Development and verification of a competing risk model for forecasting cancer-specific survival in malignant bone tumor patients: an analysis of SEER database

Introduction

Malignant bone tumors are a group of rare tumors with highly heterogeneous clinical and biological characteristics, accounting for approximately 0.2% of all malignant tumors (1). According to 2023 US data, there were approximately 3,970 new cases of malignant bone tumors and 2,140 related deaths. Despite their low incidence, they have a poor prognosis and often cause severe functional impairment, significantly affecting patients’ quality of life (2). According to histological types, malignant bone tumors mainly include osteosarcoma, chondrosarcoma, and Ewing sarcoma. Among them, osteosarcoma is the most common primary malignant bone tumor (accounting for approximately 35%), commonly occurring in children and adolescents; chondrosarcoma mostly affects the elderly and is the second most common type; Ewing sarcoma primarily occurs in adolescents. Due to the rarity and heterogeneity of these tumors, epidemiological studies based on large sample populations are crucial for revealing their clinical characteristics and prognostic factors.

Take osteosarcoma as an example: before the 1970s, the 5-year survival rate of patients was less than 20%. With the application of multidisciplinary comprehensive treatments such as neoadjuvant chemotherapy combined with limb-sparing surgery, the survival rate has increased to approximately 70% (3,4). However, the prognosis remains unsatisfactory for patients with metastatic, recurrent, and chemotherapy-resistant tumors. Other types such as chondrosarcoma (insensitive to traditional radiotherapy and chemotherapy, mainly treated with surgical resection) and Ewing sarcoma (sensitive to radiotherapy and chemotherapy but with low survival rate in advanced stages) also face clinical challenges in diagnosis and treatment. Therefore, establishing an accurate prognostic prediction model is of great significance for guiding treatment decisions and improving patient survival.

Currently, the prognostic evaluation of malignant bone tumors mainly relies on clinical and pathological factors such as age, tumor size, stage, histological grade, and site. New markers such as inflammatory indicators, nutritional parameters, and tumor microenvironment characteristics have gradually shown predictive value. As an intuitive prediction tool, nomograms have been successfully applied in various cancers and can integrate multi-factor information to quantify individual risks (5-7). However, most existing prediction models are based on traditional Cox proportional hazards regression, which does not consider the impact of competing risk events (such as non-tumor causes of death), potentially overestimating tumor-specific mortality risk.

The competing risk model (proposed by Fine and Gray) is specifically used to analyze the risk probabilities of multiple outcome events and can more accurately evaluate the cumulative incidence of specific events. In cancer research, ignoring competing risks may lead to biased estimation of the effects of prognostic factors and affect model accuracy (8,9). The US Surveillance, Epidemiology, and End Results (SEER) database covers approximately 48% of the US population, providing large-scale high-quality data for research on rare tumors such as malignant bone tumors. Constructing a competing risk model based on this database can more realistically reflect patients’ survival status and provide reliable support for clinical decisions.

This study aimed to construct and validate a competing risk nomogram using the Fine-Gray model based on the SEER database for malignant bone tumor patients. Compared with the conventional Cox-Wilson model, the Fine-Gray model addresses the limitation of ignoring non-tumor deaths (competing risks), which often leads to overestimated tumor-specific mortality. By clarifying why (to eliminate competing risk bias) and how (via subdistribution hazard modeling) the proposed model advances prognostic accuracy, this study provides a practical prognostic tool that is accessible to non-statistical biologists and supports individualized treatment decision-making. We present this article in accordance with the TRIPOD reporting checklist (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-1886/rc).

Methods

Data source and case selection

The sample size of this study was determined based on the number of complete cases of malignant bone tumors (osteosarcoma, chondrosarcoma, Ewing’s sarcoma) available in the SEER database from 2000 to 2022. Data were extracted using ICD-O-3 (International Classification of Diseases for Oncology, 3rd Edition) morphology codes. The inclusion criteria were pathologically confirmed cases, excluding those with unknown tumor site, unknown TNM (tumor, node, metastasis) stage, and autopsy diagnosis. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments, and ethical approval was not required due to the anonymized processing of the data.

Variable selection

The extracted variables included age, gender, race, primary tumor site, tumor size, clinical stage, surgery, TNM stage, radiotherapy, and chemotherapy. Race was divided into Asian/Pacific Islander, Black, White, and other; tumor sites were divided into axial bone and appendicular bone; tumor size was divided into <80 mm, 80–150 mm, >150 mm, and unknown; clinical stages were divided into Local, Regional, Distant, and unknown; surgical methods included local resection, radical resection, amputation, and no surgery/unknown; radiotherapy and chemotherapy were divided into “yes” or “no”. The outcome indicators were cancer-specific mortality (CSM) and other causes of mortality (OCM), with survival time was defined as the time elapsed between diagnosis and either death or the most recent follow-up.

