A retrospective study was conducted of patients diagnosed with EC at the Department of Obstetrics and Gynecology, Peking University People's Hospital (PKUPH), from January 2006 to December 2020. The inclusion criteria were: i) Age over 18 years; ii) histologically confirmed diagnosis of EC; iii) undergoing total hysterectomy with either systematic lymphadenectomy or sentinel lymph node dissection (10); iv) complete clinical information and postoperative pathological data. The exclusion criteria were: i) Presence of additional malignant tumors; ii) lack of medical records; iii) preoperative treatment history; iv) other serious illnesses (such as stroke and heart disease); and v) death from other causes during follow-up. Based on these criteria, a total of 834 cases were selected for subsequent analysis. The present study was approved (approval no. 2022PHB379) by the Ethics Committee Board of Peking University People's Hospital (Beijing, China), in accordance with the principles outlined in the Declaration of Helsinki (2013). Informed consent was obtained from all subjects.

All patients underwent IHC examination, with approval from the Institutional Review Board of PKUPH for tissue excision. Pathological surgical specimens were fixed in 4% paraformaldehyde at 25°C for 48 h. After dehydration in a gradient of ethanol and clarification in xylene, the tissue samples were infiltrated with paraffin and embedded. The embedded tissue blocks were then sectioned into 5 µm slices using a microtome. The sections were incubated in a 60–65°C oven for 1–1.5 h, then deparaffinized in a xylene and ethanol gradient. Antigen retrieval was performed by incubating the sections in sodium citrate buffer at 95°C for 10 min, followed by the addition of an endogenous peroxidase blocker (cat. no. BF06060; Biodragon) to the tissue. The sections were then washed with PBST (including 0.1% Tween-20) for 3 min × 3 times. The primary antibody was applied to the tissue and incubated overnight at 4°C, followed by the addition of the secondary antibody and incubation at room temperature for 30 min. Detailed information about the antibodies has been added to the supplementary materials (Table SI). The sections' color was developed with 3,3′-diaminobenzidine (DAB), and the nuclei were stained with hematoxylin at 25°C for 15 min. Two pathologists independently assessed each sample in a blinded manner, without prior knowledge of the patients' details. IHC staining for estrogen receptor (ER), PR and P53 included both the percentage of positive nuclear staining, from 0–100%, and staining intensity, which was graded on a scale from 0 to 3. On this scale, 0 indicated negative, 1 indicated weak staining (+), 2 indicated moderate staining (++), and 3 indicated strong staining (+++). Ki67 was evaluated based solely on the percentage of positive nuclear staining. The representative IHC images are attached (Fig. S1, Fig. S2, Fig. S3, Fig. S4). In summary, the expression patterns of these four IHC markers in the patients were derived from the pathology reports and were reviewed and confirmed by two experienced pathologists.

To develop a model with high accuracy and stability, 10 machine learning algorithms were integrated and 92 algorithm combinations. The algorithms included RSF, elastic net, least absolute shrinkage and selection operator (Lasso), Ridge, StepCox, CoxBoost, partial least squares regression for Cox, supervised principal components, generalized boosted regression modeling, and survival support vector machine. One algorithm to filter the variables was utilized and another to build the prognostic signature. Out of 100 possible combinations of machine learning algorithm pairs, eight were excluded because the final prognostic signature included fewer than five genes. Leave-One-Out Cross-Validation (LOOCV) is well-known for providing an unbiased estimate and allowing comprehensive testing on each data point, ensuring the accuracy of the predictive model. The principle of this algorithm is as follows: One observation is selected as the test data, while all remaining observations are used as the training data. The model is then trained, and this process is repeated for each observation in the dataset. The test error is estimated by averaging the errors across all iterations.

The procedure for generating the signature was as follows: i) The collected patient demographic and pathological staining data were organized into numerical variables [age at diagnosis, BMI, ER percentage, PR percentage, P53 percentage, Ki67 percentage, overall survival (OS) time] and categorical variables [menopause status (premenopausal, postmenopausal), diabetes mellitus (without, with), hypertension (without, with), number of ER⁺ (0, 1, 2, 3), number of PR⁺ (0, 1, 2, 3), number of P53+ (0, 1, 2, 3), survival status (alive, deceased), ascites' cytology (negative, positive), histology [endometrioid endometrial adenocarcinoma (EEA), other types], LNM (negative, positive), lymph-vascular space invasion (negative, positive), myometrial invasion (<50%, ≥50%), cervical invasion (negative, positive), FIGO stage (I, II, III, IV) and grade (G1, G2, G3)]; ii) identifying those factors highly associated with prognosis through univariate Cox regression; iii) as previously mentioned, the combination of the 92 algorithms were utilized to construct predictive models for patients with EC; and iv) the Harrell concordance index (C-index) was computed, with the model exhibiting the highest average C-index being selected as the final model.

