning in Alzheimer’s disease: a systematic literature review of current trends, methodologies, challenges, innovations, and future directions. Artif Intell Rev. 2024;58(2):44. doi:10.1007/s10462-024-11041-5

11. Sevim F, Şenol O, Akbulut F. Evaluation of the efficiency of long-term care services in oecd countries by dea method. Ebelik ve Saglik Bilimleri Dergisi. 2024;7(2):376–387. doi:10.62425/esbder.1509324

12. Top M, Konca M, Sapaz B. Technical efficiency of healthcare systems in africana countries: an application based on data envelopment analysis. Health Policy Technol. 2020;9(1):62–68. doi:10.1016/j.hlpt.2019.11.010

13. Mazon LM, Freitas SFTD, Colussi CF. Financing and management: the technical efficiency in public health expenditures of small-scale municipalities in the state of santa catarina. Ciênc Saúde Coletiva. 2021;26(4):1521–1532. doi:10.1590/1413-81232021264.09712019

14. Ngobeni V, Breitenbach MC, Aye GC. Technical efficiency of provincial public healthcare in South Africa. Cost Eff Resour Allocation. 2020;18(1):1–19. doi:10.1186/s12962-020-0199-y

15. Silber JH, Rosenbaum PR, Niknam BA, et al. Comparing outcomes and costs of medical patients treated at major teaching and non-teaching hospitals: a national matched analysis. J Gen Intern Med. 2020;35(3):743–752. doi:10.1007/s11606-019-05449-x

16. Vahedi S, Zahiri M, Pirani N, et al. Healthcare reform and productivity of hospital: a DEA-based analysis from South West of Iran. Cost Eff Resour Allocation. 2023;21(1):26. doi:10.1186/s12962-022-00403-x

17. Alshehri A, Balkhi B, Gleeson G, Atassi E. Efficiency and resource allocation in government hospitals in Saudi arabi: a casemix index approach. Healthcare. 2023;11(18):2513. doi:10.3390/healthcare11182513

18. Zhao N, Chen K. Equity and efficiency of medical and health service system in China. BMC Health Serv Res. 2023;23(1):33. doi:10.1186/s12913-023-09025-2

19. Ye Y, Tao Q. Measurement and characteristics of the temporal-spatial evolution of China’s healthcare services efficiency. Arch Public Health. 2023;81(1):197. doi:10.1186/s13690-023-01208-x

20. Jing R, Xu T, Lai X, Mahmoudi E, Fang H. Technical efficiency of public and private hospitals in Beijing, China: a comparative study. Int J Environ Res Public Health. 2020;17(1):82. doi:10.3390/ijerph17010082

21. Zhu J, Song X. Changes in efficiency of tertiary public general hospitals during the reform of public hospitals in Beijing, China. Int J Health Plann Manage. 2022;37(1):143–155. doi:10.1002/hpm.3309

22. Zeng Z, Yu X, Tao W, Feng W, Zhang W. Efficiency evaluation and promoter identification of primary health care system in China: an enhanced dea-tobit approach. BMC Health Serv Res. 2024;24(1):777. doi:10.1186/s12913-024-11244-0

23. Kohl S, Schoenfelder J, Fügener A, Brunner JO. The use of data envelopment analysis (dea) in healthcare with a focus on hospitals. Health Care Manag Sci. 2019;22(2):245–286. doi:10.1007/s10729-018-9436-8

24. Zhou M. Equity and efficiency of public health resource allocation in China. BMC Public Health. 2025;25(1):1–16. doi:10.1186/s12889-025-21447-7

25. Mei K, Kou R, Bi Y, Liu Y, Huang J, Li W. A study of primary health care service efficiency and its spatial correlation in China. BMC Health Serv Res. 2023;23(1):247. doi:10.1186/s12913-023-09197-x

26. Huang R, Li W, Shi B, et al. Evaluating China’s primary healthcare services’ efficiency and spatial correlation: a three-stage dea-malmquist model. Front Public Health. 2024;12:1366327. doi:10.3389/fpubh.2024.1366327

27. Ye Y, Evans R, Jing L, Rizwan M, Xuan Y, Lu W. The impact of county medical community reform on the medical service efficiency of county-level public general hospitals in China: a case study of Shanxi province. Int J Environ Res Public Health. 2022;19(21):13827. doi:10.3390/ijerph192113827

28. Wang J, Xie F, Guo Z. Research progress on the efficiency of health resource allocation in China. Health Econ Res. 2024;41(12):55–59.

29. Imani A, Alibabayee R, Golestani M, Dalal K. Key indicators affecting hospital efficiency: a systematic review. Front Public Health. 2022;10:830102. doi:10.3389/fpubh.2022.830102

30. Ma K, Hsu Y, Pan W, Chou M, Chung WS, Wang J. Effects of emergency department length of stay on inpatient utilization and mortality. Health Econ Rev. 2025;15(1):1–8. doi:10.1186/s13561-025-00598-8

31. Maradin D, Olgić Draženović B, Čegar S. The efficiency of offshore wind energy companies in the european countries: a dea approach. Energies. 2023;16(9):3709. doi:10.3390/en16093709

32. Andrejić M, Kilibarda M, Pajić V. Measuring efficiency change in time applying malmquist productivity index: a case of distribution centres in Serbia. Facta Universitatis Ser. 2021;19(3):499–514. doi:10.22190/FUME201224039A

33. Amore MD, Murtinu S. Tobit models in strategy research: critical issues and applications. Global Strategy J. 2021;11(3):331–355. doi:10.1002/gsj.1363

34. Liang B, Zhai S. A study of the allocation efficiency of medical and healthcare resources in the ethnic areas of China: spatial differences, dynamic evolution, and influencing factors. J Yunnan Minzu Univ. 2024;41(06):110–122.

35. Alatawi AD, Niessen LW, Bhardwaj M, et al. Factors influencing the efficiency of public hospitals in Saudi Arabia: a qualitative study exploring stakeholders’ perspectives and suggestions for improvement. Front Public Health. 2022;10:922597. doi:10.3389/fpubh.2022.922597

36. Zhong K, Chen L, Cheng S, Chen H, Long F. The efficiency of primary health care institutions in the counties of Hunan province, China: data from 2009 to 2017. Int J Environ Res Public Health. 2020;17(5):1781. doi:10.3390/ijerph17051781

37. Fu Y, Lin W, Yang Y, Du R, Gao D. Analysis of diverse factors influencing the health status as well as medical and health service utilization in the floating elderly of China. BMC Health Serv Res. 2021;21(1):438. doi:10.1186/s12913-021-06410-7

38. Nomani M, Hussain Z. Innovation technology in health care management in the context of Indian environmental planning and sustainable development. Int J Emerg Technol. 2020;11(2):560–564.

39. Alami H, Rivard L, Lehoux P, et al. Artificial intelligence in health care: laying the foundation for responsible, sustainable, and inclusive innovation in low-and middle-income countries. Globalization Health. 2020;16(1):1–6. doi:10.1186/s12992-020-00584-1

40. Yang Y, Zhu X, Song R, Zhang X, Guo F. Not just for the money? An examination of the motives behind physicians’ sharing of paid health information. J Inf Sci. 2023;49(1):145–163. doi:10.1177/0165551521991029

41. Ahmed S, Hasan MZ, Maclennan M, et al. Measuring the efficiency of health systems in Asia: a data envelopment analysis. BMJ open. 2019;9(3):e22155. doi:10.1136/bmjopen-2018-022155

42. Gavurova B, Kocisova K, Sopko J. Health system efficiency in OECD countries: dynamic network dea approach. Health Econ Rev. 2021;11(1):1–25. doi:10.1186/s13561-021-00337-9

43. Huang M, Luo D, Wang Z, et al. Equity and efficiency of maternal and child health resources allocation in Hunan province, China. BMC Health Serv Res. 2020;20(1):1–10. doi:10.1186/s12913-020-05185-7

44. Zhao H, Zheng Q. Evaluation on efficiency of resource allocation of provincial health care institutions in China from 2012 to 2021 – based on super-efficiency DEA-Malmquist model. Soft Sci Health. 2023;37(12):58–63.

45. de Oliveira LFA, Baião FA, Oliveira FLC, et al. Determinants of primary healthcare efficiency in the Brazilian Legal Amazon: a hybrid approach using data envelopment analysis and Tobit regression. Socio-Econ Plann Sci. 2025;101(1):102257. doi:10.1016/j.seps.2025.102257

46. Gcabashe S, Bayat MS, Kgobe FKL. Evaluating the effectiveness of resources allocated to rural healthcare facilities in South Africa. Social Sci Educ Res Rev. 2024;11(2):274–282.

47. Niedar A, Hafidz F, Hort K. Optimization of healthcare workers availability: increasing primary healthcare efficiency in Indonesia. Jurnal Ekonomi Kesehatan Indonesia. 2022;7(1):1. doi:10.7454/eki.v7i1.5397

48. Chen L, Chen T, Lan T, Chen C, Pan J. The contributions of population distribution, healthcare resourcing, and transportation infrastructure to spatial accessibility of health care. Inquiry. 2023;60:1438227527.

49. Yuan L, Cao J, Wang D, Yu D, Liu G, Qian Z. Regional disparities and influencing factors of high quality medical resources distribution in China. Int J Equity Health. 2023;22(1):8. doi:10.1186/s12939-023-01825-6

50. Zheng D, Gong J, Zhang C. Efficiency of medical service systems in the rural areas of mainland China: a comparative study from 2013 to 2017. Public Health. 2019;171(6):139–147. doi:10.1016/j.puhe.2019.04.003

51. Baporikar N. Role of information technology in enhancing healthcare services. Digital Technol Res Efficient Econ. 2024;5(2):63–87.

52. Wu Q, Xie X, Liu W, Wu Y. Implementation efficiency of the hierarchical diagnosis and treatment system in China: a case study of primary medical and health institutions in Fujian province. Int J Health Plann Manage. 2022;37(1):214–227. doi:10.1002/hpm.3333

53. El Husseiny IA. The efficiency of healthcare systems in the Arab countries: a two-stage data envelopment analysis approach. J Humanities Appl Soc Sci. 2023;5(4):339–358. doi:10.1108/JHASS-10-2021-0168

54. Cheng J, Kuang X, Zeng L. The impact of human resources for health on the health outcomes of Chinese people. BMC Health Serv Res. 2022;22(1):1213. doi:10.1186/s12913-022-08540-y

55. Chen Y, Wu J. The effect of the referral system on the accessibility of healthcare services: a case study of the Wuhan metropolitan development zone. Int J Environ Res Public Health. 2022;19(16):10441. doi:10.3390/ijerph191610441

56. Sankaran A. The future of smart healthcare: how ai and har are reshaping hospital workflows. Well Testing J. 2025;34(1):68–92.

57. Albarqouni L, Arab-Zozani M, Abukmail E, et al. Overdiagnosis and overuse of diagnostic and screening tests in low-income and middle-income countries: a scoping review. BMJ Global Health. 2022;7(10):e8696. doi:10.1136/bmjgh-2022-008696

58. Chen B, Jin F. Spatial distribution, regional differences, and dynamic evolution of the medical and health services supply in China. Front Public Health. 2022;10:1020402. doi:10.3389/fpubh.2022.1020402

Story Continues

Despite recent improvements in Pakistan’s macroeconomic indicators, the country continues to face deep-rooted structural challenges, with a persistently low domestic savings rate topping the list, cautioned State Bank of Pakistan (SBP) Governor Jameel Ahmad.

Speaking at a seminar in Karachi on Monday, the central bank chief emphasized the urgent need to address this issue to ensure sustainable economic growth.

Citing the latest Pakistan Economic Survey, Jameel Ahmad revealed that Pakistan’s domestic savings rate stands at a mere 7.4 percent of GDP, significantly lagging behind the South Asian average of 27 percent and East Asia’s 41 percent.

“This low savings rate forces Pakistan to rely heavily on foreign inflows to meet its development needs,” he said. “But this reliance has come at a cost, contributing to repeated balance of payment crises, instability in foreign exchange markets, and inflationary pressures, all of which have weakened our growth momentum over time.”

The SBP governor stressed that mobilizing domestic savings and channeling them into productive investments is critical to breaking the cycle of economic booms and busts.

Turning to Pakistan’s capital markets, Jameel Ahmad highlighted the progress made in the government bond market over the past two decades. The market now offers a variety of securities, including fixed-rate, floating-rate, and Sharia-compliant options with different maturities. However, he noted that the market remains heavily concentrated within the banking system, limiting its broader impact.

To address this, the SBP has introduced several reforms aimed at broadening access and deepening the bond market. These include:

• Revamping the primary dealer system.
• Expanding the scope of investor portfolios to include microfinance banks, the Central Depository Company (CDC), and the National Clearing Company of Pakistan Limited (NCCPL).
• Simplifying account opening through the Customer’s Digital Onboarding Framework, enabling nearly 100 million branchless and mobile banking users to invest in government securities.

These reforms aim to diversify the investor base, enhance liquidity, and make the sovereign segment more resilient,” he said.

The SBP governor also pointed out the glaring absence of a robust corporate debt market in Pakistan. “Outstanding corporate bonds account for less than 1% of our GDP, a stark contrast to other Asian economies,” he said.

He noted that non-financial firms, as well as key sectors like manufacturing, infrastructure, and renewable energy, remain almost entirely dependent on bank loans for financing. This over-reliance on banks, he argued, underscores the need to develop a vibrant corporate debt market to support long-term economic growth.

Jameel Ahmad concluded by reiterating the importance of addressing Pakistan’s structural challenges, particularly the low savings rate and underdeveloped capital markets. “These issues are not just economic; they are systemic. Without meaningful reforms, Pakistan will continue to face the same cycles of instability and missed opportunities,” he warned.

