Integrating epigenetic and metabolic indicators for non-invasive endometrial cancer triage: a machine learning approach

Abstract

Objective

Abnormal uterine bleeding (AUB) is a primary symptom indicative of endometrial cancer (EC), yet its diagnosis still primarily relies on invasive biopsy. This study aimed to develop and validate a non-invasive, interpretable machine learning (ML) model integrating local epigenetic (exfoliated cell CDO1 methylation) and systemic metabolic indicators for EC triage.

Methods

CDO1 was identified as a potential biomarker via bioinformatics. Its diagnostic performance was assessed using quantitative methylation-specific PCR (qMSP) in a clinical cohort of 267 women with AUB. Subsequently, a multimodal ML framework was developed that incorporated logistic regression (LR), support vector machine (SVM), Random Forest (RF), and eXtreme Gradient Boosting (XGBoost) algorithms to integrate methylation data with clinical metabolic profiles. Model interpretability was ensured by SHapley Additive exPlanations (SHAP) analysis, with final testing in an internal holdout cohort.

Results

CDO1 hypermethylation was identified as a significant epigenetic alteration associated with metabolic pathways in EC. Within our clinical cohort, CDO1 methylation emerged as a strong independent risk factor, with an odds ratio (OR) of 12.76 and a 95% confidence interval (CI) ranging from 6.31 to 25.80. Using LASSO regression, four critical predictive variables were identified: CDO1 methylation, menopausal status, diabetes and age. After integrating CDO1 methylation and metabolic profiles, the support vector machine model demonstrated stable performance in the validation set, achieving an area under the curve (AUC) of 0.775 (95% CI: 0.669–0.880), with a sensitivity of 0.800 and a specificity of 0.733. SHAP analysis was used to elucidate the contribution of each feature to the predictive model. The finalized SVM-based model’s diagnostic utility was further confirmed in an internal holdout cohort, yielding an AUC of 0.904 (95% CI: 0.838–0.969), with a sensitivity of 0.833 and a specificity of 0.905, demonstrating its preliminary clinical reliability. Using our model with a cutoff of 0.647 in the internal holdout cohort (n = 160), 24 patients were triaged for biopsy, with 2 EC cases missed, representing a clinical false-negative rate of 16.7%. The finalized SVM-based predictive model was developed into a free online risk calculator (https://phw1996.shinyapps.io/EC_risk/) for real-time clinical triage.

Conclusions

Our SVM-based model offers a non-invasive and highly specific approach for the triage of patients experiencing AUB. By correlating localized molecular changes with systemic metabolic dysregulation, this tool supports personalized risk stratification and clinical triage. It is designed for relative risk stratification, and external recalibration is required for clinical absolute risk estimation.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12905-026-04680-z.

Introduction

Endometrial cancer (EC) is one of the most prevalent malignancies of the female reproductive system globally [1]. Its incidence has been progressively increasing in recent years, thereby emerging as a significant public health issue [2]. Statistics indicate that in 2024 alone, there were approximately 66,200 new EC cases in the United States. Similarly, China has experienced a substantial increase in EC incidence, with a notable trend toward younger age at onset [3]. This rise is accompanied by an increase in EC-related mortality, further establishing it as a major public health burden. Abnormal uterine bleeding (AUB) serves as the primary early symptom, leading to clinical evaluations in millions of women annually. Conventional clinical tools have notable limitations in screening EC among patients with AUB. Serum CA125 fails to serve as an independent EC risk factor in AUB populations [4], while HE4 exhibits insufficient diagnostic specificity for EC [5]. In women with postmenopausal bleeding, adopting an endometrial thickness cutoff of 5 mm measured by transvaginal ultrasonography (TVS) yields a diagnostic sensitivity of 96.2% and a NPV of 99.3% for EC; however, the corresponding specificity is merely 51.5%. Worse still, TVS achieves even lower specificity when screening for EC in premenopausal women [6]. Moreover, the current diagnostic gold standard—office biopsy or dilation and curettage (D&C)—is inherently invasive, frequently associated with patient discomfort, and occasionally yields insufficient tissue for analysis. Given that only 5–15% of women with AUB are ultimately diagnosed with EC [7], there is an urgent clinical need for a non-invasive, high-performance triage tool to reduce unnecessary surgical procedures.

Epigenetic modifications, notably DNA methylation, have been recognized as early and stable markers of malignant transformation in the endometrium [8]. Among these, hypermethylation of the CDO1 (Cysteine Dioxygenase Type 1) promoter has been identified as a significant tumor suppressive signature. In 2017, Huang et al. screened for hypermethylated genes in endometrial tissues and cervical cytology specimens. They verified that BHLHE22, CDO1, and CELF4 showed favorable diagnostic efficacy, with CDO1 yielding the highest specificity among them [9]. Multiple subsequent studies assembled methylation panels encompassing these genes and validated their diagnostic utility for endometrial lesions using exfoliated cervical cells [10–13]. Although multi-gene methylation panels improve diagnostic accuracy via epigenetic signature combination, they incur higher costs and operational complexity, limiting their clinical scalability. Interestingly, combined CDO1 and CELF4 methylation achieves high diagnostic accuracy for distinguishing benign from malignant endometrial lesions in premenopausal women with AUB, and standalone CDO1 methylation still surpasses CELF4 alone [4]. Recently, Wang et al. recognized CDO1 and CELF4 methylation as powerful EC predictors in both premenopausal and postmenopausal women; CDO1 methylation provides superior specificity in postmenopausal patients and better sensitivity in premenopausal patients, consolidating its robust diagnostic advantages [14]. Taken together, CDO1 methylation plays a pivotal role in multi-gene combined methylation panels and has core diagnostic value in identifying EC among women with AUB. The use of exfoliated endometrial cells as a “molecular mirror” of the uterine cavity presents a minimally invasive sampling method. Nevertheless, the diagnostic sensitivity of single-gene methylation assays is frequently compromised by the biological heterogeneity of EC. EC is increasingly recognized as a disease driven by metabolic factors [15]. Robust epidemiological data establish a connection between systemic metabolic dysregulation—such as obesity, insulin resistance, and altered lipid profiles—and the pathogenesis of EC [16, 17]. Although molecular biomarkers and metabolic risk factors are typically considered separately in clinical practice, their integration has not been extensively investigated. Given these limitations, the concurrent advancement of machine learning (ML) in precision oncology presents a promising solution. ML methodologies have the potential to improve the performance of diagnostic models by integrating multidimensional clinical data. Unlike conventional statistical approaches that focus on identifying associations between individual variables and clinical outcomes, ML constructs comprehensive, individualized prediction models [18]. Moreover, interpretable ML techniques, such as SHapley Additive exPlanations (SHAP) analysis, facilitate the transparent quantification of feature contributions, thereby bridging the gap between computational capabilities and clinical interpretability.

