Disease burden, machine learning‑based risk factor identification in uterine cancer: A cross‑sectional and bibliometric analysis.

Our analysis showed that incidence and DALY rates peaked at ages 65 to 69, likely due to menopause-related hormonal decline and immune impairment. The highest number of deaths occurred at age ≥ 95, reflecting advanced age and diminished immunity. Prevalence peaked at 55 to 64 years, a critical period of perimenopause marked by rapid estrogen decline, [ 26 ] and often accompanied by reduced basal metabolic rate and elevated BMI – a well‑established UC risk factor driven by chronic inflammation, hormonal disruption, and excess energy supply to malignant cells. [ 27 , 28 ] The United States, Russia, Italy, and Portugal had high cumulative incidence and prevalence, possibly linked to high‑fat diets that promote estrogen synthesis and ovarian epithelial proliferation. [ 29 ] Based on the NHANES database and ML methodologies, this study identified molybdenum, triglycerides, cadmium, age, lead, total cholesterol, cobalt, hypertension, uric acid, and MAP as key factors associated with UC. The analysis revealed that the SHAP value for age increases with advancing age, and a higher SHAP value in older individuals may indicate an elevated risk of disease onset. This finding aligns with results from the GBD study, which reported higher mortality rates among women in older age groups. A primary underlying reason may be the hormonal imbalance associated with aging, particularly during perimenopause and postmenopause, which can promote endometrial hyperplasia and potentially carcinogenesis. [ 30 ] Furthermore, the heavy metals molybdenum, cadmium, lead, and cobalt were identified as significant influencing factors in the pathogenesis of UC. Molybdenum is an essential trace element widely distributed in nature. Molybdenum toxicity can interfere with copper absorption and elevate uric acid levels. [ 31 ] Serum uric acid, a potent antioxidant abundant in the blood that acts as a free radical scavenger, has been linked to cancer risk in epidemiological studies. A large prospective study involving 28,613 elderly women suggested that elevated uric acid levels may contribute to cancer development. [ 32 ] Uric acid, the final product of human purine metabolism, is produced from both endogenous and exogenous purines degraded by xanthine oxidase. Serum uric acid is considered a surrogate marker for metabolic disturbances. Existing evidence suggests that high serum uric acid levels may promote cancer development by inducing chronic inflammation and increasing the production of reactive oxygen species. Previous studies have indicated that women with endometrial cancer tend to have higher serum uric acid levels, suggesting that uric acid may play a role in the association between conditions like endometriosis and endometrial cancer. [ 33 ] In the context of female reproductive system diseases, uric acid may be involved in cancer pathogenesis by affecting oxidative stress and redox processes. [ 34 ] Additionally, high uric acid levels can impair mitochondrial function, increase superoxide production, subsequently lead to lipid metabolism disorders, and promote the generation of pro-inflammatory factors, all of which may further contribute to cancer development. [ 35 ] Hypertension is the most prevalent comorbidity among cancer patients, attributable to shared risk factors and cardiovascular complications induced by cancer therapies. [ 36 ] Epidemiological studies have indicated that hypertension is an established independent risk factor for endometrial cancer. [ 37 ] In other words, individuals with hypertension face an elevated risk of developing endometrial cancer. Bibliometric analysis indicates that the United States has the highest volume of publications in the field of UC research. Furthermore, institutional analysis reveals that most of the leading institutions in terms of publication output are also based in the United States, suggesting a dominant and leading position of US research in this area. Keyword analysis identified terms such as multicenter, endometrial neoplasms, disparity, race, and guideline as currently having high impact, indicating that clinical research on UC is a prominent focus in the field. Racial disparities in UC-related outcomes have been documented. Black women present with more aggressive tumors, poorer histology, and face a higher mortality rate from UC compared to any other racial group. Research exploring outcomes for Black individuals consistently shows worse results relative to all other racial groups. Black women with UC tend to have higher obesity rates, are diagnosed at a younger age, and have more comorbidities prior to diagnosis than White women. [ 38 ] These factors are all potentially associated with their ability to regain preoperative functional levels and their need for institutional care. For women with a BMI below 40, functional loss was less pronounced among Black women compared to non-Black women; however, non-Black women exhibited a relatively low risk of functional loss across all BMI ranges. Other key factors leading to the loss of functional independence include age over 60 and poorer preoperative status – both of which have been separately linked to post-hospitalization functional decline and a higher risk of post-discharge mortality. [ 39 ] These findings are consistent with the results of our ML screening, which identified age as a significant factor associated with UC onset. This study also has several limitations. First, the NHANES database is cross-sectional in design, and future research should conduct in-depth studies to enable causal inference. Second, heavy metal measurements were taken at a single time point and thus cannot reflect the effects of long-term exposure. Third, the bibliometric analysis was restricted to the WoSCC, and future studies should expand the data sources.

