The present study was approved by the Ethics Committee of The First Affiliated Hospital of Zhengzhou University (no. KEYAN-2018-LW-023; Zhengzhou, China). The data analyzed in the program were identified between January 2005 and December 2014 from the SEER database (https://seer.cancer.gov/). Only patients diagnosed with Ewing sarcoma as the primary malignancy were included. Patients with tumors that were not first tumors (patients with no prior history of tumors or patients with tumors that were not primary tumors) were excluded from the present study. The tumor size data were collected with the exception of the codes 000/888/999/989-998 (‘000’ indicated no mass or no tumor found; ‘888’ and ‘999’ indicated not applicable or unknown; ‘989-998’ indicated 989 mm or larger but not accurately recorded). The tumor extension data were also collected with the exception of the codes 000/888/999/989-998. The exclusion criteria were as follows: Patients without biopsy diagnosis, patients with the N0M0 stage and patients who did not have Ewing sarcoma as their primary malignancy, patients with unknown surgical information, and patients with unknown marital status and ethnicity. Several studies have reported that marital status is an important prognostic factor for various tumor types (14,15). Therefore, in order to include marital status in the subsequent analysis, subjects with unknown marital status were removed. The patient selection is shown in the flow chart in Fig. 1. Patient characteristics and the original dataset are available in Tables SI and SII, respectively.

A cohort was formed for validation according to the AJCC 6 and 7th edition guidelines using the same exclusion criteria from the SEER database [this was a cohort with only AJCC stage as the independent variables and survival data as the dependent variable from the same SEER database using SEER*STAT (version 8.3.6; seer.cancer.gov/seerstat), and all patients with unknown AJCC stage were excluded from this cohort] (6). In addition, non-metastatic Ewing sarcoma cases between January 2013 and December 2018 at The First Affiliated Hospital of Zhengzhou University (Zhengzhou, China) were collected for external validation using the Electronic Medical Record System using the same inclusion and exclusion criteria. Despite the rarity of the disease, an external dataset comprising 26 cases was finally collected. Patient characteristics are shown in Table SIII.

The variables analyzed in the present study included the baseline demographic features of the patients (age at diagnosis, sex and ethnicity), the tumor characteristics [size, extension, primary site, International Classification of Diseases for Oncology, 3rd Edition (ICD.O.3) histological subtype and AJCC T stage] (6,16), the medical treatment (surgery, radiation recode and chemotherapy recode) and the socioeconomic variables (marital status, degree of education, family income and employment status). Age was divided into three subgroups, namely children (<15 years), adolescents and young adults (15–39 years), and old adults (>39 years) according to the NCCN Guidelines for Adolescent and Young Adult Oncology (17). The endpoint of the present study was patient death, which was presented as OS and CSS. The follow-up endpoint of all patients was December 31, 2014.

The categorical variables were expressed as percentages and the continuous variables were presented as the median (range). Subsequently, the continuous variables were transformed into categorical variables, such as age, tumor size and tumor extension. According to the NCCN guidelines (version 2017) (6), age (<15, 15–39 and >39 years), tumor size (<80 and ≥80 mm) and tumor extension (<300 and ≥300 mm) were divided into different subgroups. In addition, patients were divided into subgroups according to the median values for socioeconomic variables. The χ² test was adopted to assess the distinction between the sub-variables of each categorical variable. The same processes were performed based on OS and CSS. To compare the OS and CSS between subtypes, a K-M curve was used. For the K-M survival analysis, a log-rank test was used to compare the survival curves of each variable.

A random forest plot was established for entire variables. The Mean Decrease Gini (MDG) was used to estimate how each variable contributed to the OS and CSS. Additionally, the out-of-bag error was used to assess the accuracy of the classification in the random forest. The Cox proportional hazards model was constructed based on the results of the univariate analysis. The significant variables of the univariate analysis were screened to structure the subsets used as the data of the full Cox model. The non-significant chemotherapy effects were also retained for subsequent analysis. As a result, a reduced Cox model was constructed by sorting out the significant predictors from the full Cox model. Although the surgical information was not significant in the full Cox model, it was reserved for subsequent analysis in the reduced Cox model. Following double selections, the Cox model with the optimum predictors was eventually selected.

Based on the reduced Cox model, the nomograms were constructed for the evaluation of OS and CSS. The discriminations between predicted and observed values were evaluated via the concordance index (C-index) of internal validation. Their calibration was subsequently estimated using the corresponding calibration curve. In addition, to assess the accuracy of the final models, the external validation and concordance of the two AJCC editions was performed as the last step.

In the present study, R version 3.5.0 software (https://www.r-project.org/) was used to identify the statistically significant variables with a two-sided P<0.05 (α=0.05). The R packages ggplot2 (version 3.3.2; http://cloud.r-project.org/web/packages/ggplot2/index.html), survminer (version 0.4.8; http://cloud.r-project.org/web/packages/survminer/index.html), survival (version 3.2–7; http://cloud.r-project.org/web/packages/survival/index.html), rms (version 6.0-1; http://cloud.r-project.org/web/packages/rms/index.html) and randomForest (version 4.6–14; http://cloud.r-project.org/web/packages/randomForest/index.html) were used to establish the model and draw the curves, as well as the nomograms.

