Our research was based on data from public databases and retrospective hospital records. The SEER database is a publicly accessible, authoritative tumor registry widely used in clinical research, containing information on millions of cancer patients in the USA, including age at onset, age at death, primary tumor site, surgical details, treatment information, and demographic characteristics. In this study, we extracted patient data from the SEER database using SEER*Stat software (version 8.4.0.1; http://seer.cancer.gov/), account number 15006-Nov2021, for cases diagnosed between 1 January 2010 and 31 December 2019. Additionally, we obtained an external validation set from Tongji Hospital, Huazhong University of Science and Technology (THAHUST), using the hospital’s electronic medical record system, for cases diagnosed between 1 January 2009 and 31 December 2018.

We retrieved a total of 1,626 available cases from the SEER database. We also identified 90 cases for an external validation set in THAHUST’s electronic medical record system. Patient data were collected based on the following criteria: (1) download data from the SEER database according to ICD-O-3.2 morphological code, including ICD-O-3 8714/3: perivascular epithelioid cell tumors, ICD-O-38896/3: mycinous leiomyosarcoma, ICD-O-3 8805/3: undifferentiated uterine sarcoma, ICD-O-38900/3: rhabdomyosarcoma, ICD-O-3 8890/3: leomyosarcoma, ICD-O-3 8891/3: epithelial leiomyosarcoma, ICD-O-3 8930/3: high-grade endometrial stromal sarcoma, ICD-O-3 8931/3: low-grade endometrial stromal sarcoma, ICD-O-3 8933/3: rhabdomyosarcoma. It should be noted that carcinosarcoma should not be included.

According to the WHO Classification of Tumors of the Female Genital Organs (fourth edition, 2014), carcinosarcoma is clearly classified as “metaplastic carcinoma” and placed under the category of endometrial cancer rather than mesenchymal tumors (17). (2) All patients were identified by pathological examination as first primary tumors, (3) underwent surgical resection, (4) had complete postoperative pathological data and general clinical characteristics, and (5) had complete follow-up data. The exclusion criteria were (1) not being the first or only primary malignancies and (2) patients with insufficient information.

Patient survival information was quantified using OS, which is defined as the time from the surgical pathological diagnosis to the date of last follow-up or death from any cause. The clinicopathological factors we collected, which may be related to patient survival, included age at diagnosis, FIGO stage, radiotherapy, chemotherapy, race, marital status, pathology, tumor grade, peritoneal cytology, pelvic lymph node log odds of positive lymph node (LODDS) grade, paraaortic lymph node LODDS grade, and tumor size. For constructing and validating the nomogram, the 1,626 patients collected from the SEER database were randomly assigned as the training set (n = 1,138) and internal validation set (n = 488) in a 7:3 split ratio, while a total of 90 cases collected from THAHUST served as an external validation set. Continuous variables, including age at diagnosis, tumor size, and lymph node LODDS grade, were analyzed using R Studio 4.2.2 (http://www.rstudio.com/) with the “survivalROC” and “survminer” packages to obtain optimal cut-off values in the training set. Categorical variables were presented as numbers with percentages. The demographic and clinical characteristic distributions between the training and validation sets were compared using the Chi-square test with SPSS. Modeling variables were obtained by univariate and multivariable Cox regression, as well as Lasso regression, and prognostic models based on the identified independent prognostic factors.

We calculated the consistency index (C-index), the area under the time-dependent receiver operating characteristic (ROC) curve (AUC) to evaluate the predictive accuracy of the nomogram. The predictive accuracy (discrimination) of a nomogram is measured by the C-index, which quantifies the level of agreement between the predicted probability and the actual probability of an event of interest occurring (18).

A C-index and AUC value above 0.90 indicate high predictive power, while values between 0.71 and 0.90 suggest moderate discrimination. We plotted a calibration curve to evaluate the model and assess potential overfitting using 1,000 bootstrap resamples. The DCA provides a clear answer to which model will bring the most significant clinical benefit, on average, to the right patient, comparing the constructed nomogram with the net clinical benefit of FIGO staging by the DCA.

Finally, the total score of each patient was calculated based on the variable scores in the nomogram. The patients were stratified into a high-risk group, a medium-risk group, and a low-risk group based on the 5-year survival probability. The Kaplan–Meier curve for OS was plotted according to risk grouping, and log-rank tests were conducted in the training and validation sets. The research process is shown in Figure 1 .

We got the optimal cut-off values based on the training set in Table 1 .

Collated clinical and demographic baseline characteristics for training sets, internal validation sets, and external validation sets are presented in Tables 2 , 3 . In the data from the SEER database, the median age at diagnosis was 55 years (quartile, 48–64 years). About half of the patients were diagnosed with stage I (55.54%).