The disease burden of UC peaks at ages 65 to 69, with the highest incidence rates in the United States, Russia, Italy, and Portugal. ML identified ten core features, including heavy metals (molybdenum, cadmium, lead) and metabolic indicators (triglycerides, hypertension, uric acid). Bibliometric analysis reveals the leadership of the United States and emerging hotspots such as disparity and multicenter research. These findings inform prevention strategies and highlight novel risk factors for future research.

Uterine cancer (UC) represents the most prevalent malignancy of the female reproductive system, primarily affecting postmenopausal women. [ 1 ] Over recent decades, its global incidence has risen by 132%, making it a significant public health concern worldwide. [ 2 ] With continued population growth, aging, and increasing prevalence of risk factors, the disease burden of UC is projected to rise further. It ranks as the most common gynecological cancer and the fourth most common cancer overall among women. [ 3 ] This growing burden is largely attributable to population expansion, aging demographics, and shifts in key risk factors, some of which are linked to socioeconomic development. [ 4 ] Globally, UC is among the top 5 cancers in women with high mortality rates, [ 5 ] and its incidence and mortality exhibit considerable geographical variation. Understanding temporal and spatial trends in the UC burden at global, regional, and national levels is therefore essential for informed policy-making and effective resource allocation. Although UC occurs mainly after menopause, increasing incidence rates have been observed across all age groups. [ 6 ] Several studies have described the epidemiological characteristics of UC at regional or national levels. [ 7 ] As an age-related disease, its epidemiology is influenced not only by chronological age but also by period effects and birth cohort effects. [ 1 ] However, the latest disease burden of UC has not yet been reported. The absence of routine screening complicates the early diagnosis of UC. Data from 2015 to 2021 indicate that 29% of cases were diagnosed only after the cancer had progressed to regional or distant stages. [ 8 ] Identifying and understanding the factors associated with the onset of UC is of great significance for its prevention and treatment. Existing studies have shown that a high body mass index (BMI), [ 9 ] obesity, metabolic syndrome (MS), diabetes mellitus, and hypertension [ 5 ] are established risk factors for UC. Although traditional statistical models like logistic regression (LR) [ 6 , 10 , 11 ] and Cox proportional hazards models, [ 12 , 13 ] have contributed to predicting UC risk, they possess inherent constraints for clinical application. LR models inherently assume linearity, potentially oversimplifying the intricate, nonlinear relationships between variables in real-world clinical contexts. Moreover, their performance can be adversely affected by multicollinearity. [ 14 ] These limitations underscore the importance of developing more sophisticated and reliable predictive tools for enhancing patient management in UC. Machine learning (ML) has emerged as a powerful, computer-assisted approach for data mining and analysis, and is now widely adopted as a predictive tool across various engineering and medical fields. [ 15 ] Compared to conventional statistical methods, ML often delivers superior predictive accuracy. [ 16 ] Furthermore, recent advances in ML – ranging from traditional linear models to complex deep learning architectures – enable the handling of large, heterogeneous datasets and can capture intricate, nonlinear relationships among variables that may remain hidden under conventional statistical analysis. [ 17 , 18 ] These capabilities support the development of precise and reliable diagnostic models for identifying individuals at high risk of UC, thereby facilitating the implementation of tailored prevention strategies and clinical interventions. Therefore, we hypothesize that ML models can outperform traditional LR in identifying key risk factors associated with the onset of UC, and that the disease burden of UC exhibits significant spatiotemporal heterogeneity, with rising trends in specific regions and age groups. This study investigates the latest disease burden of UC using the Global Burden of Disease (GBD) database. Additionally, it utilizes the National Health and Nutrition Examination Survey (NHANES) database and ML methods to identify factors associated with UC onset, and employs bibliometric analysis to clarify current research hotspots in the UC field.