Inverex and River Indus have announced that the Inverex Xio electric vehicle (EV) is now open for pre-booking in Pakistan, with a minimum advance payment of Rs. 500,000. The announcement comes as part of a campaign for Pakistan’s Independence Day.

Booking and Availability

The Xio EV can be booked at two locations:

• Karachi Showroom: 12 Merry Mansion, Frere Road, Saddar
• Hyderabad Showroom: Plot 16-B, Main Wadhoo Wah Road, Qasimabad

For inquiries, customers can contact Inverex at the following numbers:

• 0301-8227028
• 0301-8220147
• 0301-2030097
• 0301-8227024

Inverex Xio Details

The Inverex Xio is a compact electric hatchback featuring a four-door layout, suitable for city driving. The model showcased during the Independence Day campaign includes a green and white wrap bearing the Pakistani flag. The EV appears to be positioned toward urban drivers seeking affordable electric mobility.

The Xio EV is compatible with standard EV chargers and is shown with a dedicated charging station in the promotional material.

Pricing and Installments

The full pricing and delivery schedule of the Inverex Xio EV has not been officially announced. However, early pre-bookings are being accepted for Rs 500,000, suggesting limited initial availability and a potential phased rollout.

Introduction

Neonatal Hospital-Acquired Gastrointestinal Infections (NHAGIs) are defined as gastrointestinal infections that develop in newborns more than 48 hours after hospital admission. These infections represent one of the most common and severe complications in neonatal intensive care unit (NICU), characterized by high incidence rates, significant mortality, and substantial treatment costs. Epidemiological studies indicate that NHAGIs account for 10–30% of all hospital-acquired infections in NICU, constituting a major contributor to neonatal mortality.¹ Regional studies in China have reported varying NHAGI incidence rates, with tertiary hospitals showing significantly lower rates compared to resource-limited settings, consistent with global disparities in neonatal care quality.² Globally, NHAGI burden varies widely, with higher rates in low-resource settings compared to high-income countries, underscoring the need for context-specific interventions.^3,4 The pathogenesis of NHAGIs is multifactorial, with key risk factors including preterm birth, low birth weight, invasive medical procedures, antibiotic usage and the inherent risk of cross-infection within the NICU environment.^5–7 Current predictive approaches predominantly depend on clinicians empirical judgments, which are limited by subjective interpretation, inadequate timeliness, and challenges in processing complex, high-dimensional data. These limitations significantly hinder effective early warning systems and precise clinical interventions.^8,9 Consequently, the development of an objective and reliable predictive tool has emerged as a critical priority in clinical research.

Table 1 Baseline Features of the Training and Test Set

Machine learning (ML) possess the capability to autonomously extract patterns from extensive medical datasets and develop predictive models, offering significant advantages including enhanced objectivity, improved timeliness and robust processing capabilities for high-dimensional data.^10,11 The development of predictive models for NHAGIs utilizing ML enables early warning systems, facilitates personalized treatment strategies, and optimizes resource allocation. These capabilities contribute to reducing infection rates and mortality while improving treatment outcomes and enhancing the efficiency of medical resource utilization.^12,13 Specifically, Fleuren et al systematically demonstrated the utility of ML models (eg, random forests, gradient boosting) in sepsis prediction, achieving pooled AUCs of 0.82–0.88 across 28 studies, while Nemati et al developed an interpretable LSTM-based model for real-time ICU sepsis prediction with 85% sensitivity, highlighting ML’s clinical translatability for infection-related outcomes.Recent research has demonstrated the application of various algorithms, including logistic regression, support vector machines, and random forests, in constructing predictive models for neonatal outcomes, with their efficacy being empirically validated.^14–16 These findings suggest that ML holds substantial promise for broad applications in the prediction and management of NHAGIs.

To address this research gap, we conducted a comprehensive investigation of multiple obstetric and NICU-related risk factors associated with NHAGIs. Utilizing feature selection techniques to identify significant risk factors, we developed and validated a machine learning-based predictive model for NHAGIs, demonstrating strong performance in predicting infection risk. This retrospective study was designed to elucidate critical obstetric and clinical diagnostic factors influencing NHAGIs occurrence. The findings from this research are expected to significantly contribute to clinical practice by providing valuable insights for early intervention strategies. Furthermore, this study offers substantial support for enhancing neonatal healthcare outcomes and promoting healthy infant development.

Methods

Study Population

The study population comprised 851 neonates admitted to the NICU of our hospital between January 2020 and December 2024. Among these, 176 neonates were diagnosed with NHAGIs during their NICU stay, while 675 neonates without NHAGIs diagnosis during the same period were randomly selected as controls. Patient inclusion criteria were established in accordance with the Diagnostic Guidelines for Nosocomial Infections issued by the National Health Commission of China.¹⁷ Figure 1 presents the flow chart illustrating the study population screening process.

Figure 1 Flowchart of participant screening for study inclusion.

Data Collection

We retrospectively collected 29 perinatal and neonatal diagnostic and treatment-related characteristic variables from electronic medical records and nursing documentation systems.Missing data were handled using complete-case analysis, whereby any patient record with missing values for one or more variables was excluded from the analysis. This conservative approach was chosen to ensure the integrity of the analytical dataset.While we initially considered additional clinical factors such as antibiotic regimens (eg, specific agents, duration), detailed nutritional parameters (eg, caloric intake, growth velocity), and microbiome profiling, these variables were excluded due to either: (1) inconsistent documentation in medical records (>30% missing data), or (2) lack of standardized measurement protocols across cases.The included variables were:gestational age, birth weight, initial laboratory indicators upon NICU admission (WBC, HGB, PLT, NE), maternal age, maternal BMI, preeclampsia status, delivery mode, parity, timing of first feeding, milk source, nasogastric feeding, use of food additives, probiotic administration, intrauterine distress, hypothermia, hypoglycemia, asphyxia, respiratory failure, respiratory distress syndrome, ventilator use, central venous puncture, congenital heart defects, premature rupture of membranes, intrauterine infection, gestational diabetes, and pregnancy-induced hypertension. These factors were comprehensively evaluated by a multidisciplinary team comprising experienced obstetricians, infectious disease specialists, and neonatologists, ensuring their clinical relevance and reference value.

Statistical Analysis

A comprehensive descriptive analysis was performed to characterize the study population. All statistical analyses were conducted using R-studio software. Continuous variables were expressed as mean ± standard deviation (SD) and analyzed using t-test, while categorical variables were expressed as percentages and analyzed using the chi-square test. Both univariate and multivariate logistic regression analyses were employed to calculate odds ratio (OR) with corresponding 95% confidence interval (CI), with statistical significance set at p < 0.05.The dataset was randomly partitioned into training and testing sets at a 7:3 ratio. Feature selection was performed in the training set using three distinct methods: Boruta, Lasso and logistic regression. The consensus features identified through these methods were determined using Venn diagram analysis, followed by correlation heatmap visualization to assess inter-variable relationships. Subsequently, eight machine learning algorithms were implemented in the training set, including Gradient Boosting Machine (GBM), Adaptive Boosting (Adaboost), Extreme Gradient Boosting (Xgboost), Neural Network, Random Forest (RF), Logistic Regression (LR), Support Vector Machine (SVM), and K-Nearest Neighbors (KNN).For machine learning implementation, we employed a standardized training framework with 10-fold repeated (5 times) cross-validation (method=“repeatedcv”, number=10, repeats=5) and two-class summary metrics (classProbs=TRUE, summaryFunction=twoClassSummary). All models used the same random seed (set.seed(520)) for reproducibility. The eight algorithms were implemented with the following optimized hyperparameters: Logistic Regression: glm package with ridge regularization (alpha=0, lambda=0.01); SVM: svmRadial with sigma=0.01 and C=1 (optimized via grid search); GBM: n.trees=500, interaction.depth=3, shrinkage=0.01, n.minobsinnode=10; Neural Network: size=10, decay=0.01; Random Forest: mtry=sqrt(n_features), numRandomCuts=1; XGBoost: nrounds=500, max_depth=6, eta=0.01, gamma=0.1, colsample_bytree=0.8; KNN: kmax=20, distance=2; AdaBoost: mfinal=50, maxdepth=4.Model performance was evaluated and compared using multiple metrics: AUC-ROC, sensitivity, specificity, F1 score, accuracyandprecision. The testing set was utilized for model validation and performance assessment. Based on its superior predictive performance, the neural network model was selected for the development of an online risk calculator using the Shiny framework. The calculator offers a user-friendly interface where clinicians can input variables and instantly receive NHAGIs risk predictions by clicking the “Predict” button.

Results

Patient’s Characteristics and Logistic Regression Analyses

Table 1 presents the baseline characteristics of the training and validation cohorts. The patient data were randomly divided into training and validation sets at a 7:3 ratio. The training set (n = 946) underwent SMOTEprocessing, with Figure 2A illustrating the distribution of sample categories before and after SMOTE application. The validation set (n = 253) remained unprocessed for comparative analysis.Statistical analysis revealed significant differences (p < 0.05) in 12 characteristic variables between the cohorts: gestational age, birth weight, NHAGIs status, delivery mode, parity, milk source, nasogastric feeding, intrauterine distress, hypothermia, respiratory failure, respiratory distress syndrome, and ventilator use.Table 2 displays the results of logistic regression analyses performed on the training set. The analysis identified several statistically significant predictive factors (p < 0.05), including gestational age, PLT, NE, delivery mode, milk source, nasogastric feeding, probiotic administration, intrauterine distress, central venous catheterizationand pregnancy-induced hypertension.

Table 2 Univariate and Multivariate Logistic Regression Analysis of the Training Set

Figure 2 (A) The distribution of the class labels of the samples before and after applying SMOTE to the training set. (B) The result of Boruta algorithm screening important features. (C and D) The result of Lasso regression screening important features.

Broguta and Lasso Regression Analysis

The Boruta algorithm is an advanced feature selection method that effectively identifies significant predictors for modeling. In the Boruta visualization (Figure 2B), black box plots represent the minimum, average, and maximum Z-scores of shadow features, while green box plots denote confirmed features. The results demonstrate that all variables marked in green were identified as important features. Figure 2C presents the LASSO regression path diagram, displaying 29 distinct trajectories corresponding to the included variables. Each colored line represents the coefficient trajectory of an independent variable, with the y-axis indicating coefficient values and the lower x-axis showing log(λ) values. The upper x-axis displays the number of non-zero coefficients at each λ value. As log(λ) increases, the regression coefficients gradually converge toward zero. Figure 2D illustrates the cross-validation curve, where the x-axis represents log(λ) and the y-axis shows the binomial deviance. Through this process, LASSO regression identified 12 relevant variables, which were subsequently incorporated into the multivariate logistic regression analysis, as detailed in Table 3.

Table 3 The Variables Selected by Multifactor Logistic Regression After the Lasso Regression

Venn Diagram and Spearman Correlation Heatmap

Through comparative analysis of feature selection results from Boruta, Lasso and logistic regression, we identified the intersecting subset of features common to all three methods as our final predictors. Figure 3A illustrates the nine feature variables ultimately selected for inclusion in the ML prediction model: gestational age, PLT, NE, delivery mode, nasogastric feeding, probiotic administration, intrauterine distress, central venous catheterization, and pregnancy-induced hypertension. Figure 3B presents the Spearman correlation coefficient matrix heatmap, which was employed to assess inter-variable relationships.

Figure 3 (A) Venn diagram of the three feature selection methods. (B) The Spearman correlation coefficient matrix heatmap. (C and D) ROC curves for eight models in the training cohort and validation cohort.

Performance Comparison of Eight ML Algorithms

In the training set (Figure 3C), RF and KNN demonstrated perfect performance metrics (Accuracy = 1.000, Sensitivity = 1.000, Specificity = 1.000, Precision = 1.000, F1 = 1.000, AUC = 1.000), suggesting potential overfitting. GBM showed excellent performance with an AUC of 0.972 and an F1 score of 0.930, along with balanced metrics. XGBoost and Neural Network exhibited stable performance with AUCs of 0.905 and 0.892. SVM and AdaBoost demonstrated moderate performance, achieving AUCs of 0.882 and 0.864, while LR showed comparable performance with an AUC of 0.895.In the validation set (Figure 3D), LR and Neural Network emerged as the top performers, achieving AUCs of 0.877 and 0.876, with balanced sensitivity and specificity. XGBoost attained an AUC of 0.790, demonstrating the highest sensitivity (0.942) but lower precision and F1, potentially indicating overfitting. SVM maintained stable performance with an AUC of 0.875, closely matching the top performers. GBM and RF showed reduced performance in the validation set, with AUCs of 0.829 and 0.817 respectively, likely due to overfitting. KNN exhibited the poorest generalization capability, achieving only an AUC of 0.671 and sensitivity of 0.500. AdaBoost demonstrated moderate performance with an AUC of 0.789, though its F1 score remained relatively low at 0.584.In summary, while the RF and KNN models demonstrated perfect performance on the training set, their significant performance degradation on the validation set indicated severe overfitting.The calibration curves for both training and validation sets revealed that the Neural Network model’s predictions were closer to the ideal diagonal compared to other algorithms, indicating better alignment between predicted probabilities and actual outcomes (Figure 4A and B). DCA demonstrated that the RF model maintained high clinical net benefit across both datasets (Figure 4C and D). The classification performance of all ten algorithms was further evaluated through confusion matrix analysis (Figure 4E and F).Furthermore, ten-fold cross-validation performed on the training and validation sets demonstrated the Neural Network model’s robust performance, with Figure 5A and B illustrating the AUC values for each fold and the average AUC. Similarly, GBM exhibited reduced AUC and F1 scores on the validation set, likely attributable to their high model complexity. In contrast, the Neural Network model maintained consistent performance across both training and validation sets, achieving high AUC and F1 scores, which supported its selection as the final predictive model.Notably, all models showed suboptimal F1 scores on the validation set, potentially due to the limited number of positive cases. This limitation was addressed through SMOTE oversampling of the validation set, which resulted in significant improvement of F1 scores (Figure 6).Based on these comprehensive evaluations, the Neural Network was ultimately selected as the optimal prediction model.