In this study, we developed and evaluated an interpretable ML framework for EC triage in women with AUB, using an internal holdout cohort. By integrating CDO1 methylation levels with comprehensive clinical-metabolic profiles, we constructed a predictive model that significantly outperforms single-modality approaches. To ensure immediate clinical utility, we translated this model into a web-based risk calculator, offering a precision-medicine tool for personalized risk stratification and decision-making in the outpatient setting.

Methods

Study design and sample collection

This study employed a case-control design. The flowchart illustrating the study process is shown in Fig. 1. The study was structured into two primary phases: (1) The Discovery Phase, where local epigenetic signals (CDO1 methylation) and systemic metabolic features were extracted and integrated; and (2) The Validation Phase, where a validation set was utilized to test the model’s diagnostic generalizability. (3) The Practice Phase, where an internal holdout cohort was used to evaluate the clinical practicability of the model. Histopathological diagnosis served as the ground truth. To establish the biological rationale for our model, we performed immunohistochemical (IHC) scoring as a functional validation step. This was intended to verify that CDO1 promoter hypermethylation leads to protein silencing at the tissue level, thereby justifying its inclusion as a core molecular feature in our machine-learning framework. IHC scoring was performed based on staining intensity (scored 0–3) and the percentage of positive cells (0-100%), yielding a Histoscore (H-score) ranging from 0 to 300. The optimal cutoff value (H-score = 4.5) for defining low expression was determined using receiver operating characteristic (ROC) curve analysis; scores < 4.5 were classified as low expression. Two pathologists independently scored the slides in a blinded manner, with a Kappa value of 0.87 indicating substantial agreement.

Fig. 1.

Study participant enrollment flowchart

We consecutively enrolled female patients presenting with AUB at Fujian Maternity and Child Health Hospital between January 2022 and December 2023 who had an indication for endometrial biopsy. Two independent doctors performed retrospective case screening according to the standardized criteria and individually cross-checked the candidate medical records. Indications for biopsy were based on the Guidelines for Diagnosis and Treatment of Endometrial Cancer (2022 Edition) and the Chinese Guidelines for Menopause Management and Menopausal Hormone Therapy [19]. Specifically, biopsies were performed in patients with AUB whose TVS showed an endometrial thickness > 11 mm in premenopausal women or > 5 mm in postmenopausal women. The case group comprised patients with histopathologically confirmed EC (n = 100), while the control group consisted of women with normal endometrial histology or benign endometrial pathology (n = 167). Exclusion criteria were as follows: (1) patients who had undergone prior surgical treatment, radiotherapy, or chemotherapy; (2) samples with improper collection techniques, rendering effective sampling impossible; (3) repeated samples collected from the same patient at different time points; (4) samples failing to meet storage or testing requirements; and (5) patients who withdrew from the study. This study was approved by the Hospital Ethics Committee (Approval No.: 2022KYLLR03028), and written informed consent was obtained from all participants.

Bioinformatic prioritization and functional annotation of CDO1

To identify high-impact epigenetic drivers in EC, we used the UALCAN portal to analyze DNA methylation signatures in The Cancer Genome Atlas (TCGA-UCEC) dataset. Candidate genes were screened based on their methylation intensity and their association with clinicopathological parameters, including tumor stage, histologic subtype, and patient ethnicity. Methylation levels were quantified as Beta-values, and Z-score standardization was applied for sample normalization. Candidate genes were screened based on a significance threshold of FDR < 0.05 and a mean methylation difference (Delta Beta) > 0.2. The top 30 hypermethylated candidates were visualized via hierarchical clustering heatmaps.

To further elucidate the biological role of CDO1, we constructed a co-expression network and an interactome profile. Specifically, the top 100 CDO1-correlated genes were retrieved from GEPIA2, with the most significant interactors visualized using Cytoscape (v3.10.1). Protein-protein interaction (PPI) analysis was performed using the STRING database (minimum interaction score: 0.150). Finally, to explore the functional involvement of CDO1, Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis was executed via DAVID, with significance thresholds set at p < 0.05 and FDR < 0.05 to identify significantly enriched metabolic and oncogenic pathways. This bioinformatic pipeline provided the mechanistic foundation for linking local CDO1 alterations to systemic metabolic pathways.

Clinical specimen acquisition and exfoliated cell processing

Sample collection was performed by professionally trained gynecologists following a standardized procedure, as described in previous studies [20]. Briefly, patients were placed in the lithotomy position. After routine perineal and vaginal disinfection with povidone-iodine, a sterile vaginal speculum was inserted to expose the cervix, and uterine depth was measured using a uterine sound. Subsequently, sterile clinical specimens were obtained by professionally trained gynecologists using a standardized, minimally invasive protocol. Unlike traditional tissue biopsies, this approach targets the cellular debris and secretions within the uterine lumen, providing a representative molecular snapshot of the entire endometrium. Collected samples were immediately transferred into ThinPrep PreservCyt solution, transported on ice, and partitioned for molecular assays and cytological evaluation.