Leiomyosarcoma has the highest incidence (46.49%), followed by endometrial stromal sarcoma (35.73%). A total of 40.9% of the patients received postoperative chemotherapy, and only 15.93% received radiotherapy. Only 4.37% of patients were positive for peritoneal cytology. In THAHUST, the median age at diagnosis was 47 years (quartile, 39–55.25 years old), and it had a longer median survival time (SEER: 34.5 vs. THAHUST: 56.5). THAHUST data also showed a higher proportion of patients receiving postoperative adjuvant therapy.

We first explored the optimal variables through Lasso regression, which identified four variables: age at diagnosis, FIGO stage, grade, and tumor size. The results of the Lasso regression are presented in Figure 2 .

According to univariate Cox regression analysis, eight variables were identified with a p-value < 0.05, including age at diagnosis, FIGO stage, chemotherapy, race, grade, peritoneal cytology, tumor size, and pathology. According to the multivariable Cox regression analysis, age at diagnosis, FIGO stage, race, grade, peritoneal cytology, pathology, and tumor size were independent prognostic factors. Age at diagnosis (hazard ratio (HR) = 1.625 vs. age younger than 52 years, p < 0.001), FIGO stage (stage II, HR = 1.93 vs. FIGO stage I, p < 0.001; stage III, HR = 1.899, p = 0.024; stage IV, HR = 3.463, p < 0.001), race being Black (HR = 1.513, p < 0.001), tumor grade (grade 3, HR = 2.903 vs. grade 1, p < 0.001; grade 4, HR = 2.855, p < 0.001). Consistent with other studies, Black race was an independent risk factor for uterine sarcoma (16). The results of univariate and multivariable Cox regression analyses are presented in Table 4 .

In a comparison of the results of Cox regression and Lasso regression, age at diagnosis, FIGO stage, grade, peritoneal cytology, and tumor size were used to construct the nomogram for the following reasons: (1) we aimed to establish a model applicable to Asians, and Asian descent was a protective factor (HR = 0.874 [95% CI: 0.637, 1.199] vs. White, p < 0.001), with race excluded; (2) the AUC value of the model with added pathology did not show significant improvement.

Figure 3 shows a simple and efficient nomogram for predicting 1-, 3-, and 5-year OS probability. Factors of age at diagnosis, FIGO stage, grade, tumor size, and peritoneal cytology were used to construct a nomogram based on Lasso regression results. The tumor grade was set as a reference scale ranging from 0 to 100 because it had the largest absolute coefficient value. By adding the values above each variable, the total score was obtained. Drawing a vertical line down from the total score, the patient’s probability of survival at 1, 3, and 5 years can be obtained.

The C-indices in the training set, the internal validation set, and the external validation set were 0.78 [95% CI: 0.739, 0.823], 0.77 [95% CI: 0.697, 0.8255], and 0.84 [95% CI: 0.670, 0.928], respectively. The 1-, 3-, and 5-year ROC curves and AUC values were used to judge the discrimination ability of the model. In the training set ( Figure 4A ), the AUC values were 0.826, 0.847, and 0.839, respectively. The AUC values were similar in the internal validation set ( Figure 4B ). In the external validation set ( Figure 4C ), the AUC values were 0.905, 0.890, and 0.876. AUC values above 0.8 indicate strong predictive power; the external validation AUC peaked at 0.9, demonstrating excellent generalizability. Again, the calibration plots ( Figures 4D–F ) were close to the ideal 45° reference line, and the predicted values and actual results were similar.

DCA was a new method of evaluating alternative prognostic strategies that provided advantages over AUC. We compared the net clinical benefits of the model with FIGO staging, and the model showed clear advantages, especially in the external validation set ( Figure 5 ).

An appropriate threshold is set according to the patient’s 5-year OS as an inflection point of clinical treatment strategy (9). In general, an overall 5-year relative survival > 65% is good, 35%–65% is moderate, and less than 35% is very poor (19). Patients with uterine sarcoma with good 5-year OS do not always benefit from adjuvant therapy after surgery (1). In this model, a risk-averse cut-off point of 40%/80% was used for treatment decisions based on the patient’s 5-year OS. Within the clinical prognostic nomogram we constructed, each of the five predictors was assigned a standardized point score from 0 to 100, quantifying its relative contribution to the predicted outcome. The specific points for each variable, derived from the statistical algorithm using the R software, are detailed in Table 5 .

By aggregating the points from all variables, a comprehensive risk score (total points) was derived for each patient in the cohort. The results are shown in Table 6 .

All patients were divided into a high-risk group, a medium-risk group, and a low-risk group based on the tiered scores 55 and 117. The Kaplan–Meier curves and the log-rank test showed that patients with a lower total score of 55 had a higher 5-year survival probability of more than 80%, and by comparing the actual postoperative adjuvant treatment of these patients, we found that low-risk group patients did not require postoperative adjuvant therapy. Patients with a score of 55 to 117 were recommended for postoperative adjuvant treatment based on the patient’s willingness to treat. High-risk patients must undergo postoperative adjuvant therapy (p < 0.0001). The results are shown in Figure 6 .