Data curation: Hailang Wang. Investigation: Yu Li. Methodology: Haibo Wang. Software: Shuhao Wang, Zhi Chen. Supervision: Jing Dong.

The GBD database provides comprehensive estimates of incidence, prevalence, mortality, years of life lost, years lived with disability (YLD), and disability-adjusted life years (DALYs) for a wide range of health conditions. [ 19 ] GBD 2023 builds upon and updates previous estimates (including those from GBD 2021) of morbidity and mortality by geographic region, time period, age, and sex, contributing to a refined global assessment of disease burden. This iteration analyzes 371 diseases and injuries, encompassing 292 causes of death and 88 environmental, occupational, behavioral, and metabolic risk factors. [ 20 ] Data on the 2023 global incidence, mortality, prevalence, and DALYs for UC were extracted from the GBD database ( https://vizhub.healthdata.org/gbd-results/ ). Both crude and age-standardized rates were obtained. YLDs are calculated as the product of the prevalence of a specific health outcome and its corresponding disability weight. In this study, age-standardized mortality rates, DALYs, and other epidemiological indicators for UC were sourced from the GBD database. The GBD framework systematically quantifies health loss across 204 countries and territories. [ 21 ] Data were stratified into 5-year age groups from 20 to 95 years and older, and included aggregates for “all ages” as well as age-standardized metrics. Specifically, we extracted: age- and region-specific mortality and DALYs, along with age-standardized mortality rates and age-standardized DALY rates, each accompanied by 95% uncertainty intervals and socio-demographic Index scores for the 204 regions. The secondary analysis of de-identified GBD data was exempt from review by the Institutional Review Board of the University of Washington. All data are freely available through the Global Health Data Exchange. Ethical oversight – including a waiver of informed consent – was granted by the University of Washington Institutional Review Board. We conducted a cross‑sectional study using data from the NHANES public database. All participants or their proxies provided informed consent. NHANES offers comprehensive, accurate, and systematically collected data, which supports evidence‑based nutrition and public‑health policy making. The survey is maintained through a dedicated management system, with data updated regularly and freely accessible to the public. [ 14 ] The NCHS Ethics Review Board approved the NHANES protocol, and all participants gave written informed consent. Since this is a secondary analysis of de-identified public data, no further institutional review board approval was necessary. NHANES employs a multistage, stratified, cluster probability sampling design. In this study, we used all eligible female participants from the NHANES database from 1999 to 2016. Data encompassed demographic and socioeconomic variables collected via questionnaire, including gender (female), age, BMI (kg/m 2 ), race/ethnicity, educational attainment, marital status, annual family income, the family income-to-poverty ratio (FIRP), smoking, drinking, hypertension, and diabetes. This range of variables supports a thorough characterization of the cohort and facilitates analysis of how demographic and socioeconomic contexts may modify associations between environmental factors and UC. An initial 81,913 individuals were considered, but after exclusions – age < 20 years or pregnancy; no documented cancer diagnosis; missing demographic variables (income, education, marital status, or age); no heavy metal data; and incomplete BMI records, among others – the final analytical sample comprised 1346 cases. Endocrine disruption, oxidative stress, and inflammation are all closely associated with the pathogenesis of UC. Heavy metals and metabolic dysfunction are intimately linked to these factors. Therefore, we next focus on indicators related to heavy metals and metabolic function. Heavy metal analysis included 7 elements (barium, cadmium, cobalt, cesium, molybdenum, lead, tungsten) measured in urine specimens. All quantifications were performed at the National Center for Environmental Health Laboratory via inductively coupled plasma dynamic reaction cell mass spectrometry, a technique recognized for its precision in trace-level metal detection. The rigorous analytical protocol ensured high data reliability, forming a solid basis for evaluating potential links between heavy metal exposure and UC. Metabolic-related