Data processing

Clinical information was collected, and endometrial exfoliated cells were obtained prior to the biopsy procedure for DNA methylation analysis. CDO1 promoter methylation was quantified via quantitative methylation-specific PCR (qMSP) using the following primer sequences: forward 5’-CCAACTCCCACTGCTTTCTG-3’ and reverse 5’-GCA GGT TCC GTA TGG CTG AT-3’ for CDO1; forward 5’-GGT GTG AAC CAT GAG AAG TAT GA-3’ and reverse 5’-GAG TCC TTC CAC GAT ACC AAA G-3’ for glyceraldehyde-3-phosphate dehydrogenase (GAPDH). The amplification protocol consisted of an initial denaturation at 95 °C for 10 min, followed by 50 cycles of 95 °C for 10 s and 60 °C for 40 s. Systematic analytical validation was conducted for this qMSP assay to confirm its robust detection performance. Linearity was verified across a methylation range of 0.1% to 100% using serially diluted fully methylated and unmethylated human genomic DNA standards, with a coefficient of determination (R²) of 0.996 and amplification efficiency of 96.2%, for assay reproducibility. Intra-assay coefficients of variation (CVs) for methylation quantification were all below 3% across samples with low, medium, and high methylation levels, while inter-run variability, evaluated across three independent experiments on separate days, yielded CVs consistently below 5%. The stability of methylation measurements was further confirmed for both 72-hour storage at 4 °C after bisulfite conversion and up to 5 freeze-thaw cycles, with CVs of 1.12% and 2.37%, respectively, showing no significant fluctuation in detection results. To ensure methodological reliability, CpGenome Universal Methylated DNA (Millipore, USA) and healthy volunteer peripheral blood DNA were utilized as positive and negative controls, respectively. All assays were performed in triplicate with a coefficient of variation (CV) for Ct values < 5%, and the limit of detection (LOD) was validated at 1 ng. Following bisulfite conversion of DNA, the ΔCt value (ΔCt = Ct_CDO1 − Ct_GAPDH) was calculated using GAPDH as the reference gene. The continuous ΔCt was utilized for the restricted cubic spline (RCS) analysis to evaluate the dose-response relationship between CDO1 methylation and EC risk, whereas the raw Ct of CDO1 was used for the qualitative classification of methylation status. The mean Ct value for CDO1 was determined as 31.85. Accordingly, samples with a CDO1 Ct value < 31.85 were classified as hypermethylated (high methylation level), while those with a CDO1 Ct value ≥ 31.85 were classified as hypomethylated (low methylation level). Unmethylated DNA and fully methylated controls served as negative and positive controls, respectively. All procedures strictly followed the manufacturer’s protocols [21].

Clinical data included age, menopausal status, medical history (e.g., type 2 diabetes) [22], body mass index (BMI), and endometrial thickness assessed by TVS within one month before hysteroscopy. Additionally, serum glucose and lipid profiles, including fasting blood glucose (FBG), triglycerides (TG), total cholesterol (TC), high-density lipoprotein (HDL), and low-density lipoprotein (LDL), were measured within one month before hysteroscopy. Complete peripheral blood inflammatory parameters included white blood cell count (WBC), neutrophil count (Neu), lymphocyte count (Lym), and platelet count (PLT).

Feature selection

To mitigate potential collinearity and ensure the parsimony of the final diagnostic model, we implemented a Least Absolute Shrinkage and Selection Operator (LASSO) regression in the Training set. This penalized regression approach was utilized to “shrink” the coefficients of redundant or non-contributory systemic metabolic features toward zero. The optimal regularization parameter (λ) was adjudicated via 10-fold cross-validation, targeting the minimum mean cross-validated error. Only features with non-zero coefficients were retained as the most predictive subset for subsequent machine learning pipeline integration. The final multivariate model features were selected via LASSO regularization rather than p-value thresholds.

Model development and internal partitioning

The primary clinical cohort (n = 267) was partitioned into a training set and an internal validation set using an approximate 7:3 ratio with a stratified, randomized partitioning approach. To maintain a stable class distribution across both subsets, 65% of the EC cases (65/100) and 73.05% of the controls (122/167) were assigned to the training set, while 35% of the cases (35/100) and 26.95% of the controls (45/167) constituted the internal validation set. To prevent data leakage, the data were split before any feature selection or preprocessing. To ensure compatibility among the heterogeneous datasets, we implemented a structured feature normalization pipeline. For all continuous variables, Z-score standardization was employed to transform features to a mean of 0 and a standard deviation of 1, thereby preventing variables with larger numeric scales from disproportionately influencing the model weights. Categorical variables were converted into binary dummy variables (0/1). Crucially, the scaling parameters (mean and standard deviation) were derived exclusively from the training set and subsequently applied unchanged to the validation set to prevent risks of data leakage. All feature selection (LASSO), hyperparameter optimization, and threshold determinations were strictly restricted to the training set. Four candidate algorithms, namely Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF), and eXtreme Gradient Boosting (XGBoost), were constructed in parallel. Specifically, the SVM classifier was implemented in R using the ‘e1071’ package with a Radial Basis Function (RBF) kernel, with the penalty parameter (Cost/C) set to 1.0 and Gamma set to 0.1. For the alternative algorithms, the RF model was configured with 500 trees, and the XGBoost model utilized a learning rate of 0.01. All hyperparameters were optimized using a grid search strategy combined with 10-fold cross-validation on the training set, which accounted for class imbalance by preserving the original target distribution within each fold. To mitigate overfitting, the modeling pipeline incorporated several safeguards: (i) LASSO-based feature reduction was performed before model construction to restrict the predictor set to a small number of non-redundant variables; (ii) all preprocessing parameters, feature selection, hyperparameter tuning, and threshold determination were confined to the training set; (iii) stratified 10-fold cross-validation was used during hyperparameter optimization; and (iv) final model performance was evaluated in both the validation set and the internal holdout cohort without further model refitting. Continuous variables were Z-score standardized using the internal ‘scale=TRUE’ parameter, with scaling factors derived exclusively from the training set to prevent data leakage. The cohort-specific optimal cutoffs were only calculated for analytical illustration and performance comparison across different datasets, rather than intended for clinical use.

To facilitate real-world clinical adoption, the finalized SVM-based predictive architecture was operationalized into a web-based risk calculator (https://phw1996.shinyapps.io/EC_risk/). This interface was deployed via an R-Shiny framework, allowing for the instantaneous estimation of EC risk probability upon input of patient-specific parameters. This digital tool aims to assist personalized decision-making in resource-limited settings where specialized bioinformatic support is unavailable.

Model performance evaluations and interpretability framework

The discriminative power of each model was adjudicated using a multi-dimensional metric suite, including the Area Under the Receiver Operating Characteristic Curve (AUC), sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). All optimal classification thresholds were pre-specified based on the maximum Youden index derived from the training set. Beyond statistical accuracy, we prioritized model transparency using SHAP analysis [23]. Rooted in cooperative game theory, SHAP analysis deconstructed the “black-box” outputs by quantifying the average marginal contribution of each local and systemic variable across all possible feature combinations [24]. Clinical net benefit was further evaluated via Decision Curve Analysis (DCA) and calibration plots to assess the agreement between predicted risk and observed prevalence. Additionally, the Brier score was used to quantify the agreement between predicted probabilities and observed outcomes.

The internal holdout cohort

We deployed the finalized architecture on an internal holdout cohort. We enrolled women with AUB and endometrial thickening between January 2024 and December 2024 who consented to the CDO1 methylation assay and performed endometrial biopsy with available pathological results as the internal holdout cohort (n = 160) from Fujian Maternal and Child Health Hospital including 12 EC patients totally, and were processed using the same preprocessing protocols and model configurations as the practice phase. This step ensured that the model’s predictive performance was not an artifact of the primary site’s data distribution. The specific composition of the validation set, training set, and internal holdout cohort is shown in Table S3.

