A total of 202 patients diagnosed with EC between January 2017 and January 2020 were retrospectively enrolled and stratified into a good prognosis group (n = 129) and a poor prognosis group (n = 73) based on clinical outcomes. No significant differences in baseline demographic or clinical characteristics were observed between the groups (P > 0.05; Table 1). However, there was a difference in the distribution of each type between the good prognosis group and the poor prognosis group (P = 0.013; Table 1), and the CNH type accounted for a higher proportion in the poor prognosis group (57.53% vs. 35.66%) This study was approved by the Ethics Committee of The Affiliated Taizhou People’s Hospital of Nanjing Medical University, and written informed consent was obtained from all participants or their families.

Inclusion criteria comprised: (1) diagnosis of EC confirmed by postoperative pathological examination in accordance with the guidelines from the Diagnosis and Management of Endometrial Cancer (Braun et al., 2016); (2) availability of comprehensive clinical records; (3) stringent quality control (QC) of analytical data; (4) complete gene expression profiles covering all targets of interest without significant missing values or outliers; and (5) signed informed consent forms. Exclusion criteria included: (1) concurrent malignancies; (2) prior neoadjuvant therapies (e.g., chemotherapy, radiotherapy) that might alter tumor gene expression profiles; (3) suboptimal sample quality; and (4) incomplete follow-up data.

The Gene Expression Omnibus (GEO) dataset GSE120490, comprising EC patient samples, was downloaded from the National Center for Biotechnology Information (NCBI) using the GEOquery package in R. Raw data were preprocessed and normalized, followed by gene identifier (ID) matching and differential expression analysis using the stringr, limma, and tidyverse packages. Gene Ontology (GO) enrichment analysis was performed with the clusterProfiler, org. Hs.e.g.,.db and enrichplot packages. Unsupervised dimensionality reduction using PCA was performed to verify the intrinsic expression differences between tumor tissue and normal tissue, while excluding technical batch effects. To further quantify the association between gene expression and sample grouping, a supervised dimensionality reduction model was constructed using partial least squares discriminant analysis (PLS-DA) (Mallardo et al., 2024). Using ‘tumor vs. normal’ or ‘good prognosis vs. poor prognosis’ as the dependent variable and DEGs expression level as the independent variable; Calculate R² Y (model interpretability) and Q² (predictive ability) through 10 fold cross validation to verify the reliability of the model. To refine biomarker candidates, feature selection was conducted via Least Absolute Shrinkage and Selection Operator (LASSO) regression, minimizing overfitting while identifying genes with the strongest prognostic relevance.

Lavage fluid in the uterine cavity was collected from patients and labeled for subsequent processing. Total cellular RNA was extracted using an RNA extraction kit (Tiangen Biotech Co., Ltd., China) and reverse-transcribed into complementary DNA (cDNA) with a reverse transcription (RT) kit. Primers were designed for target genes identified through prior screening, and quantitative polymerase chain reaction (qPCR) was performed using the synthesized cDNA as the template. Glyceraldehyde-3-phosphate dehydrogenase (GAPDH) served as the endogenous reference gene, while peripheral blood mononuclear cells (PBMCs) from healthy individuals undergoing routine physical examinations in our hospital were used as controls. Relative gene expression levels ≥ 2-fold compared to controls were defined as significant upregulation. Primer sequences are listed in Table 2. Genes exhibiting significant upregulation or downregulation (≥2-fold change) relative to healthy controls were assigned a binary value of 1, while others were coded as 0. Postoperative follow-up was conducted for 5 years, with patients experiencing recurrence or mortality classified into the poor prognosis group, and the remaining cases categorized into the good prognosis group.

Peripheral venous blood samples (6 mL) were routinely collected from fasting patients preoperatively and centrifuged to isolate serum for subsequent analysis. Complete blood count parameters, including hemoglobin (Hb), white blood cell count (WBC), red blood cell count (RBC), and platelet count (PLT), were measured using an automated hematology analyzer. Serum levels of carbohydrate antigen 125 (CA125), carbohydrate antigen 199 (CA199), and interleukin-9 (IL-9) were quantified via chemiluminescence immunoassay (CLIA) following standardized protocols.