Model development and verification

The training set and validation set were randomly divided at a ratio of 7:3. Univariate Cox regression and bidirectional stepwise regression were used to screen variables affecting survival, for stepwise bidirectional regression, the variable entry criterion is set at P<0.05, and the variable removal criterion is P>0.10. In univariate analysis, variables with P<0.20 are included in the multivariate screening process. Finally, the predictors incorporated into the multivariate model must meet the requirement of P<0.05 and multivariate Cox regression was used to determine independent risk factors for CSM and OCM (P<0.05 as significant). A competing risk model was constructed based on independent risk factors for CSM to predict 1-, 3-, and 5-year CSS rates, and a nomogram was drawn. The discrimination ability was evaluated using area under the receiver operating characteristic curve (AUC) and C-index, and the accuracy was verified using calibration curves.

Clinical application evaluation

The clinical guidance value of the model was quantified through decision curve analysis (DCA), and compared with the TNM staging system. Based on risk scores, the optimal cut-off value (maximum Youden index) was determined by receiver operating characteristic (ROC) curve to divide patients were categorized into high- and low-risk groups, with the efficiency of risk stratification verified via cumulative incidence functions.

Statistical analysis

Measurement data were described as mean ± standard deviation (x±s), and group comparisons were performed using t-tests; count data were expressed as frequency (n) and proportion (%), and group comparisons were performed using Pearson’s chi-square test. Competing risk modeling was performed using the Fine & Gray proportional subdistribution hazard model, and differences in survival outcomes between groups were compared using Gray’s test. Statistical analysis was performed using R software (version 4.3.2), with a significance level of α=0.05. R software packages “cmprsk”, “survival”, “forestploter”, “survivalRO”, “pec”, “ggDCA”, and “riskRegression” were used in this study.

Results

Patient characteristics

A total of 3,508 patients were included (Table 1), with 2,455 in the training set and 1,053 in the validation set. The mean age was 38.3±23.3 years, 57.0% (2,000 cases) were male, 81.0% (2,840 cases) were White, and 69.1% (2,424 cases) had osteosarcoma. Tumors located in the appendicular bones accounted for 57.7%, tumors with a diameter <80 mm accounted for 40.3%, and clinical stage local accounted for 40.5% (1,422 cases). Radical surgery was performed in 42.8% (1,052 cases), and TNM stages were mainly T2 (39.1%), N0 (89.9%), and M0 (80.5%). Radiotherapy was administered in 20.6% (721 cases), and chemotherapy in 53.0% (1,858 cases). There were 964 (27.5%) CSM cases and 580 (16.5%) OCM cases. The mean overall survival time was 91.6±64.8 months, with a median survival time of 89 months. No significant differences were observed in baseline characteristics between the training set and the validation set.

Risk factor analysis

Univariate and multivariate Cox regression showed that age, gender, T3–4 stage, M stage, tumor size, primary site, Ewing sarcoma, clinical stage, surgery, radiotherapy, and chemotherapy were independent risk factors for CSM. Gender, T3–4 stage, and radiotherapy had no significant effect on OCM (Figures 1,2).

Figure 1 Multivariate Cox regression analysis of CSM. *, P<0.05; **, P<0.01; ***, P<0.001. CI, confidence interval; CSM, cancer-specific mortality; HR, hazard ratio; M, metastasis; N, node; T, tumor.

Figure 2 Multivariate Cox regression analysis of OCM. *, P<0.05; **, P<0.01; ***, P<0.001. CI, confidence interval; HR, hazard ratio; M, metastasis; N, node; OCM, other causes of mortality; T, tumor.

Model construction and validation

A competing risk model was constructed based on independent risk factors, and a nomogram was drawn (Figure 3). The sum of risk scores corresponding to each variable was associated with mortality. For the training set and validation set, the C-indexes were 0.77 [95% confidence interval (CI): 0.75–0.79] and 0.79 (95% CI: 0.77–0.82) in sequence, with AUC >0.8 in both (Figure 4). The calibration curve showed a high degree of consistency between predicted and actual survival rates (Figure 5).

Figure 3 Competing risk model nomogram for forecasting 1-, 3-, and 5-year cancer-specific survival in patients with malignant bone tumors. *, P<0.05; **, P<0.01; ***, P<0.001. M, metastasis; N, node; T, tumor.

Figure 4 The AUC of the predictive model. (A) The 1-, 3-, and 5-year cancer-specific survival AUC predicted by the nomogram in the training set; (B) the corresponding AUC in the validation set. AUC, area under the receiver operating characteristic curve.

Figure 5 The calibration curve of the prediction model. (A) In the training set, the calibration curves for predicting 1-, 3- and 5-year cancer-specific survival were highly consistent with the actual values; (B) in the validation, the calibration curves for predicting 1-, 3- and 5-year cancer-specific survival were highly consistent with the actual values.