Statistical analysis

Continuous variables are presented as median (interquartile range, IQR), and categorical variables as frequency (percentage). Group comparisons were performed using the Mann-Whitney U test or the Chi-square test. Univariate and multivariate logistic regression analyses were employed to identify risk factors for EC, and odds ratios (ORs) with corresponding 95% CIs were calculated. For the evaluation of diagnostic performance in the different sets, 95% CIs for sensitivity, specificity, PPV, and NPV were calculated using the Clopper-Pearson exact method to ensure robust estimation in a small-sample setting. Variable selection was performed using LASSO regression (10-fold cross-validation) via the “glmnet” package in R. All statistical analyses were conducted using R software (version 4.3.3). DeLong’s test was used to compare the AUC of different models based on the validation sets, to quantify the incremental discriminative value of combining CDO1 methylation with clinical variables. The finalized analytical dataset was complete with no missing clinical or molecular observations; therefore, no data imputation was required. A two-tailed p-value < 0.05 was considered statistically significant for all tests.

Results

Identification of CDO1 hypermethylation as a conserved early event in EC carcinogenesis

Initial methylomic profiling of the TCGA-UCEC dataset (438 EC tissues vs. 46 normal controls) identified a distinctive cluster of 30 hypermethylated genes (Fig. 2A). Among these candidates, CDO1 was prioritized due to its significant association with metabolic dysregulation. KEGG pathway enrichment analysis corroborated that CDO1-related differential signatures were predominantly concentrated within metabolic pathways, reinforcing its potential role as a metabolic-epigenetic driver in EC (Fig. 2B).

To provide a robust biological foundation for the subsequent machine-learning model, we first conducted a functional validation using IHC to assess CDO1 protein expression. Having identified CDO1 hypermethylation in the TCGA cohort, we next sought to biologically validate whether this epigenetic alteration functionally drives tumorigenesis in vitro. Bisulfite sequencing PCR (BSP) confirmed dense hypermethylation within the CDO1 promoter CpG islands across HEC-1 A and KLE cell lines (Fig. 2C). To establish a causal link between this epigenetic signature and gene silencing, we treated EC cells with the DNA methyltransferase inhibitor 5-Aza-2’-deoxycytidine (5-Aza-dC). Pharmacological demethylation triggered a significant restoration of CDO1 mRNA expression in both cell lines (Fig. 2D), confirming that promoter hypermethylation is the primary driver of CDO1 transcriptional repression in EC.

To further investigate the impact of CDO1 on EC, a retrospective study was conducted. Of the 300 women initially screened for AUB, 267 eligible subjects were enrolled after excluding cases with insufficient cellularity (n = 12), sample degradation (n = 3), or confounding hormonal therapy (n = 15). The final cohort comprised 100 patients with histologically confirmed EC and 167 benign controls. Baseline clinical profiles (Table 1) revealed that women in the EC group were older (median 51.0 vs. 42.0 years, p < 0.001) and had a greater metabolic burden, characterized by significantly higher BMI, elevated FBG, and a higher prevalence of diabetes (all p < 0.001). Importantly, lipid profiles (TG, TC, HDL, LDL) and the systemic immune-inflammation index (SII) showed no significant intergroup disparities, ensuring the comparability of these metabolic and inflammatory baselines.

Table 1.

Demographic and clinical characteristics of AUB patients

Variable	Total (n = 267)	EC Group (n = 100)	Control (n = 167)	P value
Age (years)	46.00 (38.00, 52.00)	51.00 (47.00, 58.00)	42.00 (34.50, 49.00)	< 0.001
BMI (kg/m2)	23.44 (20.96, 25.59)	24.62 (22.39, 26.49)	22.51 (20.55, 24.96)	< 0.001
Menopausal status (%)
premenopausal	197 (73.8%)	49 (49.0%)	148 (88.6%)
postmenopausal	70 (26.2%)	51 (51.0%)	19 (11.4%)	< 0.001
Diabetes (%)
No	252 (94.4%)	88 (88.0%)	164 (98.2%)
Yes	15 (5.6%)	12 (12.0%)	3 (1.8%)	< 0.001
SII	550.61 (379.91, 768.38)	553.27 (357.15, 963.23)	540.84 (392.60, 724.41)	0.524
FBG (mmol/L)	5.05 (4.72, 5.56)	5.37 (4.90, 5.89)	4.93 (4.63, 5.28)	< 0.001
TG (mmol/L)	1.29 (0.92, 1.79)	1.36 (0.97, 1.86)	1.27 (0.88, 1.69)	0.317
TC (mmol/L)	4.99 (4.35, 5.74)	5.02 (4.30, 5.80)	4.95 (4.38, 5.64)	0.823
HDL (mmol/L)	1.26 (1.07, 1.49)	1.19 (1.04, 1.48)	1.30 (1.10, 1.50)	0.083
LDL (mmol/L)	2.82 (2.27, 3.31)	2.98 (2.31, 3.48)	2.73 (2.25, 3.24)	0.212

Continuous variables are presented as median (Q1, Q3); categorical variables as n (%). SII is calculated as platelet count × neutrophil count / lymphocyte count (×10⁹/L)

Abbreviations: BMI Body Mass Index, SII Systemic immune inflammation index, FBG Fast blood glucose, TG Triglyceride, TC total Cholesterol, HDL High-density lipoprotein, LDL Low-density lipoprotein

Analysis of endometrial exfoliated cells revealed a profound epigenetic divergence between the two groups. The raw Ct values for CDO1 methylation were markedly lower in the EC group compared to the benign controls (median 29.40 [IQR: 27.45–31.33] vs. 35.34 [IQR: 32.45–40.00], p < 0.001), signifying a high density of promoter hypermethylation in the presence of malignancy (Fig. 2E).

Multivariate logistic regression was employed to adjudicate the independent predictive power of this epigenetic signature. The AUC of CDO1 (Ct value) in endometrial shed cells for EC diagnosis was 0.87 (95% CI: 0.83–0.92, p < 0.001), with an optimal cutoff value of Ct = 31.85 (Fig. 2F). This cutoff yielded a sensitivity of 0.75, a specificity of 0.86, a positive predictive value (PPV) of 0.57, and a negative predictive value (NPV) of 0.93 for EC diagnosis. CDO1 hypermethylation (defined by a Ct < 31.85) emerged as the most potent independent risk factor for EC (OR = 12.76, 95% CI: 6.31–25.80, p < 0.001), substantially surpassing the effect sizes of traditional clinical indicators such as menopausal status (OR = 3.03, p = 0.034) and diabetes (p = 0.064; Table 2). To further explore the biological gradient of this association, we performed an RCS analysis. The results confirmed a non-linear dose-response relationship between epigenetic silencing and oncogenic risk (P non−linear = 0.04). The probability of EC escalated sharply as the Ct value decreased (indicating increasing methylation density), suggesting a critical threshold.