Figure 4 (A and B) Calibration curves for eight models in the training cohort and validation cohort. (C and D) DCA curves for eight models in the training cohort and validation cohort. (E and F) Confusion matrix results for eight models in the training cohort and validation cohort.

Figure 5 (A) Ten-fold cross-validation of the training cohort. (B) Ten-fold cross-validation of the validation cohort.

Figure 6 The performance metrics of the eight machine learning models across the training, validation, and SMOTE-processed validation sets.

Characteristic Importance and Interpretability of Model

Figure 7A presents the feature importance scores of the Neural Network model, visualized using SHapley Additive exPlanations (SHAP) values. The marginal contributions of the nine selected features are illustrated across all samples in Figure 7B. The analysis reveals that gestational age, PLT, NE, delivery mode, and nasogastric feeding demonstrate dispersed sample distributions and wider SHAP value ranges, indicating their substantial impact on model predictions. Conversely, intrauterine distress, central venous catheterization, and pregnancy-induced hypertension show distributions concentrated near SHAP = 0, suggesting minimal influence on the model’s output.The dependency plot (Figure 7C) demonstrates that platelet count (particularly within the range of 300–400 × 10⁹/L) may significantly contribute to prediction outcomes at specific gestational ages (eg, 36 weeks). Additionally, a force plot was generated to provide detailed feature-level explanations for individual sample predictions, as shown in Figure 7D.

Figure 7 (A) Ranking of the importance scores for the eight variables. (B) Swarm map based on SHAP interpretation. (C) Dependence plot between the characteristic variables gestation week and PLT. (D) Single-sample interpretable force diagram.

Shiny Application of the Model

We developed an interactive risk calculator using Shiny in the R programming language to facilitate clinical implementation of the NHAGIs prediction model. This web-based application, accessible at [http://sh15609631795.shinyapps.io/EnteritisPredictor/], provides a user-friendly interface for clinical decision support. Figure 8 demonstrates the application interface, where clinicians can input relevant patient data and obtain the probability of NHAGIs by clicking the “Predict” button.

Figure 8 User interface for a risk calculator developed using shiny.

Discussion

Neonatal hospital-acquired gastrointestinal infections (NHAGIs) result from the complex interplay of multiple risk factors. In this study, we developed and validated eight ML algorithms to predict NHAGIs using comprehensive data encompassing various diagnostic and clinical factors. Among these models, the Neural Network demonstrated consistently superior performance across both training and testing datasets. This model’s effectiveness was further supported by DCA and calibration curve results, indicating strong clinical applicability.The clinical significance of our ML approach lies in its ability to identify critical risk factors associated with NHAGIs. Our analysis identified nine key predictive factors: gestational age, NE, PLT, central venous catheterization, nasogastric feeding, delivery mode, intrauterine distress, pregnancy-induced hypertension and probiotic administration.The clinical significance of our ML approach is underscored by its superior performance compared to existing non-ML predictive tools. Traditional clinical scoring systems like the Modified Neonatal Sepsis Score (MNSS) demonstrate limited predictive accuracy due to their reliance on 6–8 predefined variables through logistic regression.^18,19 In contrast, our neural network model achieved significantly higher discrimination while incorporating 29 input features, capturing complex interactions missed by conventional approaches. Notably, our model identified several novel predictors absent in current scores,²⁰ and overcame the classic accuracy-interpretability trade-off through SHAP analysis. The development of our Shiny application further addresses a critical implementation gap, enabling real-time risk assessment unavailable in manual scoring systems.²¹ These advancements suggest ML can overcome key limitations of rule-based tools while maintaining clinical interpretability.

Based on the permutation importance analysis of feature variables in the neural network model, gestational age, nasogastric feeding, and delivery mode emerged as crucial predictors of neonatal nosocomial gastrointestinal infections. Gestational age serves as a fundamental indicator of neonatal health, with prematurity being strongly associated with increased infection risk. Preterm infants often exhibit immature immune systems and compromised intestinal barrier function, rendering them more susceptible to gastrointestinal infections.²² Previous research has established a correlation between nasogastric feeding and neonatal gastrointestinal infections. The use of nasogastric tubes may facilitate pathogen entry into the gastrointestinal tract, particularly with prolonged usage.^23,24 Furthermore, cesarean delivery has been associated with elevated infection risk compared to vaginal delivery, potentially due to its impact on the establishment of neonatal gut microbiota, thereby increasing vulnerability to gastrointestinal infections.^25,26 Hematological indicators, NE and PLT counts, also play essential roles in predicting NHAGIs. Neutrophils, as critical components of the immune system, exhibit count variations that reflect infection or inflammatory status. The degree of neutrophil elevation in infected neonates specifically correlates with infection severity and type.²⁷ Similarly, platelet counts serve as indicators of immune status and inflammatory response. Notably, thrombocytopenia frequently occurs in neonates, particularly during infections or inflammatory states, potentially exacerbating infection risk.²⁸ Our findings also indicate that probiotic supplementation in breast milk or formula may inhibit the development of gastrointestinal infections in newborns by modulating intestinal microbiota. Previous research has demonstrated that probiotics can strengthen intestinal barrier function and suppress pathogenic growth.²⁹ Similarly, central venous catheterization, like nasogastric tube placement, represents a significant risk factor for nosocomial infections in neonates, as these devices can serve as entry points for pathogens.³⁰ Intrauterine distress has been shown to potentially compromise neonatal immune function through hypoxia, increasing infection susceptibility. The resultant intrauterine hypoxia may also impair intestinal function, further elevating the risk of gastrointestinal infections.³¹ Notably, our analysis revealed an unexpected association between pregnancy-induced hypertension (PIH) and reduced NHAGIs risk, which may be attributed to several factors:First, PIH (including preeclampsia) often necessitates medically indicated preterm delivery, resulting in more intensive medical supervision and prophylactic antibiotic use for preterm infants, potentially lowering gastrointestinal infection risk.³² The stringent infection control measures implemented in NICU, including aseptic techniques and antibiotic prophylaxis, may further contribute to this protective effect.³³ Second, PIH may influence neonatal immunity through placental transmission of immune-modulating factors, potentially enhancing immune defense mechanisms.³⁴ Third, the increased frequency of prenatal monitoring and interventions in PIH cases may indirectly reduce neonatal infection risk by enabling earlier detection and management of intrauterine infections or other complications.³⁵ Finally, mothers with PIH may be more likely to initiate breastfeeding, which has been consistently associated with reduced NHAGIs risk. Breast milk contains immunoglobulins and beneficial microorganisms that support the development of neonatal intestinal barrier function.³⁶

This study developed and evaluated eight ML algorithms to create a predictive model for NHAGIs using comprehensive hospital data. We systematically analyzed the relationship between NHAGIs and various obstetric and clinical parameters, while employing advanced techniques to interpret the model’s decision-making process and address potential issues of model interpretability. However, several limitations should be acknowledged. First, the model’s generalizability may be limited as it was developed using data from a single center. Second, despite achieving an accuracy exceeding 85%, the model requires validation through prospective studies to further establish its clinical utility and practical applicability.Third, while critical factors like antibiotic regimens and microbiome data were considered, they were excluded from the final model due to either inconsistent documentation in medical records (>30% missing data) or lack of standardized measurement protocols, which may represent potential omitted variable bias that should be addressed in future studies with more comprehensive data collection protocols.

Conclusion

In conclusion, our Neural Network model effectively predicts neonatal hospital-acquired gastrointestinal infections, identifying critical risk factors and enabling early intervention via a Shiny-based tool. This advances neonatal care by supporting timely clinical decisions. Prospective, multi-center validation is needed to confirm generalizability and optimize clinical integration, addressing limitations and enhancing the model’s impact on reducing NHAGI-related morbidity.

Abbreviations

NHAGIs,neonatalhospital-acquired gastrointestinal infections;ML,machinelearning;SHAP,shapleyadditive explanations; NICU, neonatal intensive care unit;SMOTE,syntheticminorityover-sampling technique; AUC, Area under the curve; CI, confidence interval; WBC, white blood cell; HCG, Hemoglobin;PLT, platelet; NE, neutrophil;BMI,bodymassindex;PROM, Premature rupture of membranes.

Data Sharing Statement

Relevant data from this study can be obtained from the corresponding author.

Ethics Statement

This retrospective study was conducted in accordance with the Declaration of Helsinki and was approved by the Ethics Committee of Shaoxing Maternal and Child Health Hospital (IRB-AF-022-01.5). The research involved analysis of anonymized medical records and did not include human participants or animal trials. The ethics committee waived the requirement for informed consent given the retrospective nature of the study. All data were handled in compliance with institutional guidelines and regulations for patient data confidentiality.

Author Contributions

All authors made a significant contribution to the work reported, whether that is in the conception, study design, execution, acquisition of data, analysis and interpretation, or in all these areas; took part in drafting, revising or critically reviewing the article; gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work.

Funding

There is no supporting funding.

Disclosure

No conflict of interest is declared.

References

1. Stoll BJ, Hansen NI, Bell EF, et al. Trends in care practices, morbidity, and mortality of extremely preterm neonates, 1993–2012. JAMA. 2015;314(10):1039–1051. PMID: 26348753. doi:10.1001/jama.2015.10244

2. Zaidi AK, Huskins WC, Thave D, et al. Hospital-acquired neonatal infections in developing countries. Lancet. 2005;365(9465):1175–1188. PMID: 15794973. doi:10.1016/S0140-6736(05)71881-X

3. Allegranzi B, Bagheri Nejad S, Combescure C, et al. Burden of endemic health-care-associated infection in developing countries: systematic review and meta-analysis. Lancet. 2011;377(9761):228–241. PMID: 21146207. doi:10.1016/S0140-6736(10)61458-4

4. Rosenthal VD, Al-Abdely HM, El-Kholy AA, et al. International nosocomial infection control consortium report, data summary of 50 countries for 2010–2015: device-associated module. AM J Infect Control. 2016;44(12):1495–1504. PMID: 27742143. doi:10.1016/j.ajic.2016.08.007

5. Boghossian NS, Page GP, Bell EF, et al. Late-onset sepsis in very low birth weight infants from singleton and multiple-gestation births. J Pediatr-US. 2013;162(6):1120–1124. PMID: 23324523. doi:10.1016/j.jpeds.2012.11.089

6. Patel RM, Knezevic A, Shenvi N, et al. Association of red blood cell transfusion, anemia, and necrotizing enterocolitis in very low-birth-weight infants. JAMA. 2016;315(9):889–897. PMID: 26934258. doi:10.1001/jama.2016.1204

7. Cotten CM, Taylor S, Stoll B, et al. Prolonged duration of initial empirical antibiotic treatment is associated with increased rates of necrotizing enterocolitis and death for extremely low birth weight infants. Pediatrics. 2009;123(1):58–66. PMID: 19117. doi:10.1542/peds.2007-3423

8. Saria S, Butte A, Sheikh A. Better medicine through machine learning: what’s real, and what’s artificial? PLOS Med. 2018;15(12):e1002721. PMID: 30596635. doi:10.1371/journal.pmed.1002721

9. Miotto R, Wang F, Wang S, Jiang X, Dudley JT. Deep learning for healthcare: review, opportunities and challenges. Brief Bioinform. 2018;19(6):1236–1246. PMID: 28481991. doi:10.1093/bib/bbx044

10. Esteva A, Robicquet A, Ramsundar B, et al. A guide to deep learning in healthcare. Nat Med. 2019;25(1):24–29. PMID: 30617335. doi:10.1038/s41591-018-0316-z

11. Topol EJ. High-performance medicine: the convergence of human and artificial intelligence. Nat Med. 2019;25(1):44–56. PMID: 30617339. doi:10.1038/s41591-018-0300-7

12. Fleuren LM, Klausch TLT, Zwager CL, et al. Machine learning for the prediction of sepsis: a systematic review and meta-analysis of diagnostic test accuracy. Intens Care Med. 2020;46(3):383–400. PMID: 31965266. doi:10.1007/s00134-019-05872-y

13. Nemati S, Holder A, Razmi F, et al. An interpretable machine learning model for accurate prediction of sepsis in the ICU. Crit Care Med. 2018;46(4):547–553. PMID: 29286945. doi:10.1097/CCM.0000000000002936

14. Xu H, Peng X, Peng Z, et al. Construction and SHAP interpretability analysis of a risk prediction model for feeding intolerance in preterm newborns based on machine learning. BMC Med Inform DecisMak. 2024;24(1):342. PMID: 39558307. doi:10.1186/s12911-024-02751-5

15. Tan X, Zhang X, Chai J, et al. Constructing a predictive model for early-onset sepsis in neonatal intensive care unit newborns based on SHapley Additive exPlanations explainable machine learning. TranslPediatr. 2024;13(11):1933–1946. PMID: 39649648. doi:10.21037/tp-24-278

16. Wang D, Huang S, Cao J, et al. A comprehensive study on machine learning models combining with oversampling for bronchopulmonary dysplasia-associated pulmonary hypertension in very preterm infants. Respir Res. 2024;25(1):199. PMID: 38720331. doi:10.1186/s12931-024-02797-z

17. Chinese Society of Surgical Infection and Intensive Care, Chinese Society of Surgery, Chinese Medical Association; Chinese College of Gastrointestinal Fistula Surgeons, Chinese College of Surgeons, Chinese Medical Doctor Association. [Chinese guideline for the prevention of surgical site infection]. Zhonghua Wei Chang Wai Ke Za Zhi. 2019;22(4):301–314 . doi:10.3760/cma.j.issn.1671-0274.2019.04.001