measurements and calculations in this study were conducted as follows: Blood pressure was measured at the Mobile Examination Center. Fasting plasma glucose, triglycerides, total cholesterol, and uric acid levels were determined enzymatically using an autoanalyzer, with detailed assay protocols available on the NHANES website. Mean arterial pressure (MAP) was calculated as diastolic blood pressure + 1/3 × (systolic blood pressure − diastolic blood pressure). The triglyceride-glucose (TyG) index was computed as ln[fasting triglycerides (mg/dL) × fasting glucose (mg/dL)/2]. Additionally, a metabolic score (MS) was constructed as the sum of the z -transformed values of 4 components: total cholesterol, uric acid, MAP, and the TyG index. [ 22 ] In NHANES, UC cases were identified through self-reported questionnaire data by first confirming a UC diagnosis and then excluding individuals with other cancers. K-nearest neighbor interpolation was used to impute missing data, and the Boruta algorithm performs feature selection by evaluating the importance of each attribute relative to randomly generated shadow features. During each iteration, a random forest (RF) model calculates z ‑scores for both the true features and the shadow features. A true feature is retained as significant only if its z ‑score exceeds the maximum z ‑score of the shadow features across multiple independent runs. [ 14 , 23 ] A new dataset was constructed based on the features selected by the Boruta algorithm. The dataset was randomly divided into a training subset (70% of the data) and a testing subset (the remaining 30%). The training portion was employed for both model selection and hyperparameter tuning. To support this process, a 10-fold cross-validation strategy was implemented: the training data were split into ten equally sized folds. In each round of cross-validation, 9 of these folds served as the training component, while the remaining single fold was used as the validation set. The overall performance of a given model was obtained by averaging the evaluation metrics calculated across all ten iterations. Once the models were tuned, the independent test set (30% of the original data) was used for final performance assessment. For each ML algorithm under consideration, the best‑performing model was identified based on the highest area under the receiver operating characteristic curve, decision curve analysis (DCA), and calibration curve. Refer to previous research. [ 24 ] For model training and evaluation, 11 ML methods were applied, consisting of CatBoost, Lasso, Gradient Boosting Tree (GBM), Decision Tree (DT), naive bayes (NB), support vector machines (SVM), LightGBM (LGB), neural network (NN), XGBoost, LR, and RF. A literature search was performed in the Web of Science Core Collection (WoSCC) [ 25 ] on December 20, 2025, ensuring contemporaneity of the retrieved records. Using the topic search “TS = uterine cancer,” 2041 relevant articles published from 2005 to 2025 are identified and included in the analysis. All relevant article data were obtained from WoSCC and compiled in Excel 2016. Extracted variables encompassed authors, affiliations, countries/regions, journals, paper and citation counts, publication year, keywords, and references. Screening and data extraction were performed independently by 2 researchers. The bibliometric analysis was conducted using VOSviewer (v1.6.10.0) and CiteSpace (v6.2.R6). Common bibliometric indicators, such as publication and citation counts, were analyzed to assess scholarly output. VOSviewer was employed to generate and visualize co‑occurrence networks, where node size corresponded to publication frequency, line thickness indicated association strength, and node color denoted thematic clusters or time periods. Citespace was used to perform cluster analysis, display timeline views, and detect citation bursts in references and keywords, which helped identify emerging trends and core research themes in the field of UC. Participant demographics were presented as mean ± SD for continuous variables and as counts (percentages) for categorical variables. Statistical analyses were performed using EmpowerStats (X&Y Solutions) and R (The University of Auckland), applying the Kruskal–Wallis test for non‑normally distributed data and the chi‑square test for categorical data, with statistical significance defined as P  < .05. Using a generalized linear regression model to calculate the odds ratios (OR) and 95% confidence intervals for the associations between key features and UC.