Fig. 2.

CDO1 Methylation Analysis in Endometrial Cancer Tissues from TCGA Database. (A) Heatmap of the top 30 hypermethylated genes in endometrial cancer (EC) compared to normal endometrial tissues. (B) KEGG pathway enrichment analysis showing CDO1-associated metabolic pathways. (C) Bisulfite sequencing (BSP) analysis of CpG sites within the CDO1 promoter region in HEC-1 A and KLE cell lines. Each row represents a single cloned allele, and each circle represents a CpG site. Filled and open circles indicate methylated and unmethylated CpGs, respectively. (D) CDO1 mRNA expression levels in HEC-1 A and KLE cells after treatment with 5 µM 5-Aza-2’-deoxycytidine (5-Aza-dC) or vehicle control (DMSO) (P < 0.0001). (E) Comparison of CDO1 methylation in endometrial exfoliated cells between EC group and control group

Table 2.

Univariate and multivariate Logistic regression analysis for predicting EC

Variable	Univariate		Multivariate
	OR (95%CI)	p value	OR (95%CI)	p value
Age (years)	1.12 (1.08, 1.15)	< 0.001	1.05 (1.00, 1.10)	0.052
BMI(kg/m2)	1.12 (1.04, 1.20)	0.003	1.00 (0.90, 1.11)	0.953
Menopausal status (postmenopausal vs. premenopausal)	8.05 (4.34, 14.94)	< 0.001	3.03 (1.09, 8.44)	0.034
Diabetes (Yes vs. No)	7.41 (2.04, 26.95)	0.002	5.43 (0.91, 32.50)	0.064
SII	1.00 (1.00, 1.00)	0.007	1.00 (1.00, 1.00)	0.190
FBG (mmol/L)	1.24 (1.01, 1.50)	0.038	0.975 (0.78, 1.23)	0.833
TG (mmol/L)	1.16 (0.89, 1.49)	0.267	-	-
TC (mmol/L)	1.01 (0.80, 1.27)	0.953	-	-
HDL (mmol/L)	0.58 (0.25, 1.33)	0.196	-	-
LDL (mmol/L)	1.07 (0.82, 1.39)	0.618	-	-
CDO1 methylation			-	-
Low (≥ 31.85)			-	-
High (< 31.85)	17.60 (9.44, 32.83)	< 0.001	12.76 (6.31, 25.80)	< 0.001

Abbreviations: BMI Body Mass Index, SII Systemic immune inflammation index, FBG Fasting blood glucose, TG Triglyceride, TC total Cholesterol, HDL High-density lipoprotein, LDL Low-density lipoprotein, Ct Cycle threshold

The final multivariate model features were selected via LASSO regularization rather than p-value thresholds

Refinement of clinical feature selection using LASSO regression for EC analysis

A total of 267 patients were included in this study. Table S1 presents basic patient information. Significant differences were observed among the groups across multiple key indicators, including patient age, CDO1 methylation status, BMI, menopausal status, history of diabetes, and inflammatory indices, as well as indicators related to glucose and lipid metabolism, such as FBG, TG, TC, HDL, and LDL.

By applying LASSO regression analysis, which is renowned for its variable selection and regularization capabilities, the baseline dataset was effectively optimized. Based on this LASSO regression analysis, four key predictors were identified: CDO1 methylation (coefficient = 1.202), menopausal status (0.834), diabetes (0.290), and age (0.248) (Fig. 3A-B). These were incorporated into four ML models (LR, SVM, RF, XGB).

Fig. 3.

Machine Learning Model Construction. (A-B) Feature selection using LASSO regression analysis in the Training set. (A) LASSO coefficient profiles of the clinical variables. (B) Ten-fold cross-validation for tuning parameter (lambda) selection in the LASSO model. (C-H) Performance evaluation of SVM machine learning models in the training and validation sets: (C) ROC curves in the training set. (D) DCA curves in the training set. (E) calibration curve in the training set. (F) ROC curves in the validation set. (G) DCA curves in the validation set. (H) Calibration curve in the validation set

Construction and validation of the machine learning models

We conducted a comprehensive analysis of four machine learning algorithms, comparing their performance based on accuracy, sensitivity, specificity, and AUC. In the training set, the SVM model achieved an accuracy of 0.872. ROC analysis yielded an AUC of 0.926 at a cutoff value of 0.121, reflecting the model’s strong ability to distinguish between normal endometrium and EC. DCA indicated a superior net benefit within a threshold probability range of 0.03 to 0.83, supporting favorable clinical utility. The calibration curve showed moderate alignment with the ideal diagonal, suggesting some divergence between predicted and actual probabilities. These findings were corroborated in the validation set, where the model attained an accuracy of 0.762 and an AUC of 0.775 (cutoff: 0.649), confirming its discriminative capacity. DCA further demonstrated clinical applicability across threshold probabilities ranging from 0.09 to 0.64, although the calibration curve exhibited a suboptimal fit, indicating notable miscalibration (Fig. 3C-H).

Comparing and evaluating the machine learning models

The LR model showed consistent discriminative ability with AUCs of 0.925 and 0.831 in the training and validation sets, respectively, yet its calibration was only moderate and generally inferior to that of the SVM model. And the specificity of the LR model was 0.667, which was significantly lower than the specificity of 0.733 for the SVM model. The RF model achieved perfect performance in the training set (accuracy = 1.00, AUC = 1.00) but suffered from marked overfitting, as seen in its substantial drop in validation accuracy (0.725) and AUC (0.740), along with generally poor calibration—highlighting its instability compared to the more generalized SVM. The XGB model exhibited strong diagnostic performance (training AUC = 0.956, validation AUC = 0.775) and favorable calibration in the validation set, yet its clinical net benefit range was narrower than that of the SVM (Table 3).

Table 3.