18. Escobar GJ, Puopolo KM, Wi S, et al. Stratification of risk of early-onset sepsis in newborns ≥ 34 weeks’ gestation. Pediatrics. 2014;133(1):30–36. PMID: 24366992. doi:10.1542/peds.2013-1689

19. Wynn JL, Wong HR, Shanley TP, et al. Time for a neonatal-specific consensus definition for sepsis. Pediatr Crit Care Me. 2014;15(6):523–528. PMID: 24751791. doi:10.1097/PCC.0000000000000157

20. Zea-Vera A, Ochoa TJ. Challenges in the diagnosis and management of neonatal sepsis. J Trop Pediatrics. 2015;61(1):1–13. PMID: 25604489. doi:10.1093/tropej/fmu079

21. Wiens J, Saria S, Sendak M, et al. Do no harm: a roadmap for responsible machine learning for health care. Nat Med. 2019;25(9):1337–1340. PMID: 31427808. doi:10.1038/s41591-019-0548-6

22. Chawanpaiboon S, Vogel JP, Moller AB, et al. Global, regional, and national estimates of levels of preterm birth in 2014: a systematic review and modelling analysis. Lancet Glob Health. 2019;7(1):e37–e46. PMID: 30389451. doi:10.1016/S2214-109X(18)30451-0

23. Sitomorang I, Garna H. Risk factors for nosocomial infections in neonatal intensive care unit. Paediatr Indones. 2018;34(1–2):48–56. doi:10.14238/pi34.1-2.1994.48-56

24. Auriti C, Maccallini A, Di Liso G, Di Ciommo V, Ronchetti MP, Orzalesi M. Risk factors for nosocomial infections in a neonatal intensive-care unit. J Hosp Infect. 2003;53(1):25–30. PMID: 12495682. doi:10.1053/jhin.2002.1341

25. Shao Y, Forster SC, Tsaliki E, et al. Stunted microbiota and opportunistic pathogen colonization in caesarean-section birth. Nature. 2019;574(7776):117–121. PMID: 31534227. doi:10.1038/s41586-019-1560-1

26. Neu J, Rushing J. Cesarean versus vaginal delivery: long-term infant outcomes and the hygiene hypothesis. Clin Perinatol. 2011;38(2):321–331. PMID: 21645799. doi:10.1016/j.clp.2011.03.008

27. Li Z, Yuan T. Neutrophil extracellular traps in adult diseases and neonatal bacterial infectious diseases: a review. Heliyon. 2024;10(1):e23559. PMID: 38173520. doi:10.1016/j.heliyon.2023.e23559

28. Zhou E, Wang K, Gu Y. Research Progress in Neonatal Alloimmune Thrombocytopenia: a Narrative Review. Altern Ther Health Med. 2023;29(6):77–81. PMID: 37318890.

29. Dermyshi E, Wang Y, Yan C, et al. The “Golden Age” of probiotics: a systematic review and meta-analysis of randomized and observational studies in preterm infants. Neonatology. 2017;112(1):9–23. PMID: 28196365. doi:10.1159/000454668

30. O’Grady NP, Alexander M, Dellinger EP, et al. Guidelines for the prevention of intravascular catheter-related infections. The Hospital Infection Control Practices Advisory Committee, Center for Disease Control and Prevention, U.S. Pediatrics. 2002;110(5):e51. PMID: 12415057. doi:10.1542/peds.110.5.e51

31. Sharma D, Shastri S, Sharma P. Intrauterine growth restriction: antenatal and postnatal aspects. Clin Med Insights Pediatr. 2016;10:67–83. PMID: 27441006. doi:10.4137/CMPed.S40070

32. Steer P. The epidemiology of preterm labour. BJOG Int J Obstet Gynaecol. 2005;112(Suppl 1):1–3. PMID: 15715585. doi:10.1111/j.1471-0528.2005.00575.x

33. Goldenberg RL, Culhane JF, Iams JD, Romero R. Epidemiology and causes of preterm birth. Lancet. 2008;371(9606):75–84. PMID: 18177778. doi:10.1016/S0140-6736(08)60074-4

34. Redman CW, Sargent IL. Immunology of pre-eclampsia. Am J Reprod Immunol. 2010;63(6):534–543. PMID: 20331588. doi:10.1111/j.1600-0897.2010.00831.x

35. Gestational Hypertension and Preeclampsia: ACOG Practice Bulletin Summary, Number 222. Obstet Gynecol. 2020;135(6):1492–1495. PMID: 32443077. doi:10.1097/AOG.0000000000003892

36. Kramer MS, Kakuma R. Optimal duration of exclusive breastfeeding. Cochrane Database Syst Rev. 2012;8:CD003517. PMID: 22895934. doi:10.1002/14651858.CD003517.pub2

Russia’s main stock index fell more than 2% on Monday as Ukrainian President Volodymyr Zelensky and European allies prepared for talks in Washington aimed at ending the war.

The Moscow Exchange index, which tracks about 40 of Russia’s largest companies, dropped to 2,945 points at the open. The RTS index slipped to 1,160.

Shares of energy giants Gazprom, Rosneft and Tatneft led the decline, along with shipping firm Sovcomflot and Bank St. Petersburg, all losing more than 3%.

Meanwhile, the ruble inched higher, rising 0.07% to 80.2 against the U.S. dollar, 0.09% to 93.83 against the euro and 0.10% to 11.17 against the Chinese yuan, according to Reuters data. Analysts predict the currency will likely remain volatile ahead of the Trump-Zelensky meeting later Monday.

The Russian market is also expected to face “heightened volatility” in the coming days amid uncertainty over ongoing talks on ending the war in Ukraine, said Yaroslav Kabakov, strategy director at the investment firm Finam.

If negotiations succeed, “active market participants are likely to start pricing in the prospect of future trilateral talks,” said BKS analyst Mikhail Zeltser, who predicted the MOEX could climb above 3,000.

U.S. President Donald Trump and Russia’s Vladimir Putin held a high-stakes summit in Alaska on Friday, though they failed to reach the landmark Ukraine peace deal that the American leader had been hoping for.

Introduction

UC is a chronic, idiopathic inflammatory bowel disease (IBD) characterized by persistent mucosal inflammation. It affects millions of individuals worldwide, with its prevalence steadily increasing in recent years.¹ The development of UC is multifactorial, involving genetic susceptibility, impaired epithelial barrier integrity, dysregulated immune responses, and environmental triggers.² Among these factors, the essential trace element Se has attracted growing attention due to its antioxidant³ and anti-inflammatory properties, as well as its potential therapeutic value in UC.^4,5

Selenium exerts its biological effects primarily through selenoproteins—proteins that incorporate the amino acid Sec.⁶ The human genome encodes 25 selenoproteins,⁷ many of which, such as glutathione peroxidases (GPx), thioredoxin reductases (TrxR), and iodothyronine deiodinases (DIO), serve as intracellular antioxidants with well-established oxidoreductase functions.⁸ Recent studies have shown that selenoproteins play a critical role in the onset and progression of UC. In knockout mouse models, both GPx-1 and GPx-2 have demonstrated protective roles against intestinal inflammation; their simultaneous deletion results in ileocolitis.⁹ Although GPx-3 deletion alone does not induce colitis, it significantly exacerbates disease severity in dextran sulfate sodium (DSS)-treated GPx-3⁻/⁻ mice.¹⁰ Additionally, Selenoprotein S (SELENOS) expression increases in vivo alongside the endoplasmic reticulum (ER) stress marker GRP78 following DSS treatment.¹¹ In patients with UC, Selenoprotein P (SELENOP) is downregulated in colon biopsy samples, with its expression inversely correlated with disease severity as measured by endoscopy.¹² Specific deletion of Selenoprotein I (SELENOI) in intestinal epithelial cells triggers ferroptosis, impairs intestinal regeneration, and reduces colonic tumor growth.¹³ Furthermore, Selenoprotein W (SELENOW) is essential for resolving experimental colitis by regulating the epidermal growth factor receptor (EGFR) and Yes-associated protein 1 (YAP1) signaling pathways.¹⁴

In this study, we investigated the role of SeMet-related genes in UC using integrative bioinformatics and experimental approaches. Gene expression and clinical data were obtained from the GEO database, and SeMet-related gene sets were collected from MSigDB. WGCNA and differential expression analysis were performed to identify key modules and DRGs. Based on eleven upregulated SeMet-related genes, 161 UC patients were stratified into two molecular subtypes. Machine learning algorithms were applied to identify six candidate signature genes with high diagnostic potential, which were then used to construct a UC risk prediction model. Single-sample gene set enrichment analysis (ssGSEA) revealed strong correlations between signature genes and immune cell infiltration. ScRNA-seq analysis showed upregulation of several selenoproteins in epithelial cells and downregulation of SELENOP in immune cells. We further validated the elevated expression of selenoprotein M (SELENOM) and selenoprotein N (SELENON) in UC tissues and demonstrated that WARS1 responds to oxidative stress, with its knockdown increasing inflammatory cytokine levels. These findings enhance our understanding of SeMet-related genes in UC and highlight their diagnostic and therapeutic relevance.

Materials and Methods

Sources and Processing of Datasets

Eight UC datasets were obtained from the Gene Expression Omnibus (GEO, https://www.ncbi.nlm.nih.gov/geo/) database, including GSE75214,¹⁵ GSE87466,¹⁶ GSE47908,¹⁷ GSE206171, GSE48958,¹⁸ GSE9452,¹⁹ GSE38713,²⁰ and GSE13367.²¹ Next, the GSE75214 and GSE87473 datasets were combined, and batch effects in the gene expression data were addressed using the “comBat” function from the “sva” package in R. The datasets GSE47908, GSE206171, and GSE48958 were used as test sets for machine learning. The gene set “Selenium Metabolism and Selenoproteins” was obtained from the Molecular Signatures Database v5.0 (http://software.broadinstitute.org/gsea/msigdb/index.jsp).

Immune Cell Infiltration Analysis

ssGSEA²² was employed to estimate the relative infiltration levels of immune cell types in the UC microenvironment. Gene sets specific to various immune cell types were obtained from the study by Charoentong et al.²³ The analysis was conducted using the R package GSVA, which calculates enrichment scores for individual gene sets across samples based on transcriptome data.

Defining SeMet-Associated Molecular Subtypes in Ulcerative Colitis

Unsupervised hierarchical clustering analysis was then conducted on 161 UC samples using the “ConsensusClusterPlus” R package, based on the significantly upregulated DRGs.²⁴ Molecular pathways with significant enrichment were identified based on gene set variation analysis (GSVA) scores across distinct subtypes. Immune cell infiltration variations between these subtypes were then analyzed.

Gene Co-Expression Network Construction

To identify co-expressed gene modules and investigate their relationships with various traits or phenotypes, we employed WGCNA analysis.^25,26 Initially, the necessity for filtering gene samples was assessed using the “goodSamplesGenes” function from the “WGCNA” package (v1.69). Next, we constructed an adjacency matrix by calculating Pearson’s correlation coefficients between all gene pairs. This matrix was then used to build a scale-free co-expression network, applying a soft-thresholding technique that amplifies strong gene correlations and suppresses weaker ones. Finally, the adjacency matrix was transformed into a topological overlap matrix (TOM), which quantitatively reflects the similarity between node pairs based on their weighted correlations.

Identification of Feature Genes and Construction of a Risk Model Based on Machine Learning Algorithms

In order to identify key genes with potential diagnostic value, we applied two advanced feature selection methods: the Least Absolute Shrinkage and Selection Operator (LASSO) and Support Vector Machine-Recursive Feature Elimination (SVM-RFE). Genes identified by both algorithms were considered as potential candidates for further investigation. Following this, we developed a UC risk prediction model by integrating 15 different machine learning algorithms and evaluating 207 distinct combinations of these models. The algorithms incorporated included: neural networks, logistic regression, linear and quadratic discriminant analysis, K-nearest neighbors (KNN), decision trees, random forests, XGBoost, ridge regression, LASSO regression, elastic net regression, support vector machines, gradient boosting machines, stepwise logistic regression, and naive Bayes.^27,28

Gene Set Enrichment Analysis (GSEA) and Correlation Analysis

To explore the functional relevance of specific genes in the context of UC, we analyzed their expression profiles in relation to other mRNAs through Pearson correlation using transcriptomic datasets from the GEO database. To gain insights into the potential biological pathways involved, Gene Set Enrichment Analysis (GSEA) was carried out utilizing the R package “clusterProfiler”.²⁹ For this study, the “c2.cp.kegg.v7.5.1.entrez.gmt” collection was used in GSEA.

Single-Cell Analysis

To further investigate the signature genes and DRGs at the single-cell level in UC, the GSE214695³⁰ and GSE231993³¹ datasets were downloaded from the GEO database for single-cell sequencing analysis. High-throughput sequencing data were analyzed using the “Seurat” package in R to construct single-cell expression profiles. To reduce dimensionality and identify clusters based on highly variable genes, principal component analysis (PCA) was initially performed. For further nonlinear dimensionality reduction and visualization, the UMAP algorithm was applied. To correct for batch effects between control and UC samples, the “Harmony” R package was utilized. Cell type annotation was performed using the “SingleR” R package in combination with manual curation based on published literature, ensuring both computational accuracy and biological relevance. Marker genes for each cell population were identified using the FindAllMarkers function. The expression patterns and distribution of both signature genes and DRGs were examined across individual cell clusters. To explore intercellular communication patterns within the UC microenvironment, the “CellChat” R package was employed. Finally, the Monocle2 package was employed to perform pseudotime analysis on macrophages.

Cell Culture

The HCT116 human colon cancer cell line and NCM460 normal human colonic epithelial cell line were obtained from the American Type Culture Collection. The cells were cultured in complete DMEM medium (Thermo Scientific, Waltham), supplemented with 10% fetal bovine serum (FBS, Gibco, Thermo Scientific, Waltham), in a CO2 incubator set at 5% CO2 and 37°C.