The cumulative incidence rate (CIR) and cross-sectional prevalence rate (CPR) of UC worldwide in 2023 were shown in Figure 1 A, B, respectively. The results indicated that in 2023, the United States, Russia, Italy, and Portugal had the highest levels of CIR and CPR for UC, suggesting that these countries should strengthen disease prevention and control measures. This was followed by the majority of European countries. The DALYs, deaths, incidence, and prevalence across different age groups in various countries were presented in Figure 1 C–F. The results showed that Oceania had the highest DALYs in the age group 65 to 69, followed by the Caribbean. Western Europe had the highest number of deaths in the age group 95 and above, followed by Andean Latin America. High-income North America reported the highest incidence in the age groups 65 to 69 and 70 to 74, followed by the groups 75 to 79 and 60 to 64. High-income North America also had the highest prevalence in the age group 65 to 69, followed by the groups 70 to 74 and 60 to 64. Analysis of the disease burden of UC in 2023. (A, B) Global spatial distribution of age-standardized CIR (A) and CPR (B). Age-specific patterns of (C) DALYs, (D) deaths, (E) incidence, and (F) prevalence across different geographic regions. CIR = cumulative incidence rate, CPR = cross-sectional prevalence rate, DALYs = disability-adjusted life years, UC = uterine cancer. Subsequently, we selected 12 representative countries or regions (Canada, China, France, Germany, India, Italy, Netherlands, Russia, South Africa, Spain, the United Kingdom, and the United States) to analyze DALYs (Fig. 2 ), deaths (Fig. 3 ), incidence (Fig. 4 ), and prevalence (Fig. 5 ) across different age groups. Across the 12 selected countries, the highest DALYs and incidence consistently clustered in the 55 to 79 years age range, peaking at 65 to 69 years. The US recorded the highest national incidence in the 65 to 69 years group (18,005 cases) and the highest prevalence in the 60 to 64 years group (1,50,288 cases). Deaths were most frequent at ages 70 to 84 years, with the US reporting 2320 deaths in the 70 to 74 years subgroup. DALYs analysis for different countries: (A) Canada; (B) China; (C) France; (D) Germany; (E) India; (F) Italy; (G) Netherlands; (H) Russia; (I) South Africa; (J) Spain; (K) the United Kingdom; (L) the United States. DALY = disability-adjusted life year. Deaths analysis for different countries: (A) Canada; (B) China; (C) France; (D) Germany; (E) India; (F) Italy; (G) Netherlands; (H) Russia; (I) South Africa; (J) Spain; (K) the United Kingdom; (L) the United States. Incidence analysis for different countries: (A) Canada; (B) China; (C) France; (D) Germany; (E) India; (F) Italy; (G) Netherlands; (H) Russia; (I) South Africa; (J) Spain; (K) the United Kingdom; (L) the United States. Prevalence analysis for different countries: (A) Canada; (B) China; (C) France; (D) Germany; (E) India; (F) Italy; (G) Netherlands; (H) Russia; (I) South Africa; (J) Spain; (K) the United Kingdom; (L) the United States. Table S1 , Supplemental Digital Content 1 summarizes baseline characteristics of the 1345 participants (1160 without UC, 185 with UC). The UC group was significantly older (60.83 vs 47.37 years, P  < .001) and had higher BMI (30.31 vs 29.10, P  = .037), MAP (89.38 vs 86.82, P  = .008), and TyG index (8.89 vs 8.71, P  < .001). Metabolic parameters (total cholesterol, glucose, uric acid, MS) were also elevated in the UC group (all P  < .05). For heavy metals, levels of barium, cadmium, cesium, molybdenum, lead, and tungsten were generally lower in UC cases, with several differences reaching significance. Socioeconomic factors: the UC group had lower FIRP (2.10 vs 2.58, P  < .001) and higher proportions of lower education (e.g., less than ninth grade: 21.08% vs 9.82%, P  < .001). Comorbidities were more frequent in UC: hypertension (57.84% vs 29.46%), diabetes (21.62% vs 11.20%), current smoking (47.57% vs 35.75%; all P  < .01). Marital status showed more widowed individuals in UC (24.86% vs 10.42%, P  < .001), and racial composition had a higher proportion of non-Hispanic White individuals in UC (52.97% vs 43.76%, P  = .014). The results of feature selection using the Boruta algorithm are displayed in Figure 6 . The 20 variables most strongly associated with UC, ranked by their z -value in descending order, are: education level, FIRP, MAP, uric acid, drinking, glucose, annual