Prediction results for multiple machine learning algorithm models in the test sets

Parameter	Training set (n = 187)				Validation set (n = 80)
	LR	SVM	RF	XGB	LR	SVM	RF	XGB
Accuracy	0.861	0.872	1.00	0.914	0.775	0.762	0.725	0.750
Kappa	0.712	0.732	1.00	0.702	0.560	0.525	0.468	0.511
Sensitivity	0.938	0.938	1.00	0.886	0.914	0.800	0.914	0.886
Specificity	0.820	0.836	1.00	0.926	0.667	0.733	0.578	0.644
PPV	0.735	0.753	1.00	0.866	0.681	0.700	0.627	0.660
NPV	0.962	0.962	1.00	0.942	0.909	0.825	0.900	0.879
Recall score	0.938	0.938	1.00	0.892	0.914	0.800	0.914	0.886
F1 score	0.824	0.835	1.00	0.879	0.780	0.747	0.744	0.756
AUC	0.925	0.926	1.00	0.956	0.831	0.775	0.740	0.775
Cutoff	0.221	0.121	0.440	0.512	0.506	0.649	0.225	0.444

Abbreviations: PPV positive predictive value, NPV negative predictive value, LR Logistic regression, SVM Support Vector Machine, RF random forest, XGB XGBoost

However, LR, RF, and XGB had their own characteristics and limitations. Statistical comparison of ROC curves was performed using the DeLong test. The main difference was in the comparison between the SVM and the LR models. The difference in AUC was − 0.057, with the adjusted p-value of 0.290 (> 0.05). Although the AUC of SVM was lower than that of LR, its 95% CI contained 0, indicating no statistically significant difference (Table S4). We cannot conclude that the LR model is significantly superior to the SVM model. In the management of AUB populations in which 85%–95% of patients are ultimately diagnosed with benign conditions, our primary objective was to maximize specificity, provided that a clinical safety threshold for sensitivity (≥ 80%) was maintained. Given that this model is intended for risk stratification of patients with EC, we prioritized the model with higher specificity. The SVM model’s higher specificity than LR model (0.733 vs. 0.667) and stable net benefit in DCA curves indicate its superior ability to reduce unnecessary invasive procedures.

Based on the above analysis, the SVM model demonstrated the highest specificity among all models. Therefore, SHAP analysis provides individualized feature attribution, showing which biomarkers and clinical factors drive each patient’s risk prediction. SHAP analysis was performed on the SVM model, which. confirmed that CDO1 gene methylation was the most influential feature in predicting EC, followed by menopausal status, diabetes and age (Fig. 4A-B). SHAP dependence analyses demonstrated that CDO1 hypermethylation, postmenopausal, diabetes and advanced age independently increased the predicted risk of EC. Samples carrying high CDO1 methylation were concentrated in women with conventional clinical risk factors (elderly, postmenopausal, diabetic individuals), revealing synergistic pro-cancer effects between CDO1 hypermethylation and clinical risk factors (Fig. S2).

Fig. 4.

Machine Learning Model Evaluation and Internal Holdout Cohort. (A) SHAP summary plot showing the overall feature importance for the SVM model. (B) SHAP force plot illustrating the interpretation of an individual prediction from the SVM model. (C) ROC curve of SVM model in the internal holdout cohort. (D) Calibration curve of the SVM model in the internal holdout cohort. The dotted line represents the ideal fit where predictions perfectly match observations. The solid line represents the performance of the current model. (E) DCA of the SVM model in the internal holdout cohort, showing the net clinical benefit across different threshold probabilities

Validation for assessing model generalizability

To objectively evaluate the generalizability of the SVM-based triage model, a single-center holdout was incorporated. This cohort consisted of 160 patients presenting with abnormal uterine bleeding, among whom 12 were pathologically confirmed with EC and 148 were diagnosed with benign endometrial conditions. Endometrial cytology samples were collected from all patients prior to biopsy to assess CDO1 methylation status. Clinical variables, including age, menopausal status and history of diabetes, were also documented.

After applying the same preprocessing procedures and model configurations used in the training set, the SVM model demonstrated favorable discriminative ability in the internal holdout cohort. The area under the ROC curve was 0.904 (95% CI: 0.838–0.969). Detailed performance metrics showed a sensitivity of 0.833 (95% CI: 0.516–0.979), a specificity of 0.905 (95% CI: 0.846–0.947), a PPV of 0.417 (95% CI: 0.221–0.634), and an NPV of 0.985 (95% CI: 0.948–0.998) (Fig. 4C). In contrast to its discriminative performance, calibration assessment indicated non-negligible miscalibration in this low-prevalence holdout cohort. After applying the probability-calibration mapping derived from the internal validation set, without any refitting, to the holdout cohort, the Brier score was 0.169. The calibration curve showed a deviation between predicted and observed probabilities, indicating that the absolute risk estimates should be interpreted cautiously and require further recalibration in larger prospective cohorts (Fig. 4D). Decision curve analysis suggested potential net clinical benefit across a range of threshold probabilities, supporting the preliminary applicability of the model for risk stratification rather than definitive absolute risk prediction (Fig. 4E). According to current clinical guidelines, final enrolled patients suspected of EC (n = 160) would receive biopsy referrals. Using our model’s cutoff of 0.647, 24 patients were identified as candidates for biopsy referral, and 2 cases were misclassified as benign endometrium. Although 2 EC cases were missed (16.7% missed-diagnosis rate), the model achieves an 85% reduction in unnecessary biopsies, delivering significant cost-saving and clinical efficiency benefits for real-world AUB women triage.

Development of a web-based clinical translation tool

To provide a more convenient, dynamic, and personalized risk assessment, we deployed the optimal-performing SVM model as a publicly accessible online web tool (available at https://phw1996.shinyapps.io/EC_risk/). This tool enables users to input or select key clinical and molecular parameters—such as the CDO1 methylation, age, menopausal status, and history of diabetes. The fixed threshold of 0.647 (derived from the internal holdout cohort) is the cutoff embedded and deployed in the web-based risk calculator for all clinical predictions. The backend model then performs real-time calculations and returns an individualized probability of EC risk. The resulting risk output can directly support clinical decision-making, aiding physicians in stratifying patients with AUB by risk. This tool helps identify low-risk patients who may avoid unnecessary invasive biopsies while prompting more definitive pathological evaluation for those at high risk. The implementation of this tool marks a transition in our research from algorithm development to clinical application. It offers an immediately usable, practical means for precise, non-invasive risk assessment of EC.