Western Blotting

The cells were lysed with ice-cold RIPA lysis buffer (Servicebio, China) containing protease inhibitors and then centrifuged at 4°C (12,000 rpm for 20 min). The protein supernatant was then quantified using a BCA protein assay kit (Biyuntian, China). After protein denaturation, 30 μg of protein were separated on 10% gels by SDS-PAGE and transferred to a PVDF membrane (Millipore, USA). After blocking with 5% skim milk in TBS-T, the membrane was incubated overnight at 4°C with the following antibodies: anti-SELENOM (SANTA, 1:400); anti-SELENON (SANTA, 1:400); anti-GPX1/2 (SANTA, 1:400); anti-WARS1 (Proteintech, 1:1000); anti-GAPDH (Proteintech, 60004-1-Ig). The membranes were then incubated with goat anti-rabbit (Proteintech, RGAR001, 1:5000) or mouse IgG secondary antibodies (Proteintech, RGAM001, 1:5000) for 1 hour. The membranes were then washed with TBS-T three times (5 min per wash) and finally visualized using enhanced chemiluminescence substrate.

RNA Transcription and Real-Time PCR

Total RNA was extracted from tissue and cell samples of UC using the RNAiso kit (Vazyme RC201). mRNA was then reverse transcribed into complementary DNA (cDNA) using the cDNA synthesis kit (Vazyme R222-01). RT-qPCR was subsequently performed using the SYBR Green Master Mix reagent (Vazyme Q131). Relative expression levels were determined using the 2−∆∆Ct method, with GAPDH used as an internal control.

Histology and Immunohistochemistry

Colon tissue samples were preserved in 4% paraformaldehyde, embedded in paraffin, and sectioned into 5 μm slices for histological analysis. For immunofluorescence staining, antigen retrieval was performed by heating the sections in a sodium citrate buffer solution. To suppress endogenous peroxidase activity, tissue slices were treated with 3% hydrogen peroxide. Subsequently, non-specific binding sites were blocked using 3% bovine serum albumin (BSA) for 30 minutes at ambient temperature. Primary antibodies were applied and incubated overnight at 4 °C. The next day, after applying the corresponding secondary antibodies, the sections were developed with DAB, counterstained with hematoxylin, and examined under a standard brightfield microscope.

Animals and Animal Models

All animal experiments in this study were conducted in accordance with the welfare and ethical guidelines for experimental animals established by Zhejiang University and approved by the Animal Experimental Ethics Committee. Wild-type C57BL/6 mice (6 weeks old) were acquired from GemPharmatech (Jiangsu, China) and maintained in a specific-pathogen-free (SPF) animal facility. The mouse colitis model was induced by administering 3% dextran sulfate sodium (DSS, Yeasen Biotech, Shanghai, China) dissolved in filter-purified and sterilized water ad libitum to the Wild-type C57BL/6 mice for 7 days.

Sample Collection

Five patients diagnosed with UC and five healthy controls (HC) were recruited from the Fourth Affiliated Hospital of Zhejiang University. Intestinal mucosal biopsies were obtained during endoscopic examination. All participants provided written informed consent prior to anesthesia. The study was approved by the Medical Ethics Committee of the Fourth Affiliated Hospital of Zhejiang University and conducted in accordance with the Declaration of Helsinki.

Results

Variations of SeMet-Related Genes in Active Ulcerative Colitis

To investigate alterations in SeMet-related genes associated with active UC, datasets GSE75214 and GSE87466 were merged, and batch effects were removed. Differentially regulated genes (DRGs) were then identified using the “limma” package in R. As shown in Figure 1A, 27 DRGs displayed distinct expression patterns between UC and healthy controls. Specifically, genes such as FOS, SELENOM, SELENON, DIO2, GPX1, GPX2, SELENOK, RELA, SELENOS, CREM, and TRNAU1AP were significantly upregulated, while SELENBP1, SELENOW, GPX4, TXNRD2, and others were notably downregulated. Genes with |log₂FC| > 0.5 and adjusted P-value < 0.05 were selected for downstream analysis. Figure 1B presents the correlation matrix among the 11 most significantly dysregulated DRGs. To explore immune dysregulation in UC and the involvement of DRGs, we applied ssGSEA to quantify the relative abundance of various immune cell types in both UC and control samples. As shown in Figure 1C, immune cell populations such as activated dendritic cells, CD8⁺ T cells, B cells, natural killer (NK) cells, macrophages, and neutrophils were significantly enriched in UC tissues. Correlation analysis between DRGs and immune infiltration revealed distinct association patterns (Figure 1D). SELENOM and SELENON were positively correlated with central memory CD4⁺ T cells, effector memory CD8⁺ T cells, and NK cells, whereas SEPSECS, SEPHS2, and SELENBP1 showed inverse correlations with neutrophils and macrophages. These findings suggest that DRGs may influence disease progression in UC through modulation of immune cell infiltration and activity.

Figure 1 Variations of SeMet-Related Genes in Active Ulcerative Colitis. (A) Scatter plots showing differential expression of SeMet-related genes between UC and control groups. (B) Pie chart size reflects correlation coefficients among genes. (C) Violin plots comparing immune cell infiltration between UC and controls. (D) Correlation analysis of 11 differentially expressed DRGs and immune cell types. (E) Consensus clustering heatmap for k = 2. (F) Heatmap of differential expression of 11 DRGs across identified subtypes. (G) Boxplots displaying subtype-specific expression of 11 DRGs. (H) Violin plots illustrating immune cell infiltration differences between subtypes. (I) GSVA results comparing pathway enrichment between cluster 1 and cluster 2. Statistical significance is indicated as follows: ns, not significant; *P < 0.05; **P < 0.01; ***P < 0.001; ****P < 0.0001.

Consensus clustering analysis of 161 active UC samples was performed based on the expression profiles of eleven upregulated DRGs. The optimal number of clusters was determined to be k = 2, supported by multiple evaluation metrics, including the structure of the consensus matrix (Figure 1E), the shape of the cumulative distribution function (CDF) curve (Figure S1A), and the inflection point in the CDF area change plot (Figure S1B). Consequently, two molecular subtypes associated with selenoproteins and selenium metabolism were identified and designated as Cluster 1 (n = 73) and Cluster 2 (n = 88). The heatmap and boxplot (Figure 1F and G) revealed significantly higher expression levels of SELENOK, SELENOM, SELENOS, CREM, and DIO2 in Cluster 1 compared to Cluster 2. Immune profiling demonstrated that activated CD4⁺ T cells, activated dendritic cells, effector memory CD8⁺ T cells, neutrophils, and macrophages were significantly enriched in Cluster 1 relative to Cluster 2 (Figure 1H). Gene set variation analysis (GSVA) further indicated that Cluster 1 was positively enriched in multiple immune-related pathways, including natural killer cell-mediated cytotoxicity, JAK-STAT signaling, toll-like receptor signaling, and NOD-like receptor signaling, whereas it was negatively enriched in selenocysteine metabolism and the PPAR signaling pathway (Figure 1I). These findings suggest that patients in Cluster 1 are likely in a stage characterized by pronounced inflammatory responses and impaired selenoprotein function. Given the classification of selenoproteins into housekeeping and stress-related types,³² we performed differential expression analysis comparing selenoprotein expression between inactive and active colitis patients using the GSE75214 dataset. This analysis identified significant upregulation of SELENOM, SELENOS, and DIO2 in active colitis (Figure S1C), supporting their involvement during periods of inflammatory activity.

Verification of Signature Genes and Construction of a Prediction Model

To identify genes strongly associated with clinical traits, we selected the top 25% of genes showing the highest variability between UC and healthy control samples for WGCNA. Outlier samples were removed, and a soft-thresholding power of β = 18 was selected based on the scale-free topology criterion (cutoff = 0.8; Figure S1D). Using the dynamic tree cut algorithm, nine distinct gene modules were identified, each represented by a unique color. Correlation analysis between module eigengenes and clinical traits revealed that the pink module (containing 202 genes) exhibited the strongest association with UC status (Figures 2A and S1E). A separate WGCNA was conducted based on the two UC subtypes identified previously. With a soft-thresholding power of β = 17, eight distinct modules were detected (Figure S1F). Correlation analysis with subtype features indicated that the brown module showed the strongest positive correlation with Cluster 1, comprising 216 genes (Figures 2B and S1G). By intersecting the gene sets from the pink and brown modules, 19 overlapping candidate signature genes were identified for further analysis (Figure 2C). To reduce dimensionality and eliminate redundancy, least absolute shrinkage and selection operator (LASSO) regression with tenfold cross-validation was employed, resulting in the selection of 10 genes (Figure 2D and E). Subsequently, a support vector machine (SVM) model was applied to the overlapping gene set for further screening, and the root mean square error (RMSE) reached its minimum when six genes were included (Figure 2F and G). Combining the results from both approaches, six potential core genes were identified: WARS1, Kynureninase (KYNU), Chitinase 3 Like 1 (CHI3L1), Plasminogen Activator, Urokinase (PLAU), Granzyme B (GZMB), and Caspase-4 (CASP4) (Figure 2H). To construct a predictive model for UC, 15 machine learning algorithms were utilized and evaluated using the merged dataset as well as three external validation cohorts (GSE47908, GSE48958, and GSE206171). The top 50 models, ranked by area under the curve (AUC), are presented in Figure 2I. Gene feature importance across the different models was calculated (Figure S2A), and a comprehensive integration of multiple models yielded the final ranking of predictive genes (Figure 2J).

Figure 2 Verification of Signature Genes and Construction of a Prediction Model. (A) Heatmap displaying correlations between module eigengenes and UC status; red and blue indicate positive and negative correlations, respectively. (B) Correlation heatmap between module eigengenes and molecular subtypes. (C) Shared genes identified between disease-related and subtype-associated modules. (D and E) Feature selection using the LASSO regression model. (F and G) Key gene identification through SVM-RFE analysis. (H) Venn diagram showing common candidate genes identified by both machine learning approaches. (I) Integrated heatmap presenting the top 50 AUC values across datasets for each algorithm. (J) Bar plot ranking the importance scores of the six selected feature genes.

Functional Analysis of Signature Genes in Active UC

Subsequent analyses were performed to validate the expression and diagnostic potential of the six identified signature genes. In independent test sets, all six genes—WARS1, KYNU, CHI3L1, PLAU, GZMB, and CASP4—were significantly upregulated in UC samples compared to healthy controls (Figure 3A), with AUC values exceeding 0.80, indicating robust diagnostic performance (Figure 3B). Similarly, in the external validation set, the six genes remained significantly upregulated (Figure S2B), and their respective AUC values also demonstrated high diagnostic accuracy (Figure S2C). To elucidate the potential biological functions of these genes in active UC, Gene Set Enrichment Analysis (GSEA) was conducted. All six signature genes were positively enriched in multiple immune-related signaling pathways, including cytokine–cytokine receptor interaction, B cell receptor signaling, T cell receptor signaling, Toll-like receptor signaling, and the JAK-STAT pathway (Figure 3C). In contrast, these genes were negatively enriched in pathways associated with neurodegenerative disorders, such as Alzheimer’s disease, Huntington’s disease, and Parkinson’s disease (Figure 3C). Given the enrichment of immune-related pathways, we further investigated the association between the six signature genes and the infiltration of 28 immune cell types. A strong positive correlation was observed between the expression of these genes and various immune cell populations (Figure 3D), suggesting their involvement in immune activation. Furthermore, correlation analysis among the six signature genes revealed potential co-regulatory relationships (Figure 3E). Collectively, these results suggest that the signature genes may contribute to UC pathogenesis by modulating immune cell responses, potentially through synergistic mechanisms.

Figure 3 Functional Analysis of Signature Genes in Active UC. (A) Expression profiles of WARS1, KYNU, CHI3L1, PLAU, GZMB, and CASP4 in the training set. (B) ROC curves evaluating the diagnostic performance of signature genes in the training set. (C) GSEA results illustrating pathway enrichment of individual signature genes. (D) Correlation matrix showing associations between signature genes and 28 immune cell types. (E) Inter-gene correlation among the six signature genes. *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001.

Single-Cell Transcriptomic Analysis of Selenoprotein-Related Genes in Ulcerative Colitis

Single-cell RNA sequencing analysis was performed on six UC tissue samples from the GSE214695 dataset using the Seurat R package. Through clustering and dimensionality reduction using principal component analysis (PCA) and uniform manifold approximation and projection (UMAP), 13 distinct cell types were identified and annotated: B cells, CD4⁺ T cells, CD8⁺ T cells, colonocytes, endothelial cells, fibroblasts, glial cells, goblet cells, macrophages, mast cells, neutrophils, plasma cells, and tuft cells (Figure 4A). Given that SELENOP is a key selenoprotein responsible for transporting approximately 60% of total serum selenium, further analysis focused on selenoproteins that were elevated in bulk transcriptome data, including SELENOP. The expression profiles of selenoprotein genes across the 13 identified cell types are shown in Figure 4B. Differential expression analysis between UC and healthy control samples was conducted using the FindMarkers function in Seurat. Results revealed significant upregulation of multiple selenoprotein genes in epithelial cell populations, particularly in colonocytes and goblet cells (Figure 4C).

Figure 4 Single-Cell Transcriptomic Analysis of Selenoprotein-Related Genes in Ulcerative Colitis. (A) UMAP analysis showing identified cell types. (B) Feature plots showing selenoprotein genes expression in the 13 cell types. (C) Heat map showing the expression disparity of selenoprotein genes between UC and control groups. (D) UMAP of macrophage subtypes: M0, M2, IDA, and M1. (E) Monocle 2 pseudotime analysis reveals macrophage differentiation trajectories and transcriptional dynamics across subtypes. (F) Pseudotime-dependent expression of selenoproteins (SELENOP, SELENOK) across macrophage polarization states. (G) CellChat analysis reveals that intestinal epithelial cells regulate macrophage function through the LGALS9–HAVCR2 signaling axis. (H) Differential SELENOP expression in epithelial cells with varying LGALS9 expression levels. “×” denotes genes with no detectable expression in the corresponding cell type. Statistical significance is indicated as follows: ns, not significant; *P < 0.05; **P < 0.01; ***P < 0.001.