family income, hypertension, MS, TyG, total cholesterol, triglycerides, age, barium, tungsten, cesium, cobalt, cadmium, molybdenum, and lead. This study is an exploratory analysis using ML methods, where the sample size requirement is typically based on the events per variable rule of thumb or model stability considerations. After feature selection, approximately 20 core features were used for modeling. According to the rule of thumb that the number of events for a binary outcome should be at least 10 times the number of candidate variables, the number of UC cases is sufficient to support ML modeling with 20 variables. Implementation of Boruta-based feature selection. BMI = body mass index, FIRP = family income-to-poverty ratio, MAP = mean arterial pressure, MS = metabolic syndrome, TyG = triglyceride-glucose. Using the features selected by the Boruta algorithm, we conducted comprehensive model training and evaluation. Calibration curves for the training set (Fig. 7 A) showed that LR and Lasso exhibited excellent calibration properties. Across all models, training‑set AUCs ranges from 0.821 to 1.0, with RF, XGBoost, and LGB achieving the highest performance (AUC = 1.0; Fig. 7 B), suggesting some degree of overfitting. DCA of the training set (Fig. 7 C) indicated that LGB and RF yielded larger net benefit areas. For the test set, the calibration curve (Fig. 7 D) demonstrated optimal performance for the LR model. Test‑set AUCs spanned 0.768 to 0.964, with RF achieving the highest value (0.964; Fig. 7 E), while DCA (Fig. 7 F) showed that GBM has a larger net benefit area. The full list of testing‑set AUCs for the 11 algorithms was as follows: CatBoost (0.964), DT (0.852), GBM (0.962), LGB (0.929), LR (0.768), Lasso (0.790), NB (0.888), NN (0.774), RF (0.964), SVM (0.795), and XGBoost (0.952). Integrating the results from receiver operating characteristic, DCA, and calibration curves, we selected RF, LGB, and XGBoost as the top‑performing algorithms, which were subsequently employed for feature importance analysis and selection. Presentation of the performance assessment and interpretability analysis of 11 machine learning models for depression identification. The training set results (A–C) and test set results (D–F) each display a full suite of evaluation curves: calibration curve, ROC, and decision curve analysis. DT = decision tree, GBM = Gradient Boosting Tree, LGB = lightGBM, LR = logistic regression, NB = naive bayes, NN = neural network, RF = random forest, ROC = receiver operating characteristic, SVM = support vector machine. The important features identified by the 3 models were shown in Figure 8 . For the LGB model, the top 5 features were molybdenum, triglycerides, cadmium, age, and lead ( Fig. S1A , Supplemental Digital Content 2). Among the 15 key features, molybdenum had the most significant impact (Fig. 8 A). Using 1 sample as an example, its predicted value f ( x ) was 0.997, lower than the average expected value E ( f ( x )) (1.14), indicating that this individual’s UC risk was below the average level. This result was primarily influenced by features such as triglycerides, molybdenum, cadmium, total cholesterol, age, lead, hypertension, cobalt, and BMI ( Fig. S1B , Supplemental Digital Content 2). Identification of significant features: (A) SHAP beeswarm plot for LGB; (B) SHAP beeswarm plot for RF; (C) SHAP beeswarm plot for XGBoost. BMI = body mass index, FIRP = family income-to-poverty ratio, LGB = lightGBM, MAP = mean arterial pressure, MS = metabolic syndrome, RF = random forest, SHAP = shapley additive explanations, TyG = triglyceride-glucose. The RF model identified the top 5 features as BMI, drinking, glucose, hypertension, and age ( Fig. S1C , Supplemental Digital Content 2). BMI had the most significant impact (Fig. 8 B). For the same sample, its predicted value f ( x ) was −1362, lower than the average expected value E ( f ( x )) (0), further confirming the individual’s lower UC risk. The main influencing factors included drinking, BMI, glucose, hypertension, age, MS, uric acid, MAP, and annual family income ( Fig. S1D , Supplemental Digital Content 2). The top 5 features selected by the XGBoost model were triglycerides, cadmium, tungsten, molybdenum, and age ( Fig. S1E , Supplemental Digital Content 2). Among the key features, triglycerides had the greatest impact, indicating a correlation with an increased risk of UC (Fig. 8 C). For this sample, the model output a