Discussion

Instead of relying solely on a single biological marker, this study evaluates an integrative diagnostic framework that combines localized molecular alterations with systemic clinical and metabolic risk factors. First, our findings support an association in this cohort between CDO1 promoter hypermethylation, along with clinical indicators of glucose and lipid metabolism, and the presence of EC in women presenting with AUB. Second, we have successfully developed and validated a multimodal machine learning prediction model that innovatively combines this molecular marker with essential clinical and metabolic indicators. Compared with conventional unidimensional diagnostic approaches, this integrated model, particularly the SVM algorithm, demonstrates favorable discriminative power and consistent net clinical benefit in both the validation set and the internal holdout cohort. While further optimization is required to refine its calibration accuracy across diverse clinical populations, the model currently enables reliable, personalized risk assessment for EC triage. Finally, to facilitate the translation of research findings into clinical practice, we have operationalized the optimal SVM model into an interactive online risk calculator. This digital deployment significantly enhances the model’s accessibility and generalizability, providing frontline and primary care physicians with an immediate, non-invasive decision-support tool for EC risk stratification.

This study systematically investigated the role of CDO1 gene methylation in the pathogenesis and progression of EC and evaluated its potential as a diagnostic biomarker. By integrating analyses of the TCGA database, ex vivo functional experiments, and a clinical case-control study, we demonstrated across multiple levels of evidence that CDO1 promoter hypermethylation is a frequent event in EC. Critically, CDO1 hypermethylation in exfoliated cells emerged as the most potent independent risk factor for EC (OR = 12.76, 95% CI: 6.31–25.80, P < 0.001). This epigenetic signature demonstrates robust diagnostic efficacy in our cohort, reinforcing its potential as a core molecular feature for non-invasive risk stratification. This finding holds significant clinical implications, particularly for high-risk patients with abnormal TVS findings who decline invasive biopsies. The current paradigm for diagnosing EC in women presenting with AUB relies heavily on imaging and invasive biopsy, methods hampered by suboptimal specificity or inherent invasiveness [8, 25]. Although transvaginal ultrasound demonstrates a sensitivity and negative predictive value as high as 96.2% and 99.3%, respectively, its specificity is relatively low, necessitating additional examinations to exclude endometrial malignancies. However, outpatient endometrial sampling demonstrates high diagnostic accuracy for EC (sensitivity 90%–100%, specificity 98%–100%); its failure rate is as high as 42% due to insufficient sample volume, which limits its reliability as a first-line screening tool [26]. The selection of the SVM-based model involved a critical trade-off between sensitivity and specificity. While the LR model exhibited higher peak sensitivity, the SVM model (sensitivity 80.0%, specificity 73.3% in the validation set) aligns more closely with the goals of secondary triage in AUB populations. By maximizing specificity, this approach directly addresses the clinical need to minimize unnecessary invasive biopsies in the 85%–95% of symptomatic women who are ultimately diagnosed with benign conditions, thereby enhancing the efficiency of outpatient EC risk stratification. Although the sensitivity and specificity of the present model are slightly lower, its non-invasive nature offers a novel diagnostic approach for patients who are unwilling or unable to undergo invasive procedures. Furthermore, traditional serum markers such as CA125 and HE4 have limited clinical utility due to their insufficient sensitivity and specificity [6]. This diagnostic dilemma underscores the urgent need for highly accurate, non-invasive tools. Most existing research relies on tissue samples, and tissue biopsy is invasive, making it unsuitable for large-scale screening. Wever et al. detected cellular DNA methylation in urine, cervicovaginal self-samples, and cervical smears, finding significantly higher levels of CDO1, ADCYAP1, and BHLHE22 methylation across all three sample types from 103 EC patients compared with normal controls [27]. Wang et al. further demonstrated that detecting CDO1 hypermethylation in cervical cytology specimens effectively distinguishes between benign and malignant endometrial conditions, suggesting that CDO1 methylation in cervical cells is a promising biomarker for EC [11]. Although non-invasive testing using cervical exfoliated cells and urine samples. However, their sensitivity for detecting endometrial lesions remains relatively low. In contrast, endometrial exfoliated cells originate directly from the endometrium and may more accurately reflect features of early-stage lesions. The minimally invasive nature of endometrial brushing—offering substantially better tolerability than dilatation and curettage—combined with the molecular stability of methylation markers, circumvents limitations of tissue-based diagnostics. Unlike histopathology, methylation analysis does not require intact tumor cells and can be measured robustly from tumor-derived DNA fragments, enhancing its utility in resource-constrained settings where pathological expertise may be limited.

EC is well recognized as a metabolism-driven malignancy with obesity [28], insulin resistance [29], and dyslipidemia [30] as major risk factors. As we know, CDO1 is the rate-limiting enzyme for cysteine breakdown and endogenous taurine synthesis. Abundant clinical evidence has confirmed the correlation between CDO1 methylation and EC onset. The biological plausibility of our findings is underscored by CDO1’s role as a tumor suppressor that regulates cysteine metabolism, oxidative stress response, and ferroptosis [31, 32]. Via a taurine-mediated mechanism, CDO1 ameliorates metabolic disorders by enhancing lipid metabolism and insulin sensitivity [33]. Mechanistically, adipose-tissue-specific knockout and overexpression mouse models further confirmed that CDO1 activates lipolysis to reverse high-energy diet-induced obesity and associated metabolic complications in vivo [34]. Hypermethylation-induced silencing of this gene likely promotes carcinogenesis by reducing cellular susceptibility to oxidative damage and altering the tumor microenvironment. Our data align with TCGA profiles showing a strong inverse correlation between CDO1 methylation and expression in EC tissues (r = -0.68, P < 0.001), confirming the functional impact of this epigenetic alteration.

Machine learning integration further elevated the clinical utility of CDO1 methylation. Among evaluated algorithms, the SVM model demonstrated optimal performance in the validation set, achieving 73.3% specificity and significant net clinical benefit across probability thresholds of 0.09–0.64. While random forest showed perfect training performance (AUC = 1.00), its validation AUC (0.74) indicated marked overfitting. In contrast, the SVM model did not exhibit perfect training-set performance and was selected based on a pre-specified clinical utility criterion that prioritized higher specificity while maintaining a sensitivity threshold of ≥ 80%.To mitigate overfitting, we used a parsimonious four-variable predictor set selected by LASSO, restricted all preprocessing, feature selection, hyperparameter tuning, and threshold determination to the training set, and evaluated the finalized model in both the validation set and the internal holdout cohort without further refitting. Moreover, SHAP interpretability and SHAP dependence analyses confirmed CDO1 methylation as the dominant predictive feature, followed by menopausal status, highlighting its centrality to EC pathogenesis. For one thing, the molecular signature and clinical covariates provide complementary predictive signals rather than redundant information, which widens the predictive score gap between high-risk and low-risk populations and improves discrimination performance. For another, this synergistic pattern endows the model with biologically interpretable outputs, consistent with the underlying pathogenesis of EC, and strengthens its translational value for clinical triage. In addition, the scattered distribution of dots indicates only trend-level enrichment without severe multicollinearity, ensuring stable model performance.