Data processing and analyses were performed using SPSS Statistics 27.0 and R software version 4.4.3. Normally distributed continuous variables were expressed as mean ± standard deviation ( x ¯ ± s), while categorical variables were presented as n (%). Intergroup comparisons were conducted using independent samples t-tests (for continuous variables) and chi-square tests (for categorical variables). Multivariate binary logistic regression analysis was employed to identify independent risk factors, using variance inflation factor for collinearity screening. For gene feature selection, LASSO regression was implemented via the glmnet package in R. Kaplan-Meier (K-M) survival curves and Cox proportional hazards regression forest plots were generated using the forestplot, survminer, and survival packages in R. Random forest models were constructed with the randomForestSRC, ggRandomForests, pdp, and GGally packages to evaluate variable importance and partial dependence. Statistical significance was defined as a P < 0.05.

As shown in Figure 1, box plots (Figure 1A) and principal component analysis (PCA) plots of the GEO dataset GSE120490 (n = 145 samples) revealed no significant batch effects or inter-sample variability (P > 0.05). PCA analysis showed that the GSE120490 samples were clustered into two main clusters (Figure 1B), and the dataset included tumor tissue and normal tissue data. Further analysis confirmed that this grouping was driven by sample type (PCA1 contribution rate of 32.7%), rather than technical batch effects. Meanwhile, after correcting for batch effects using the SVA package in R language, the grouping trend remained significant, indicating biological differences. This is consistent with the inherent differences in gene expression profiles between tumor tissue and normal tissue. The PLS-DA (Figure 1C) score plot shows complete separation of tumor and normal tissue samples on the first principal component (explanatory power 41.2%) (R² Y = 0.89, Q² = 0.82), indicating a strong explanatory power of gene expression for sample types. Volcano plot (Figure 2A) was used to screen DEGs based on the statistical difference of gene expression level only, which was used to distinguish the abnormal expression genes in tumor tissues from normal tissues, and did not involve the association analysis with clinical prognosis outcome; The core genes related to prognosis were gradually screened through lasso regression (Section 2.2) and multiple regression (Section 2.5). Differential expression analysis (adjusted P < 0.05) identified 178 significantly upregulated genes, 200 downregulated genes, and 23,142 non-DEGs (Figure 2B). GO enrichment analysis demonstrated distinct functional annotations across biological processes (BP), molecular functions (MF), and cellular components (CC) (Figure 3).

The 378 significant DEGs were subjected to LASSO regression to identify key predictive features. During variable selection, the penalty coefficient (λ) was systematically compressed across all 378 initial predictors. Optimal λ (λ = 0.013) was determined via cross-validation, minimizing the mean squared error while balancing model parsimony and goodness-of-fit. This process yielded a refined predictive model incorporating eight genes: Erb-B2, CCND1, PIK3CA, VEGF, KIT, PTEN, E-cadherin, and p53 (Figure 4).

Based on qPCR results, gene expression changes were assessed by comparing each group’s data with healthy controls. Independent samples t-tests revealed significant differences in the expression of Erb-B2, PIK3CA, VEGF, KIT, PTEN, E-cadherin, and p53 (P < 0.05), whereas CCND1 showed no significant variation (P > 0.05). Similarly, tumor markers CA125 and CA199, RBC, and the inflammatory cytokine IL-9 exhibited statistically significant intergroup differences (Table 3).