In contrast, SELENOP expression was significantly downregulated in several immune cell populations, including macrophages, CD4⁺ T cells, CD8⁺ T cells, and neutrophils (Figure 4C). Given the critical role of macrophage polarization in modulating inflammatory responses in UC, we further explored the functional roles of key selenoproteins—specifically SELENOP, SELENOK, SELENOM, SELENOS, and GPX2—within macrophage populations. Based on the classification by Alba et al³⁰ we annotated macrophages into four distinct subtypes: M0 macrophages, M2 macrophages, inflammation-dependent alternative (IDA) macrophages, and M1 macrophages (Figure 4D). Recent studies suggest that resident macrophages typically fall into M0 or M2 subsets, whereas inflammatory macrophages are classified as either M1 or IDA types.³⁰ To better understand the transcriptional relationships among macrophage populations, we applied the unsupervised inference method Monocle 2 to construct potential transitional trajectories (Figures 4E and S3A) and generated faceted pseudotime plots illustrating the distribution of cells from each patient (Figure S3B). Previous studies have demonstrated that SELENOK plays a critical role in the migration of immune cells, including T cells, neutrophils, macrophages, and dendritic cells (DCs).^33–36 By contrast, macrophages lacking SELENOP show markedly reduced migratory capacity.¹² Moreover, impaired SELENOP function correlates with enhanced M2 polarization.³⁷ Interestingly, SELENOP may suppress pro-inflammatory immune polarization, thereby mitigating inflammation-driven tumorigenesis.³⁸ This suggests a complex regulatory role in macrophage behavior and immune homeostasis. Additionally, SELENOS deficiency in macrophages promotes M1 polarization by targeting Ubiquitin A-52 Ribonucleoprotein (Uba52) to inhibit YAP ubiquitination and degradation.³⁹ In our analysis, the differentiation trajectory revealed notable changes in SELENOK and SELENOP expression (Figure 4F). After classifying macrophages into resident and inflammatory subtypes, we observed that inflammatory macrophages from UC patients significantly upregulated SELENOK and SELENOS, while downregulating SELENOP (Figure S3C). Collectively, these findings suggest that multiple selenoproteins cooperatively modulate macrophage polarization and function, playing pivotal roles in UC-associated inflammation and shaping the immune microenvironment.

To investigate the potential crosstalk between immune cells and epithelial cells, we conducted further analysis using CellChat. The results revealed that intestinal epithelial cells may regulate macrophage function through the LGALS9–HAVCR2 signaling axis (Figure 4G). Notably, epithelial cells with higher SELENOP expression also exhibited elevated LGALS9 levels, suggesting that the activity of this signaling pathway may be more pronounced in these cells (Figure 4H).

We applied the same algorithm to analyze the expression of signature genes at the single-cell level (Figure S3D–F). Among them, WARS1, CASP4, KYNU, and PLAU were broadly upregulated across most cell types. GZMB was mainly upregulated in immune cells, while CHI3L1 showed the highest expression in fibroblasts. Notably, inflammatory macrophages from UC patients displayed significant upregulation of WARS1, CASP4, and KYNU.

To further validate these results, we analyzed 12 additional samples from the GSE231993 dataset, which included UC and normal tissue samples from eight patients. After dimensionality reduction and clustering, we used the SingleR package to identify nine distinct cell types (Figure S4A). Subsequent refinement of monocyte clustering revealed four subsets: resident macrophages, infiltrating macrophages, DCs, and a minor proliferating macrophage population (Figure S4B). We then assessed expression differences and distribution patterns of selenoprotein genes and signature genes across these nine cell types (Figure S4C–F). Consistent with earlier findings, most selenoproteins were significantly upregulated in epithelial cells, while SELENOP showed marked downregulation in most non-epithelial cell types. Signature genes were also broadly elevated across nearly all cell types. Finally, differential analysis between resident and inflammatory macrophages yielded results similar to those previously described (Figure S4G). Further studies are required to elucidate the biological functions underlying these differences.

Validation of Differential Expression of Selenoproteins in Ulcerative Colitis

To investigate the differential expression of selenoprotein genes and six signature genes, we established a murine colitis model induced by DSS. After induction, we collected colonic tissues from both control and DSS-treated mice for molecular and histological analyses. Additionally, colonic tissue samples from five UC patients and five healthy individuals were obtained for further validation. qPCR revealed significant upregulation of several key selenoprotein genes, including DIO2, GPX2, SELENOM, SELENON, and SELENOS, in the DSS group compared to controls (Figure 5A). Among these, GPX2, SELENOM, and SELENON were further validated at the protein level by IHC (Figures 5B and S5A–B) and Western blotting (Figure 5C), both of which consistently confirmed their elevated expression in inflamed colonic tissues. Given the established roles of selenoproteins in maintaining redox homeostasis and mitigating oxidative stress, we next examined their inducibility under oxidative stress in vitro. Cells were treated with hydrogen peroxide (H₂O₂) to mimic oxidative damage, and gene expression dynamics were monitored over time. Notably, protein levels of GPX1/2, SELENOM, and SELENON increased in a time-dependent manner following H₂O₂ stimulation, suggesting their involvement in cellular antioxidant defense mechanisms (Figure 5D).

Figure 5 Validation of Differential Expression of Selenoproteins in Ulcerative Colitis. (A) The result of Real-time PCR analysis illustrated the expression levels of several key selenoprotein genes. (B) Immunohistochemistry experiment validating the expression of GPX1/2, SELENOM and SELENON in the colon of WT mice with UC. (C) GPX1/2, SELENOM and SELENON protein levels in the colon of WT mice with UC. (D) Western blot analysis of the dynamic expression of GPX1/2, SELENOM and SELENON in HCT116 cells and NCM460 cells during H2O2 (200μM) stimulation. Statistical significance is indicated as follows: ns, not significant; *P < 0.05; **P < 0.01.

WARS1 as a Potential Therapeutic Target in Ulcerative Colitis

The analysis of signature genes validated the bioinformatic predictions, revealing significant upregulation in the DSS-treated group (Figure 6A). Given that WARS1 ranked highest in gene importance across multiple machine learning algorithms, we performed further investigation on this gene. Intracellularly, WARS1 primarily ligates tryptophan (Trp) to its corresponding tRNA for protein synthesis. Extracellularly, it also serves as an innate immune activator.⁴⁰ Previous studies have reported that IFN-γ-mediated secretion of WARS1 from mesenchymal stem cells (MSCs) contributes to the suppression of excessive inflammation and the progression of IBD,⁴¹ highlighting its immunomodulatory potential. Recent studies have shown that WARS1 downregulation affects cytosolic translation and mitochondrial protein synthesis, activating the mitochondrial unfolded protein response (UPRmt), which is an important pathway in cellular stress and inflammation.⁴² Overall, these findings suggest that WARS1 may play a significant role in UC.

Figure 6 WARS1 as a Potential Therapeutic Target in Ulcerative Colitis. (A) The result of Real-time PCR analysis illustrated the expression levels of signature genes. (B) Expression levels of WARS1 mRNA in multiple GEO datasets. (C) Immunohistochemical staining images of WARS1 expression in inflamed intestinal tissues from UC patients and normal tissues from control subjects. (D) WARS1 protein levels in the colon of WT mice with UC. (E) Western blot analysis of dynamic expression of WARS1 in HCT116 and NCM460 cells during H2O2 (200μM) stimulation. (F) Western blot confirming WARS1 knockdown in HCT116 and NCM460 cells. (G) qPCR showing changes in inflammatory factors in WARS1 knockdown and control HCT116 and NCM460 cells. Statistical significance is indicated as follows: ns, not significant; *P < 0.05; **P < 0.01; ***P < 0.001; ****P < 0.0001.

To further explore the function of WARS1, we analyzed gene expression profiles from colonic mucosal samples of UC patients and healthy controls using four publicly available datasets (GSE75214, GSE13367, GSE38713, and GSE9452) from the Gene Expression Omnibus (GEO). WARS1 expression was significantly higher in active UC patients compared to both healthy controls and those with inactive UC (Figure 6B). We further confirmed the elevated WARS1 protein levels in the DSS-treated group by immunohistochemistry (Figure 6C and S5C) and Western blotting (Figure 6D). Notably, WARS1 expression progressively increased in cells exposed to H₂O₂, suggesting a role in oxidative stress responses (Figure 6E). In addition, we observed a strong positive correlation between WARS1 and inflammatory cytokines including TNF-α, IL-1β, and IL-6 in active UC patient samples (Figure S5D), implying its involvement in the inflammatory cascade. To validate this, we knocked down WARS1 in HCT116 and NCM460 cell lines (Figure 6F), which resulted in a marked upregulation of TNF-α, IL-1β, and IL-6 (Figure 6G). Taken together, these findings indicate that WARS1 may function as a regulator of inflammation and oxidative stress in UC and could serve as a promising therapeutic target. However, further studies are needed to elucidate its precise role and underlying mechanisms.

Discussion

The incidence of UC has increased globally in recent years. Its pathogenesis involves genetic predisposition, epithelial barrier disruption, immune dysregulation, and environmental factors.² Current treatments aim to induce remission and improve quality of life, but often cause side effects and fail in some patients.^43,44 Thus, novel therapies that resolve inflammation and promote intestinal healing are urgently needed.

Selenium and selenoproteins, due to their antioxidant and anti-inflammatory properties, are thought to protect against UC. The exacerbation of experimental colitis in selenium or selenoprotein deficiency highlights selenium’s role in regulating inflammatory pathways and oxidative stress in the gut.¹⁰ Although several studies have reported on the roles of various selenoproteins in UC, a comprehensive analysis of the selenoprotein family in relation to UC remains lacking. Therefore, for the first time, we aimed to analyze the potential association of selenoproteins with UC from a bioinformatics perspective.

Intestinal barrier dysfunction and immune imbalance play central roles in UC onset and progression. Therefore, protecting intestinal epithelial cells and modulating immune nd inflammatory responses represent two key therapeutic strategies. Our analysis indicates that the selenoprotein family may contribute to both processes. We identified six selenoproteins upregulated in active UC patients compared to controls. Single-cell data showed that most of these genes had significantly increased expression in epithelial cells. Conversely, SELENOP expression decreased markedly in most immune cells. SELENOP primarily functions as a selenium transporter and also acts as an extracellular antioxidant.⁴⁵ It is mainly produced by the liver and secreted into plasma, where it delivers selenium to tissues through its ten selenocysteine residues.⁴⁶ Within tissues, cells uptake and degrade SELENOP to release selenocysteines for the synthesis of other selenoproteins. Plasma SELENOP levels, together with selenium and GPx measurements, commonly assess selenium status.⁴⁷ Studies show that impaired SELENOP function promotes polarization toward M2 macrophages.³⁷ SELENOP deficiency compromises macrophage migration and disrupts intracellular selenoprotein balance.¹² The intestinal macrophage pool includes resident and infiltrating populations.⁴⁸ Resident macrophages generally suppress inflammation during UC,⁴⁹ whereas infiltrating monocytes differentiate into pro-inflammatory effectors.⁵⁰ Our results reveal a significant reduction of SELENOP in inflammatory macrophages relative to resident macrophages, possibly due to its consumption during immune activation. The broad decline of SELENOP across immune cells suggests it acts as a critical regulator during UC-associated immune responses.

Although we consistently observed SELENOP downregulation at the transcriptomic level in macrophages, its regulatory mechanism remains unclear. Whether this decrease results from transcriptional repression, defective uptake, or increased protein degradation requires further investigation. Future research using transcriptional reporter assays, mRNA stability studies, and protein turnover experiments will clarify how SELENOP expression is controlled in macrophages and how this regulation affects intestinal immune homeostasis. Moreover, this study offers preliminary evidence of immune–epithelial crosstalk via the LGALS9–HAVCR2 signaling axis, especially in epithelial cells expressing high levels of SELENOP. While these findings provide novel mechanistic insights, additional functional experiments are necessary to confirm the biological significance of this pathway and elucidate the role of selenoproteins in intercellular communication during UC.

In this study, WARS1, CHI3L1, GZMB, KYNU, PLAU, and CASP4 were identified as critical biomarkers for the diagnosis of UC. Previous studies reported downregulation of WARS1 in a DSS-induced experimental colitis model.⁵¹ Furthermore, it has been demonstrated that WARS1 inhibits the proliferation of CD4+ T cells derived from hUCB-MSCs by promoting apoptosis.⁵¹ However, recent research suggests that secreted WARS1, an endogenous ligand for Toll-like receptor (TLR) 2 and TLR4 involved in infection response, is a key activator of genes characteristic of a hyperinflammatory sepsis phenotype.⁵² Our study shows that WARS1 is significantly upregulated in active UC and gradually increases with prolonged oxidative stress. Knockdown of WARS1 leads to an increase in the expression of inflammatory factors, suggesting that it is more likely to act as a regulator in response to inflammation and oxidative stress. CHI3L1, a glycoprotein implicated in various diseases, including IBD, exerts its influence on multiple components of both the innate and adaptive immune responses.⁵³ GZMB, a serine protease extensively studied for its role in cytotoxic lymphocyte-mediated apoptosis, has been identified as a promising biomarker for detecting active IBD and predicting treatment response.⁵⁴ KYNU is a hydrolase involved in tryptophan metabolism. The silencing of KYNU has been shown to suppress inflammatory responses in intestinal epithelial cells under IL-1β stimulation.⁵⁵ PLAU encodes the urokinase-type plasminogen activator (uPA), which converts plasminogen to plasmin, a potent protease involved in fibrinolysis and extracellular matrix (ECM) degradation. A recent study suggests that local nutrient deprivation may serve as a candidate mechanism for PLAU upregulation in intestinal fibroblasts, a process that could be further amplified by IBD risk factors.⁵⁶ The CASP4 gene encodes a protein that plays a role in immunity and inflammatory processes.⁵⁷ Human caspases-4 and −5 have been shown to be elevated in the stromal tissue of patients with UC, and their expression levels are correlated with disease activity and inflammation scores.⁵⁸ In conclusion, the six signature genes identified in this study are closely linked to the pathogenesis of UC and could be potential targets for early diagnosis and treatment of the disease.