We compared our SVM model with existing non-invasive tools based on published data, as a formal head-to-head analysis was not possible in this cohort. For transvaginal ultrasound, a 2024 meta-analysis reported a pooled sensitivity of 76.6% (95% CI: 70.9–81.4%) and a specificity of 87.4% (95% CI: 80.6–92%) [35]. For endometrial cytology, a 2023 meta-analysis (26 studies) showed a pooled sensitivity of 84% (95% CI: 83–86%), specificity of 98% (95% CI: 98–98%), and summary AUC of 0.967 [36]. For a CDO1/CELF4 methylation panel in cervical scrapings, a 2025 prospective study (n = 675) reported a sensitivity of 84.1%, specificity of 98.8%, and AUC of 0.92 [12]. In our validation set, the SVM model achieves a higher specificity of 0.733 and a sensitivity of 0.800, while the LR model achieves a higher sensitivity of 0.914 but a lower specificity of 0.667. The low specificity of the LR model would substantially raise the false-positive rate and lead to numerous unnecessary biopsies among benign cases. Given that most enrolled participants are outpatients presenting with AUB and ambiguous imaging findings, excessive false positives would greatly increase patients’ psychological burden, medical costs, and the overall workload of clinical departments. Notably, in our internal holdout cohort, the SVM model achieved a sensitivity of 83.3%, specificity of 90.5%, and AUC of 90.4%. The model showed 80% sensitivity in the primary validation set, and exceeded 83% in the additional internal holdout cohort, demonstrating improved diagnostic efficacy in real-world clinical populations. These results offer an important reference for the clinical translation of this model. In fact, previous studies have reported that the false-negative rate of endometrial biopsy is approximately 10% [37]. Our model achieves a marked improvement in specificity without substantially increasing the number of missed cancer cases.

These comparisons are indirect and should be interpreted with caution due to differences in study populations and designs. The lower specificity of our model relative to multi-gene methylation panels or cytology may reflect the inclusion of clinical variables (age, menopausal status, diabetes) rather than a purely molecular test, as well as the wide confidence intervals due to the small number of EC cases (n = 12) in the holdout cohort. Nevertheless, our model offers an integrative approach—combining a local epigenetic marker from exfoliated endometrial cells with systemic metabolic indicators—and is available as a public online risk calculator. Future head-to-head studies are needed to directly compare its performance with existing non-invasive tools.

This study establishes CDO1 hypermethylation as a key event in endometrial carcinogenesis and a powerful diagnostic biomarker. The non-invasive detection of CDO1 methylation in exfoliated endometrial cells offers a viable alternative to tissue biopsy, and the developed predictive model provides a potentially useful tool for clinical risk assessment that warrants further validation. These findings suggest that CDO1 methylation may serve as a promising biomarker for early detection of EC. Although these safeguards reduced the risk of overfitting, the possibility of residual overfitting cannot be fully excluded because of the limited sample size and single-center nature of the development cohort. Therefore, the SVM model should be regarded as a preliminary triage model, and its generalizability requires further confirmation in larger prospective multi-center cohorts. To facilitate the practical application of these findings, we developed a web-based tool (https://phw1996.shinyapps.io/EC_risk/) that integrates CDO1 methylation with readily available clinical variables to streamline risk stratification.

Limitation

As a single-center case-control study focusing on high-risk patients with AUB, selection bias was inevitable including demographics, lifestyle, race and ethnicities etc., limiting the generalizability of our results. Limited advanced cases restrict the analyses of EC invasiveness. Variable patient cohorts and laboratory environments may cause model drift. While we confirmed the relationship between CDO1 hypermethylation and EC, their potential mechanisms remain unclear and require further exploration. Besides, high costs and strict pre-test requirements limit its application at primary medical institutions. Multicenter prospective studies with diverse populations are needed to recalibrate the model, verify its stability, broaden its applicability, and explore more applications.

Conclusion

In this study, we verified that promoter hypermethylation silences CDO1 transcription in EC. Rescue of CDO1 expression effectively inhibits malignant phenotypes in EC cells, suggesting that epigenetic downregulation of CDO1 contributes to EC development. CDO1 hypermethylation in exfoliated endometrial cells is independently associated with early EC risk, and its detection offers better diagnostic sensitivity and specificity, as well as non-invasiveness, relative to tissue immunohistochemistry. The SVM diagnostic model incorporating CDO1 methylation and clinical metabolic features exhibited promising performance for personalized risk evaluation and triage in high-risk AUB populations. This tool aids clinical decision-making through relative risk stratification, but external recalibration is essential before it can be used to assess absolute clinical risk. Overall, combining molecular markers from exfoliated cells and systemic metabolic signatures using machine learning is a promising non-invasive strategy for EC triage, though it still needs to be validated in larger prospective cohorts.

Abbreviations

AUB

Abnormal Uterine Bleeding

CDO1

Cysteine Dioxygenase Type 1

Ct

Cycle Threshold

EC

Endometrial Cancer

FIGO

International Federation of Gynecology and Obstetrics

Authors’ contributions

Conceived and designed the analysis: Peng X. Collected the data: Mao X, Shi Y, Zhao J, Xue H, Wu X, Sun P, Chen G. Contributed data or analysis tools: Lin X, Shi Y, Zhao J, Wu X. Performed the analysis: Mao X, Lin X, Shi Y, Xue H. Wrote the paper: Mao X, Lin X. Revised the paper: Mao X, Peng X, Sun P, Chen G.

Funding

This study was funded by the National Nature Science Foundation of China (82203739) the National Key Research and Development Program of China (2022YFC2704303), and Major scientific research projects for Fujian Province health medicine innovation project (2024CXB008). The sponsors have no involvement in study design, data collection, analysis, and interpretation or in writing of the manuscript.

Data availability

The original data can be directed to the corresponding authors on reasonable request.

Declarations

Ethics approval and consent to participate

The study was reviewed and approved by the Ethics Committee of Fujian Maternity and Child Health Hospital (Approval No. 2022KYLLR03028). The participants provided their written informed consent to participate in this study. This study was performed in accordance with the Helsinki Declaration.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.