To explore the associations between molecular subtypes, clinical indicators, gene expression, and prognosis, subgroup analyses were performed based on four molecular subtypes of endometrial carcinoma (EC): POLE mutation (POLE), microsatellite instability (MSI), copy number low (CNL), and copy number high (CNH).No statistically significant differences were observed in clinical indicators [TIME, CA125, CA199, hemoglobin (Hb), red blood cell count (RBC), platelet count (PLT), white blood cell count (WBC), and interleukin-9 (IL-9)] among the four molecular subtypes (all P > 0.05; Table 4). Differential expression patterns of key genes (identified via LASSO regression) were observed across molecular subtypes, with two genes showing statistically significant associations. The proportion of ERBB2-positive cases varied significantly among subtypes (P = 0.0151),the CNL subtype had the lowest ERBB2 positivity (19%), while the POLE subtype showed the highest (50%); VEGF positivity also differed significantly across subtypes (P = 0.0103). The CNL subtype had the lowest VEGF positivity (17.5%), whereas the MSI subtype showed the highest (46.2%). For other genes (PIK3CA, CCND1, KIT, PTEN, E-cadherin, p53), no statistically significant differences in expression distribution were observed across subtypes (all P > 0.05), though trends were noted (e.g., CCND1 positivity was highest in MSI (56.4%) and lowest in POLE (30%); Table 4). Prognostic outcomes (good vs. poor) showed a trend across subtypes, though not statistically significant (P = 0.2721). The CNL subtype had the highest proportion of good prognosis cases (73%), while the CNH subtype had the lowest (57.8%). This is consistent with prior observations that CNH is overrepresented in the poor prognosis group (57.53% vs. 35.66% in good prognosis group).

Factors demonstrating significant differences underwent collinearity diagnostics, revealing a variance inflation factor (VIF) > 5 for the gene PTEN and the clinical indicator RBC, indicating substantial multicollinearity. These variables were subsequently excluded from further analysis. Multivariate binary logistic regression analysis of the remaining factors identified Erb-B2 as a protective factor against poor prognosis in EC patients, while E-cadherin3 and p53 emerged as risk factors for adverse outcomes. Clinical markers CA125, CA199, and IL-9 were also significantly associated with increased risk of poor prognosis (Table 5).

To verify the clinical effectiveness of prognostic grouping, K-M survival analysis was conducted. The results showed that the 5-year survival rate of the good prognostic group (62.0%–100%) was significantly higher than that of the poor prognostic group (45.2%–100%). The Log rank test confirmed that there was a statistically significant difference in survival curves between the two groups, which proves that the initial prognostic grouping has clinical significance. (P < 0.05; Figure 5). Cox proportional hazards regression analysis incorporating factors identified by logistic regression revealed hazard ratios (HR) < 1 for Erb-B2 and p53, indicating protective effects, while E-cadherin exhibited an HR > 1, signifying increased risk. Clinical markers CA125, CA199, and IL-9 also showed HR > 1, correlating with elevated risk of adverse outcomes (Table 6; Figure 6).

A random forest model was constructed using the screened clinical indicators and genetic markers. Feature importance analysis revealed that CA199 exhibited relatively prominent positive importance in the model, suggesting its potential role as a key variable in correlation analyses or predictive modeling. Random forest feature importance ranking: CA199 (0.035)>IL-9 (0.028)>CA125 (0.022)>Erb-B2 (0.018); The decision tree is based on clinically detectable indicators to construct a hierarchical rule (Figure 7C) when CA199 > 56.8 U/mL (the clinical routine detection threshold is about 37 U/mL), the risk of poor prognosis is 2.3 times higher than below the threshold (68.7% vs. 29.4%); If both CA199 > 56.8 U/mL and CA125 > 51.2 U/mL are met, the risk of poor prognosis further increases to 72.3% (much higher than the overall poor prognosis rate of 36.1%) In contrast, E-cadherin, p53, and Erb-B2 displayed balanced but lower importance scores, indicating weaker contributions to outcome prediction. Directional importance analysis demonstrated variability in the magnitude and direction of variable impacts across events. CA199 showed higher positive importance for Event two (poor prognosis, represented by longer blue bars), whereas IL-9 exerted notable negative importance for Event 1 (favorable prognosis, indicated by longer red bars).

The decision tree diagram illustrated branching rules based on thresholds of CA125 and CA199 levels. Starting from the root node, data partitioning proceeded through sequential splits determined by these biomarkers, ultimately forming terminal leaf nodes for outcome classification. Interaction matrix analysis highlighted strong negative correlation (r = −0.473) between Erb-B2 and IL-9, reflecting their interconnected roles in prognostic stratification (Figure 7).