To summarize, our study highlights the potential contribution of SeMet-related genes and selenoproteins to UC pathogenesis, proposing new molecular targets for intervention. Nevertheless, our findings are primarily based on bioinformatics and in vitro analyses, with limited in vivo validation. Importantly, our in vivo data are restricted to the DSS-induced colitis model, which—despite its reproducibility—mainly reflects epithelial injury and innate immune responses. Given the multifactorial nature of UC and the involvement of adaptive immunity, future studies should incorporate complementary models, such as the T cell transfer colitis model, to validate the universality of selenoprotein alterations and further elucidate their immunoregulatory roles.

Conclusion

In summary, this study identified 11 SeMet-related genes that were significantly upregulated in UC patients, revealing their association with immune cells. Additionally, six signature genes (WARS1, KYNU, GZMB, CHI3L1, PLAU, and CASP4) were identified, and based on these genes, a highly accurate predictive model was developed. Among these, WARS1 was notably upregulated in response to oxidative stress, and its knockdown resulted in elevated levels of inflammatory cytokines, underscoring its critical role in the pathogenesis of UC. Single-cell RNA sequencing demonstrated that selenoproteins were predominantly expressed in epithelial cells and may protect epithelial cells from oxidative stress. These findings provide new insights into the early diagnosis and potential therapeutic targets for UC.

Abbreviations

AUC, Area Under the Curve; CASP4, Caspase-4; CDF, Cumulative Distribution Function; CHI3L1, Chitinase 3 Like 1; DCs, Dendritic Cells; DIO, Iodothyronine Deiodinases; DRGs, Differentially Regulated Genes; DSS, Dextran Sulfate Sodium; ECM, Extracellular Matrix; EGFR, Epidermal Growth Factor Receptor; ER, Endoplasmic Reticulum; GPx, Glutathione Peroxidases; GSEA, Gene Set Enrichment Analysis; GSVA, Gene Set Variation Analysis; GZMB, Granzyme B; H₂O₂, Hydrogen Peroxide; IBD, Inflammatory Bowel Disease; IDA, Inflammation-Dependent Alternative; IHC, Immunohistochemistry; LASSO, Least Absolute Shrinkage and Selection Operator; MSCs, Mesenchymal Stem Cells; MSigDB, Molecular Signatures Database; PCA, Principal Component Analysis; PLAU, Plasminogen Activator, Urokinase; qPCR, Quantitative Polymerase Chain Reaction; RMSE, Root Mean Square Error; scRNA-seq, Single-Cell RNA Sequencing; Se, Selenium; Sec, Selenocysteine; SeMet, Selenium metabolism and selenoproteins; ssGSEA, Single-Sample Gene Set Enrichment Analysis; SVM, Support Vector Machine; TLR, Toll-like Receptor; Trp, Tryptophan; TrxR, Thioredoxin Reductases; Uba52, Ubiquitin A-52 Ribonucleoprotein; UC, Ulcerative Colitis; UMAP, Uniform Manifold Approximation and Projection; UPRmt, Mitochondrial Unfolded Protein Response; uPA, Urokinase-type Plasminogen Activator; WARS1, Tryptophanyl-tRNA Synthetase 1; WGCNA, Weighted Gene Co-expression Network Analysis; YAP1, Yes-associated Protein 1.

Author Contributions

All authors made a significant contribution to the work reported, whether that is in the conception, study design, execution, acquisition of data, analysis and interpretation, or in all these areas; took part in drafting, revising or critically reviewing the article; gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work.

Funding

The work was supported by the Zhejiang Provincial Natural Science Foundation (LHDMZ23H160003).

Disclosure

The authors declare that they have no competing interests.

References

1. Vegh Z, Kurti Z, Lakatos PL. Epidemiology of inflammatory bowel diseases from west to east. J Dig Dis. 2017;18(2):92–98. doi:10.1111/1751-2980.12449

2. Ungaro R, Mehandru S, Allen PB, Peyrin-Biroulet L, Colombel JF. Ulcerative colitis. Lancet. 2017;389(10080):1756–1770. doi:10.1016/S0140-6736(16)32126-2

3. Papp LV, Lu J, Holmgren A, Khanna KK. From selenium to selenoproteins: synthesis, identity, and their role in human health. Antioxid Redox Signal. 2007;9(7):775–806. doi:10.1089/ars.2007.1528

4. Barrett CW, Singh K, Motley AK, et al. Dietary selenium deficiency exacerbates DSS-induced epithelial injury and AOM/DSS-induced tumorigenesis. PLoS One. 2013;8(7):e67845. doi:10.1371/journal.pone.0067845

5. Krehl S, Loewinger M, Florian S, et al. Glutathione peroxidase-2 and selenium decreased inflammation and tumors in a mouse model of inflammation-associated carcinogenesis whereas sulforaphane effects differed with selenium supply. Carcinogenesis. 2012;33(3):620–628. doi:10.1093/carcin/bgr288

6. Labunskyy VM, Hatfield DL, Gladyshev VN. Selenoproteins: molecular pathways and physiological roles. Physiol Rev. 2014;94(3):739–777. doi:10.1152/physrev.00039.2013

7. Gladyshev VN, Arner ES, Berry MJ, et al. Selenoprotein gene nomenclature. J Biol Chem. 2016;291(46):24036–24040. doi:10.1074/jbc.M116.756155

8. Tinggi U. Selenium: its role as antioxidant in human health. Environ Health Prev Med. 2008;13(2):102–108. doi:10.1007/s12199-007-0019-4

9. Esworthy RS, Aranda R, Martin MG, Doroshow JH, Binder SW, Chu FF. Mice with combined disruption of Gpx1 and Gpx2 genes have colitis. Am J Physiol Gastrointest Liver Physiol. 2001;281(3):G848–855. doi:10.1152/ajpgi.2001.281.3.G848

10. Speckmann B, Steinbrenner H. Selenium and selenoproteins in inflammatory bowel diseases and experimental colitis. Inflamm Bowel Dis. 2014;20(6):1110–1119. doi:10.1097/MIB.0000000000000020

11. Speckmann B, Gerloff K, Simms L, et al. Selenoprotein S is a marker but not a regulator of endoplasmic reticulum stress in intestinal epithelial cells. Free Radic Biol Med. 2014;67:265–277. doi:10.1016/j.freeradbiomed.2013.11.001

12. Short SP, Pilat JM, Barrett CW, et al. Colonic epithelial-derived selenoprotein P is the source for antioxidant-mediated protection in colitis-associated cancer. Gastroenterology. 2021;160(5):1694–1708e1693. doi:10.1053/j.gastro.2020.12.059

13. Huang X, Yang X, Zhang M, et al. SELENOI functions as a key modulator of ferroptosis pathway in colitis and colorectal cancer. Adv Sci. 2024;11(28):e2404073. doi:10.1002/advs.202404073

14. Nettleford SK, Liao C, Short SP, Rossi RM, Singh V, Prabhu KS. Selenoprotein W ameliorates experimental colitis and promotes intestinal epithelial repair. Antioxidants. 2023;12(4). doi:10.3390/antiox12040850

15. Vancamelbeke M, Vanuytsel T, Farre R, et al. Genetic and transcriptomic bases of intestinal epithelial barrier dysfunction in inflammatory bowel disease. Inflamm Bowel Dis. 2017;23(10):1718–1729. doi:10.1097/MIB.0000000000001246

16. Li K, Strauss R, Ouahed J, et al. Molecular comparison of adult and pediatric ulcerative colitis indicates broad similarity of molecular pathways in disease tissue. J Pediatr Gastroenterol Nutr. 2018;67(1):45–52. doi:10.1097/MPG.0000000000001898

17. Bjerrum JT, Nielsen OH, Riis LB, et al. Transcriptional analysis of left-sided colitis, pancolitis, and ulcerative colitis-associated dysplasia. Inflamm Bowel Dis. 2014;20(12):2340–2352. doi:10.1097/MIB.0000000000000235

18. Van der Goten J, Vanhove W, Lemaire K, et al. Integrated miRNA and mRNA expression profiling in inflamed colon of patients with ulcerative colitis. PLoS One. 2014;9(12):e116117. doi:10.1371/journal.pone.0116117

19. Olsen J, Gerds TA, Seidelin JB, et al. Diagnosis of ulcerative colitis before onset of inflammation by multivariate modeling of genome-wide gene expression data. Inflamm Bowel Dis. 2009;15(7):1032–1038. doi:10.1002/ibd.20879

20. Planell N, Lozano JJ, Mora-Buch R, et al. Transcriptional analysis of the intestinal mucosa of patients with ulcerative colitis in remission reveals lasting epithelial cell alterations. Gut. 2013;62(7):967–976. doi:10.1136/gutjnl-2012-303333

21. Bjerrum JT, Hansen M, Olsen J, Nielsen OH. Genome-wide gene expression analysis of mucosal colonic biopsies and isolated colonocytes suggests a continuous inflammatory state in the lamina propria of patients with quiescent ulcerative colitis. Inflamm Bowel Dis. 2010;16(6):999–1007. doi:10.1002/ibd.21142

22. Barbie DA, Tamayo P, Boehm JS, et al. Systematic RNA interference reveals that oncogenic KRAS-driven cancers require TBK1. Nature. 2009;462(7269):108–112. doi:10.1038/nature08460

23. Charoentong P, Finotello F, Angelova M, et al. Pan-cancer immunogenomic analyses reveal genotype-immunophenotype relationships and predictors of response to checkpoint blockade. Cell Rep. 2017;18(1):248–262. doi:10.1016/j.celrep.2016.12.019

24. Wilkerson MD, Hayes DN. ConsensusClusterPlus: a class discovery tool with confidence assessments and item tracking. Bioinformatics. 2010;26(12):1572–1573. doi:10.1093/bioinformatics/btq170

25. Langfelder P, Horvath S. WGCNA: an R package for weighted correlation network analysis. BMC Bioinf. 2008;9(1):559. doi:10.1186/1471-2105-9-559

26. Sharma R, Bhullar A, Bansal P, Kaur G, Gupta VJM. Bioinformatics analysis identifies potential hub genes, therapeutic agents, and crucial pathways in the pathogenesis of refsum’s disease. Medinformatics. 2023. doi:10.47852/bonviewMEDIN32021516

27. Alhatemi RAJ, Savaş SJM, van Veldhoven RPJ, Verhagen E, Fiore A. A weighted ensemble approach with multiple pre-trained deep learning models for classification of stroke. Npj Nanophotonics. 2024;1(1):10–19. doi:10.1038/s44310-024-00011-y

28. Charles S, Natarajan JJM. Identification of key gene modules and novel transcription factors in tetralogy of fallot using machine learning and network topological features. Medinformatics. 2024;1(1):27–34.

29. Yu G, Wang LG, Han Y, He QY. clusterProfiler: an R package for comparing biological themes among gene clusters. OMICS. 2012;16(5):284–287. doi:10.1089/omi.2011.0118

30. Garrido-Trigo A, Corraliza AM, Veny M, et al. Macrophage and neutrophil heterogeneity at single-cell spatial resolution in human inflammatory bowel disease. Nat Commun. 2023;14(1):4506. doi:10.1038/s41467-023-40156-6

31. Du J, Zhang J, Wang L, et al. Selective oxidative protection leads to tissue topological changes orchestrated by macrophage during ulcerative colitis. Nat Commun. 2023;14(1):3675. doi:10.1038/s41467-023-39173-2

32. Carlson BA, Moustafa ME, Sengupta A, et al. Selective restoration of the selenoprotein population in a mouse hepatocyte selenoproteinless background with different mutant selenocysteine tRNAs lacking Um34. J Biological Chem. 2007;282(45):32591–32602. doi:10.1074/jbc.M707036200

33. Huang Z, Hoffmann FW, Norton RL, Hashimoto AC, Hoffmann PR. Selenoprotein K is a novel target of m-calpain, and cleavage is regulated by toll-like receptor-induced calpastatin in macrophages. J Biol Chem. 2011;286(40):34830–34838. doi:10.1074/jbc.M111.265520

34. Meng XL, Chen CL, Liu YY, et al. Selenoprotein SELENOK enhances the migration and phagocytosis of microglial cells by increasing the cytosolic free Ca(2+) level resulted from the up-regulation of IP(3)R. Neuroscience. 2019;406:38–49. doi:10.1016/j.neuroscience.2019.02.029

35. Norton RL, Fredericks GJ, Huang Z, Fay JD, Hoffmann FW, Hoffmann PR. Selenoprotein K regulation of palmitoylation and calpain cleavage of ASAP2 is required for efficient FcgammaR-mediated phagocytosis. J Leukoc Biol. 2017;101(2):439–448. doi:10.1189/jlb.2A0316-156RR

36. Verma S, Hoffmann FW, Kumar M, et al. Selenoprotein K knockout mice exhibit deficient calcium flux in immune cells and impaired immune responses. J Immunol. 2011;186(4):2127–2137. doi:10.4049/jimmunol.1002878

37. Barrett CW, Reddy VK, Short SP, et al. Selenoprotein P influences colitis-induced tumorigenesis by mediating stemness and oxidative damage. J Clin Investigat. 2015;125(7):2646–2660. doi:10.1172/JCI76099