A total of 1,471 cases from the SEER database were selected between 2005 and 2014 as primary malignant cases. These cases were filtered, and an entire cohort containing 627 patients was included based on the inclusion and exclusion criteria. The patient characteristics are listed in Table SI, and the raw dataset of the entire primary cohort comprising 1,471 patients with non-metastatic Ewing sarcoma is shown in Table SII. At the end of the follow-up period, 502 patients were alive and 125 patients did not survive. The median age was 17 years (range, 0–85 years) and the highest proportion was noted in the 15–39-year-old subgroup. There were 252 women and 375 men. The median tumor size was 70 mm (range, 5–950 mm) and the median tumor extension was 400 mm (range, 100–820 mm). With regard to the primary tumor site, the distribution was as follows: Head/face/neck (8.0%), limbs (48.8%), thorax/abdomen (19.9%) and trunk (23.3%). The ICD.O.3 histology groups were adamantinoma of long bones (4.5%) and Ewing sarcoma (95.5%). In view of the limited number of AJCC T3 (0.8%) cases, the main comparison χ² test was only performed in the AJCC T1 and AJCC T2 cases (53.1 vs. 46.1%). AJCC T was the only variable to do this filtration in the χ² test, and T3 cases were not deleted in all subsequent analysis processes. Surgery was performed in 76.6% of patients, with chemotherapy recorded in 92.0% of the cases, whereas radiation was noted in 40.7% of the cases. The median OS was 40.00 months (range, 0.00–119.00 months).

The K-M survival analysis for OS and CSS is summarized in Table I. Patients who were >39 years exhibited the poorest OS and CSS, whereas patients whose age was <15 years exhibited the best prognosis (P<0.001, Fig. 2A; P<0.001, Fig. 2B; P-values apply to the log-rank tests among all categories, not the comparison between any two categories). Improved OS/CSS was achieved in patients who underwent surgery compared with patients who did not undergo surgery (P<0.001, Fig. 2C; P<0.001, Fig. 2D). In addition, the MDG values of the random forests were estimated for OS [out of bag (OOB)=21.37%] and CSS (OOB=19.62%), along with the outcomes of the parametric or the non-parametric tests (Table I). Variance homogeneous and normal distributed continuous variables could be compared by Student's t-test, otherwise, the Mann-Whitney U-test or Kruskal-Wallis H-test should be used. The Student's t-test had conditions of use. The c² test was used for discontinuous variables. In Table II, tumor features and medical treatments were significantly associated with OS and CSS, with the exception of the chemotherapy recode. Chemotherapy recode was also associated with the survival ending (the endpoint of the present study was patient death, which was presented as OS and CSS, and the follow-up endpoint of all patients was December 31, 2014). The parameters age, sex and marital status (data not shown) significantly influenced the prognosis of the patients. Finally, 11 elements (age, sex, tumor size, tumor extension, primary site, ICD.O.3 histology, AJCC T, surgery information, radiation recode, chemotherapy recode and marital status) were further analyzed using the Cox regression model.

In the Cox proportional hazards model, a higher hazard ratio (HR) was noted for the parameter of older age for OS [15–39 years: HR, 2.154; 95% confidence interval (CI), 1.375–3.375; P<0.001; >39 years: HR, 4.334; 95% CI, 2.417–7.770; P<0.001] and CSS (15–39 years: HR, 2.084; 95% CI, 1.316–3.300; P=0.002; >39 years: HR, 4.121; 95% CI, 2.248–7.552; P<0.001) compared with younger age (<15 years). Larger tumor extension was also a risk factor for OS (≥300.00 mm: HR, 2.404; 95% CI, 1.366–4.230; P=0.002) and CSS (≥300.00 mm: HR, 2.652; 95% CI, 1.465–4.803; P=0.001). In addition, male subjects appeared to have poorer OS (HR, 1.654; 95% CI, 1.124–2.434; P=0.011) and CSS (HR, 1.675; 95% CI, 1.120–2.506; P=0.012) than female subjects. The ICD.O.3 histological results for the higher risk subtype (Ewing sarcoma) indicated associations with OS (HR, 19.137; 95% CI, 2.466–148.492; P=0.005), but not with CSS (data not shown). The patients who had undergone surgery exhibited favorable OS (HR, 0.593; 95% CI, 0.407–0.865; P=0.007) and CSS (HR, 0.561; 95% CI, 0.378–0.831; P=0.004). The patients who were treated with chemotherapy exhibited improved OS (HR, 0.308; 95% CI, 0.164–0.581; P<0.001), whereas the effects of chemotherapy on CSS (HR, 0.780; 95% CI, 0.400–1.521; P=0.466) were not identified to be statistically significant.

Based on the reduced Cox models, nomograms were built to predict the probability of 3- and 5-year OS (Fig. 3A and C; C-index=0.791) and CSS (Fig. 3B and D; C-index=0.813). The calibration plots demonstrated the agreement between predictions and observed values due to the 45-degree-line nearby points (Fig. 3C and D). Information of the 26 cases for external validation is shown in Table SIII and the median age was 18 years (range, 1–51 years). The median survival time was 22 months (range, 3–54 months; Table SIII). The concordances of external validation based on the cohort of The First Affiliated Hospital of Zhengzhou University (Zhengzhou, China) were 0.834 for OS and 0.825 for CSS. Although the results of the K-M survival analysis indicated significant prognostic values of AJCC stage, the C-indices were 0.531 (OS) and 0.534 (CSS) for AJCC 6th edition guidelines, and 0.547 (OS) and 0.561 (CSS) for AJCC 7th edition guidelines, respectively, which were lower than the nomograms (Fig. S1). Additionally, the results of the K-M survival analysis indicated that chemotherapy could not significantly prolong the survival of the patients (Fig. S2). In the subgroup analysis, older patients were more likely to undergo chemotherapy rather than surgery, which might explain the poor prognosis in older patients (Table SIV).