The χ² test was applied to compare categorical variables and the Wilcoxon rank-sum test or the unpaired t-test were used to assess continuous variables. Fisher's exact test was employed for the analysis of sample data with theoretical frequencies <5. The correlation between two continuous variables was evaluated using the Pearson correlation coefficient. The optimal cut-off value was determined with the survminer package. C-indices were compared using the Compare C package. Cox regression and Kaplan-Meier analyses followed by the log-rank test were conducted with the survival package. ROC analysis was performed with the pROC package, and the area under the curve (AUC) for survival variables was assessed using the time ROC package. All data analyses were conducted with R version 4.3.2 (http://www.R-project.org; The R Foundation) and EmpowerStats (http://www.empowerstats.com; X&Y Solutions, Inc.). A two-tailed significance level of P<0.05 was considered to indicate a statistically significant difference.

In the present risk model, a total of 834 patients with EC were randomly assigned to two groups: The training cohort (n=566) and the validation cohort (n=278), in a 2:1 ratio. The clinical baseline features and clinicopathological characteristics of patients are presented in Tables I and II. Based on the P-values obtained from the unpaired t-test and Fisher's exact test, the differences between the two groups were found to be statistically non-significant. Both cohorts predominantly consist of middle-aged and overweight patients, with mean ages of 56.49 and 55.85 years, and mean body mass indexes of 26.35 and 26.18 in the training and validation cohorts, respectively. In the training cohort, 364 patients (65.47%) are postmenopausal, while 192 patients (34.53%) are not. By contrast, the validation cohort includes 101 patients with premenopausal (36.33%) and 177 patients with non-premenopausal (63.67%). Most patients are staged as FIGO Stage I, comprising 77.34% (430/556) of the training cohort and 85.25% (237/278) of the validation cohort.

The present study analyzed 19 characteristic factors of patients with EC. Except for menopausal status, diabetes and hypertension, univariate Cox analysis revealed that the impact of the remaining factors on OS was statistically significant (Table III). Additionally, ROC curves were plotted for models incorporating four factors, three factors, and two factors, respectively (Figs. S5 and S6), demonstrating that the model including four IHC factors had the best predictive performance (AUC=0.951). A machine learning-based pathology-related model incorporating these 16 selected factors was developed.

In the EC dataset, 92 prediction models were applied using the LOOCV framework and the C-index for each model was calculated (Fig. 1A). The Lasso and stepwise Cox models were selected, which revealed the highest average C-index of 0.923. In Lasso regression, the optimal λ value was identified by minimizing the partial likelihood deviance using the LOOCV framework (Fig. 1B). Through stepwise Cox proportional hazards regression, a final set of 11 factors were determined from the original 16 factors (Fig. 1C). A risk score was calculated for each patient using the regression coefficients (Fig. 1D). The median risk score was used in each cohort as the threshold to stratify patients (Fig. 1E). As risk scores increased, survival time decreased, and the mortality rate increased (Fig. 1F). By combining multiple machine learning algorithms, the accuracy of the present study's predictive model has been significantly improved.

Kaplan-Meier plots and ROC curves were used to evaluate the relationship between risk scores and prognosis in patients with EC. The model demonstrated superior accuracy according to ROC analysis. In the training cohort, the AUC for predicting OS at 1, 3 and 5 years was 0.918, 0.893 and 0.853, respectively (Fig. 2A). In the validation cohort, the AUCs were 0.995, 0.757 and 0.719, respectively (Fig. 2C). Furthermore, the OS rate for the high-risk group was significantly lower compared with the low-risk group, with P=5.615×10⁻⁷ (Fig. 2B) for the training cohort and P=1.374×10⁻³ (Fig. 2D) for the validation cohort. The present study's model clearly assessed patient risk severity effectively and demonstrated strong predictive capability for recent events.

A similar approach was used to develop a prognostic model for RFS in patients with EC. The integration of Lasso and stepwise Cox methods demonstrated superior statistical power, achieving a C-index of 0.896 for the training group (Fig. 3A). After filtering out 11 factors (Fig. 3B and C), patients were categorized into high-risk and low-risk groups using the new risk calculation formula (Fig. 3D and E). Patients in the high-risk group exhibited a shorter time to recurrence, as demonstrated by a denser concentration of red dots in the lower right corner (Fig. 3F).

The present study's model demonstrated exceptional predictive capability, with AUC values exceeding 0.85 for both cohorts over a 5-year period. Notably, the training and validation cohorts achieved an AUC value of 0.99 and 0.925 at 1 year, respectively (Fig. 4A and C, respectively). In the training cohort, patients in the low-risk group had a significantly improved RFS compared with those in the high-risk group, with a P=9.41×10⁻⁸ (Fig. 4B). The validation cohort revealed similar outcomes, with a P=2.715×10⁻³ (Fig. 4D). These results indicated that the present study's model provides outstanding predictive performance for both OS and RFS.

To evaluate the enhanced predictive efficacy of incorporating IHC markers for OS and RFS in patients with EC, the IHC-related indicators were removed and the impact on curve values for both scenarios was assessed, pre- and post-exclusion. It was found that including these four factors, actually improved diagnostic accuracy, with one AUC value increasing from 0.765 to 0.872 (Fig. 5A) and another from 0.791 to 0.882 (Fig. 5B).

In summary, the model of the present study demonstrated robust predictive performance, demonstrating high accuracy and reliability in forecasting both OS and RFS across patients with EC. This predictive capability highlights its potential utility in clinical decision-making and personalized treatment planning.