38. Barrett CW, Short SP, Williams CS. Williams CSJC, sciences mL. selenoproteins and oxidative stress-induced inflammatory tumorigenesis in the gut. Cell Mol Life Sci. 2017;74(4):607–616. doi:10.1007/s00018-016-2339-2

39. Yao Y, Xu T, Li X, et al. Selenoprotein S maintains intestinal homeostasis in ulcerative colitis by inhibiting necroptosis of colonic epithelial cells through modulation of macrophage polarization. Theranostics. 2024;14(15):5903–5925. doi:10.7150/thno.97005

40. Ahn YH, Park S, Choi JJ, et al. Corrigendum: secreted tryptophanyl-tRNA synthetase as a primary defence system against infection. Nature Microbiol. 2017;2(3):17015. doi:10.1038/nmicrobiol.2017.15

41. Kang I, Lee B-C, Lee JY, et al. Interferon-γ-mediated secretion of tryptophanyl-tRNA synthetases has a role in protection of human umbilical cord blood-derived mesenchymal stem cells against experimental colitis. BMB Reports. 2019;52(5):318. doi:10.5483/BMBRep.2019.52.5.134

42. Pontanari F, Demagny H, Faure A, et al. Wars1 downregulation in hepatocytes induces mitochondrial stress and disrupts metabolic homeostasis. Metabolism. 2025;162:156061. doi:10.1016/j.metabol.2024.156061

43. Turner D, Ricciuto A, Lewis A, et al. STRIDE-II: an update on the selecting therapeutic targets in inflammatory bowel disease (STRIDE) initiative of the international organization for the study of IBD (IOIBD): determining therapeutic goals for treat-to-target strategies in IBD. Gastroenterology. 2021;160(5):1570–1583. doi:10.1053/j.gastro.2020.12.031

44. Vulliemoz M, Brand S, Juillerat P, et al. TNF-alpha blockers in inflammatory bowel diseases: practical recommendations and a user’s guide: an update. Digestion. 2020;101 Suppl 1(Suppl. 1):16–26. doi:10.1159/000506898

45. Burk RF, Hill KE. Selenoprotein P-expression, functions, and roles in mammals. Biochim Biophys Acta. 2009;1790(11):1441–1447. doi:10.1016/j.bbagen.2009.03.026

46. Burk RF, Hill KE. Selenoprotein P: an extracellular protein with unique physical characteristics and a role in selenium homeostasis. Annu Rev Nutr. 2005;25(1):215–235. doi:10.1146/annurev.nutr.24.012003.132120

47. Short SP, Pilat JM, Williams CS. Roles for selenium and selenoprotein P in the development, progression, and prevention of intestinal disease. Free Radic Biol Med. 2018;127:26–35. doi:10.1016/j.freeradbiomed.2018.05.066

48. Sheng J, Ruedl C, Karjalainen K. Most tissue-resident macrophages except microglia are derived from fetal hematopoietic stem cells. Immunity. 2015;43(2):382–393. doi:10.1016/j.immuni.2015.07.016

49. Zigmond E, Bernshtein B, Friedlander G, et al. Macrophage-restricted interleukin-10 receptor deficiency, but not IL-10 deficiency, causes severe spontaneous colitis. Immunity. 2014;40(5):720–733. doi:10.1016/j.immuni.2014.03.012

50. Zigmond E, Varol C, Farache J, et al. Ly6C hi monocytes in the inflamed colon give rise to proinflammatory effector cells and migratory antigen-presenting cells. Immunity. 2012;37(6):1076–1090. doi:10.1016/j.immuni.2012.08.026

51. Kang I, Lee BC, Lee JY, et al. Interferon-gamma-mediated secretion of tryptophanyl-tRNA synthetases has a role in protection of human umbilical cord blood-derived mesenchymal stem cells against experimental colitis. BMB Rep. 2019;52(5):318–323.

52. Kim YT, Huh JW, Choi YH, et al. Highly secreted tryptophanyl tRNA synthetase 1 as a potential theranostic target for hypercytokinemic severe sepsis. EMBO Mol Med. 2024;16(1):40–63. doi:10.1038/s44321-023-00004-y

53. Deutschmann C, Roggenbuck D, Schierack P. The loss of tolerance to CHI3L1 – A putative role in inflammatory bowel disease? Clin Immunol. 2019;199:12–17. doi:10.1016/j.clim.2018.12.005

54. Heidari P, Haj-Mirzaian A, Prabhu S, Ataeinia B, Esfahani SA, Mahmood U. Granzyme B PET imaging for assessment of disease activity in inflammatory bowel disease. J Nucl Med. 2024;65(7):1137–1143. doi:10.2967/jnumed.123.267344

55. Shi Y, Luo S, Zhai J, Chen Y. A novel causative role of imbalanced kynurenine pathway in ulcerative colitis: upregulation of KMO and KYNU promotes intestinal inflammation. Biochim Biophys Acta Mol Basis Dis. 2024;1870(2):166929. doi:10.1016/j.bbadis.2023.166929

56. Secchia S, Beilinson V, Chen X, et al. Nutrient starvation activates ECM remodeling gene enhancers associated with inflammatory bowel disease risk in fibroblasts. BioRxiv. 2024. doi:10.1101/2024.09.06.611754

57. McIlwain DR, Berger T, Mak TW. Caspase functions in cell death and disease. Cold Spring Harb Perspect Biol. 2015;7(4):a026716. doi:10.1101/cshperspect.a026716

58. Flood B, Oficjalska K, Laukens D, et al. Altered expression of caspases-4 and −5 during inflammatory bowel disease and colorectal cancer: diagnostic and therapeutic potential. Clin Exp Immunol. 2015;181(1):39–50. doi:10.1111/cei.12617

Use of smoked tobacco

Our analysis determined that tobacco was smoked by 13.10% of the middle aged and elderly individuals in India with a rural-urban absolute difference of 7.32%. Among adults aged 15 years and above, GATS 2 found the prevalence of smoked tobacco to be 10.7%. The rural-urban difference determined was 3.6% [21]. The National Non-communicable Disease Monitoring Survey (NNMS) found the prevalence of smoked tobacco to be 12.6%, with the prevalence in rural areas being 1.5% more than urban areas among individuals aged 18–69 years [32].

In our analysis, in rural and urban areas, a higher prevalence of tobacco smokers was found among males, the elderly, Muslims, those from scheduled castes, those with less than a primary school education, those in an unskilled profession, those not living alone and North Indians. In rural areas, the highest prevalence of smoking was among the richer wealth quintiles while in urban areas, the highest prevalence was among the poorest. Previous studies too showed similar determinants of smoked tobacco use [12, 14].

The concentration index of smoked tobacco use showed a slightly higher concentration among the lower wealth quintiles. This is in line with previous literature where the concentration indices calculated for smoked tobacco using data from the National Family Health Survey (NFHS) 5 found a higher concentration among the poor as compared to the rich [33]. A compilation of 4 rounds of NFHS data from 1998 to 2021 found the prevalence of smoked tobacco consumption to be consistently higher among the lowest wealth quintiles [12].

The decomposition analysis of smoked tobacco showed that rural-urban differences in gender composition, educational status and caste were the determinants which contributed the most to the rural-urban difference in the prevalence of smoking. These factors contributed more to the difference in smoked tobacco, than smokeless tobacco, indicating that improving the educational status of the rural population or increasing their sex ratio would decrease the use of smoked tobacco in rural areas more than that of smokeless tobacco.

Gender disparities in smoking can be largely attributed to the socio-cultural beliefs prevalent across India. Among men, smoking is often a communal activity, with practices such as sharing a hukkah viewed as a sign of comradery [27]. It is also perceived as a way to build social connections and is even associated with masculinity. In contrast, women who smoke often face social stigma, as smoking is considered unconventional and inappropriate for women [27, 34]. This social judgment serves as a strong deterrent for women, contributing to the gender gap in smoking prevalence. The situation is changing presently, possibly due to shifting gender roles and changing societal norms [35]. The prevalence of smoking among females has fluctuated minorly since 2005, in contrast to the steep decline among males [36]. This transition may alter the determinants of tobacco use in the future. The prevalence of smoking among rural women is considerably higher than urban women, possibly due to their lower educational status and the prevalent use of inexpensive bidis (traditional hand-rolled cigarettes) [35, 36].

Use of smokeless tobacco

In this analysis, the prevalence of the use of smokeless tobacco in middle aged and elderly Indians was found to be 20.43%, with the rural prevalence being 23.82%, urban prevalence being 13.03% and difference 10.79%. The NNMS determined the prevalence of use of smokeless tobacco to be 24.7%, with a rural (28.3%)-urban (17.6%) difference of almost 10% among individuals 18 to 69 years of age [32]. As per GATS-2, the prevalence of smokeless tobacco use among Indians more than 15 years of age was 21.4%, with a rural prevalence of 24.6%, an urban prevalence of 15.2% and a difference of almost 10% [21].

In our study, the middle aged and elderly individuals with a higher prevalence of smokeless tobacco use as compared to their counterparts were males, Muslims, scheduled tribes, those from the poorer and poorest wealth indices, those with less than a primary education, Northeast Indians, and those who had no media exposure. In rural areas, those with a skilled profession and in urban areas, those with an unskilled profession had higher prevalence of tobacco use. In urban areas, smokeless tobacco use was more among the elderly. A significantly higher concentration of smokeless tobacco users was found among the poor as compared to the rich.

An analysis of GATS 2 data of Indian women found that smokeless tobacco consumption was more among those with lesser education, those from scheduled tribes, those from the poor wealth quintiles and those from Northeast India [17]. An analysis of GATS 2 data by Nair et al. from Northeast Indian states found similar determinants [37]. A decomposition analysis of data from Northeast India found that 90% of the difference in the prevalence of smokeless tobacco use between men and women could be attributed to differences in the age, employment status, education and wealth status [37], emphasising the high contribution of socioeconomic determinants to smokeless tobacco use.

The decomposition analysis showed that differences in the regions of domicile, occupation and education between rural and urban populations contributed the most to the difference in the prevalence of smokeless tobacco use. This re-emphasises the fact that socio-economic factors and regional cultural differences are important determinants of smokeless tobacco consumption.

Smokeless tobacco use is socially accepted in several rural areas of India due to longstanding traditional beliefs and practices [7]. These cultural norms contribute to the wide variation in both the prevalence and types of smokeless tobacco products used across states. In certain eastern and northeastern states, its use remains deeply rooted in local customs [37], and offering betel nut with betel leaf, for instance, is regarded as a sign of hospitality [38]. In these regions, high rates of smokeless tobacco use are also observed among women and youth [17, 39]. In South India, particularly among some indigenous rural tribes, chewing tobacco holds such cultural significance that it is prioritized over food and is commonly used during weddings, festivals, funerals, and even pregnancy [7]. However, studies on tobacco-related beliefs and practices from other states are lacking, highlighting the need for further research to inform culturally tailored interventions.

Policy implications and recommendations

Since the prevalence of smoked and smokeless tobacco consumption was found to be higher in rural areas than urban areas, control efforts should be focussed more there, through cessation programs and strengthened law enforcement. To address the varying prevalence of tobacco use across different states in the country, we propose differential taxing by states, as well as a more decentralized tobacco cessation program design and implementation at the state or regional level. Social and behaviour change communication should be strengthened to address the socio-cultural factors of tobacco use.

In line with previous literature [13], our results emphasise the importance of focussing on the underlying inequities which contribute to differential tobacco consumption. Improving the educational status of the rural population could decrease the difference in the prevalence of smoked tobacco consumption by 41% and smokeless tobacco by 20%. Given the significant negative concentration index, tobacco cessation activities should target poorer wealth quintiles.

The distinct cultural habits and beliefs which have led to regional variations in smokeless tobacco consumption should be explored using qualitative research methods and addressed on an individual basis.

Strengths and limitations

To the best of our knowledge, this is the first study which could be found which delineated the factors contributing to the difference in the rural-urban prevalence of smoked and smokeless tobacco consumption among middle aged and elderly adults in India. The results are generalisable to the population aged 45 years and above in India, given the nationally representative sample included and the appropriate sample weights used.

However, this study has some limitations as well. Given the cross-sectional nature of the primary data-set, temporality could not be determined and there is a chance of reverse-causality. Bonferroni correction was not used while determining the factors associated with tobacco use. Since the data collected was self-reported, there may be underreporting of tobacco use because of social desirability bias. Additional factors which could potentially influence tobacco use, such as knowledge about the toxic effects of tobacco, were not determined for this analysis. The sample chosen for this analysis was middle aged and elderly Indians, due to the lack of data openly available for the entire adult population, which would have made the results more generalisable.

Kia Motors Port Qasim is offering a limited-time maintenance discount in connection with Pakistan’s Independence Day. The dealership is providing a 30% reduction on car maintenance services, valid until August 31, 2025.

Maintenance Services Covered

The discounted service package includes the following:

• Free Vehicle Wash: Customers can avail a complimentary exterior car wash as part of the promotion.
• Free AC System Check: Kia Port Qasim will conduct a no-cost air conditioning inspection to assess cooling performance and address any minor issues.
• Discounted Periodic Maintenance: The 30% discount applies to routine maintenance services, which may include engine oil replacement, filter changes, brake checks, and other standard inspections depending on the vehicle’s service schedule.

Location and Contact Details

The dealership is located at Plot 164, Jogi Mor, Main National Highway, near Abbott Laboratory, Karachi. Customers may contact Kia Port Qasim through the following numbers to book an appointment:

• +92 333 1037938
• +92 333 1037952

Additional Notes

• The promotion is only applicable at the Kia Port Qasim branch.
• It is recommended to schedule service appointments in advance to avoid delays due to increased seasonal demand.
• The discount applies only to labor and service charges; spare parts or accessories are not included unless otherwise specified.

This initiative supports routine vehicle upkeep while offering temporary financial relief during the national holiday season.