predicted value f ( x ) of 1.01, still lower than the average expected value of 1.14, suggesting a below-average risk level. This was mainly attributed to the effects of features such as cadmium, triglycerides, molybdenum, lead, total cholesterol, age, cesium, TyG, and tungsten ( Fig. S1F , Supplemental Digital Content 2). Following overlap analysis, a set of hub features was identified, comprising molybdenum, triglyceride, cadmium, age, lead, total cholesterol, cobalt, hypertension, uric acid, and MAP ( Fig. S2 , Supplemental Digital Content 3). The relationship between the levels of these hub features and their corresponding shapley additive explanations (SHAP) values was displayed in Figure S3 , Supplemental Digital Content 4. The presence of hypertension was associated with elevated SHAP values. For MAP and age, SHAP values demonstrated a positive correlation with increasing feature levels. In contrast, the SHAP values for the remaining variables exhibited fluctuating or non-monotonic trends across their respective ranges. Interactions between the SHAP values of different features were presented in Figure S4 , Supplemental Digital Content 5. Notable interaction pairs were observed, including: MAP and cadmium; age and cobalt; cadmium and triglyceride; cobalt and molybdenum; molybdenum and cadmium; lead and age; total cholesterol and triglyceride; uric acid and lead; and hypertension and age. The correlations between important features and UC risk were shown in Table 1 , and the results indicated that hypertension had the highest OR value for the development of UC. Relationships between covariates and UC risk. CI = confidence intervals, OR = odds ratio, UC = uterine cancer. The annual publication output in the field of UC research exhibited a 3-stage pattern: a phase of slow development from 2005 to 2014, with annual publications generally remaining below 100; a period of steady growth from 2015 to 2019, during which annual outputs approached or exceeded 100; and a phase of rapid expansion from 2020 to 2025, characterized by annual publications consistently surpassing 150 (Fig. 9 ). Annual publication volume within UC research. UC = uterine cancer. The country co-occurrence network was presented in Figure 10 A, revealing close collaborative relationships among different countries in this research domain. The United States was represented by the largest node, indicating it has the highest publication volume, followed by Japan, China, and Germany. The institutional co-occurrence network, shown in Figure 10 B, similarly demonstrated extensive collaboration. Leading institutions in terms of publication output include the University of Texas MD Anderson Cancer Center, Columbia University, Memorial Sloan Kettering Cancer Center, the National Cancer Institute, and Stanford University. The author co-occurrence network (Fig. 10 C) identified approximately 10 major collaborative teams led by influential researchers. Authors with the highest publication counts were Jason D. Wright, Dawn L. Hershman, June Y. Hou, John K. Chan, and Ling Chen. Bibliometric analysis in the field of UC. Network of co-occurring countries (A), institutions (B), authors (C), keywords (D), keywords network by average published year (E). (F) The top 25 most influential keywords. (G) Cluster map of keywords over time. UC = uterine cancer. The keyword co-occurrence network was displayed in Figure 10 D. Significant keywords were grouped into 5 main clusters. Cluster 1 (red, 59 keywords) primarily foused on the treatment of UC. Cluster 2 (green, 52 keywords) centers on research related to risk factors and disease burden. Cluster 3 (blue, 44 keywords) is mainly concerned with clinical aspects. Cluster 4 (yellow, 39 keywords) concentrated on comorbidities. Cluster 5 (purple, 4 keywords) addressed research on UC across different racial groups. The keyword cluster analysis based on average publication year (Fig. 10 E) identified pembrolizumab, proliferation, multicenter, and guideline as relatively novel and emerging terms. The 25 most influential keywords (Fig. 10 F) currently included multicenter, endometrial neoplasms, disparity, race, and guidelines. A temporal span analysis of keywords formed 6 clusters (ovarian cancer, mutations, laparoscopy, chemotherapy, tamoxifen, endometrial hyperplasia), all of which maintained high relevance and influence in current research.

We are grateful to every participant involved in this research.