Clinical application efficiency

DCA showed that the nomogram had a higher clinical net benefit at all time points than the TNM staging system (Figure 6). A risk threshold of 1.84 was determined by ROC to split patients into high-risk (total score ≥1.84) and low-risk (total score <1.84) groups (Figure 7). The cumulative incidence curve showed a significantly higher CSM rate in the high-risk group (P<0.001, Figure 8).

Figure 6 DCA of the prediction model. (A) DCA for forecasting 1-, 3-, and 5-year cancer-specific survival in the training set; (B) DCA for forecasting the same time points of cancer-specific survival in the validation set. DCA, decision curve analysis; TNM, tumor-node-metastasis.

Figure 7 The optimal cut-off value (maximum Youden index) was determined by ROC curve to divide patients into high- and low-risk groups. ROC, receiver operating characteristic.

Figure 8 Cumulative incidence curves. (A) In the training set, the cumulative incidence curve revealed that cancer-specific mortality was significantly higher in the high-risk group than in the low-risk group; (B) in the validation set, the cumulative incidence curve showed the same trend of significantly higher cancer-specific mortality in the high-risk group compared to the low-risk group.

Discussion

This study constructed a competing risk model for CSS in patients with malignant bone tumors based on the SEER database, integrating multi-dimensional prognostic factors and verifying its clinical value. Traditional Cox regression may lead to biased risk assessment due to ignoring OCM. This study is the first in this field to include both CSM and OCM outcomes, which is closer to clinical reality (8,9). The incidences of OCM in the training set and validation set were 17.3% and 14.7%, respectively, indicating that non-tumor deaths cannot be ignored. By using the subdistribution hazard model to isolate the independent influencing factors of CSM, it avoids misjudgment of risks by traditional methods and improves prediction accuracy.

The study found that age increase was a common risk factor for both CSM and OCM. Patients aged ≥50 years had a 3.2-fold higher CSM risk than those <25 years [hazard ratio (HR) =3.2], consistent with previous studies (2,10,11). This may be related to the pro-inflammatory characteristics of the tumor microenvironment in the elderly, decreased DNA damage repair capacity, and abnormal activation of the p53 pathway (12-14). The CSM risk in males was significantly higher than in females (HR =1.2), possibly due to anatomical differences (such as long bone volume) and hormone levels (androgens promote tumor proliferation, while estrogens inhibit tumor growth) (15,16).

Tumors originating from the axial bone had a 20% higher CSM risk than those from the appendicular bones (HR =1.2), possibly due to the difficulty of surgical resection (R0 resection rate 55% in axial bone vs. 78% in appendicular bone) and high early micrometastasis rate (17). Tumor size was strongly correlated with prognosis (HR =1.9 in the >150 mm group), reflecting tumor proliferation activity and angiogenesis capacity [such as high vascular endothelial growth factor (VEGF) expression and hypoxic microenvironment] (18-20).

Treatment methods significantly affected survival: the CSM risk in patients who did not undergo surgery was 1.6 times higher than that in the local resection group, indicating the core role of surgery; the risk in amputated patients was higher (HR =1.6), possibly due to extensive tumor invasion, requiring preoperative precise evaluation. Chemotherapy was positively correlated with CSM risk (HR =2.1), possibly because advanced patients needed more chemotherapy rather than chemotherapy toxicity itself, and the same applied to radiotherapy. It should be noted that the SEER database uses the same code for “not received” and “unknown” radiotherapy and chemotherapy, which may introduce bias.

In terms of pathological types, the CSM risk of Ewing sarcoma was lower than that of osteosarcoma (HR =0.7), related to the enhanced immunogenicity of the EWS-FLI1 fusion gene (21,22); chondrosarcoma has no driver mutations, and its prognosis depends on the thoroughness of surgery. In this study, there was no significant difference in survival between chondrosarcoma and osteosarcoma.

The model performance was significantly better than the TNM staging system (C-index 0.77–0.79, AUC >0.8, DCA net benefit increased by 30–40%), with high calibration curve fitting and significant risk stratification efficiency (5-year CSM rate 65% in high-risk group vs. 22% in low-risk group). The nomogram integrates multi-dimensional factors. For example, a <25-year-old male with appendicular primary tumor <80 mm, T0–1N0M0 stage, and treated with surgery combined with radiotherapy and chemotherapy for Ewing sarcoma, the nomogram predicts a 5-year CSS probability of approximately 12.2%, suggesting the need for intensified palliative treatment; low-risk patients can extend the follow-up period to optimize medical resource allocation.

The competing risk nomogram developed in this study provides a practical and quantifiable tool for individualized prognosis assessment. In clinical practice, by summing the scores corresponding to each patient’s characteristics (e.g., age, tumor size, pathological type, treatment modalities) on the nomogram, clinicians can readily obtain the predicted probability of 1-, 3-, and 5-year CSS. This facilitates several key applications: (I) risk stratification: identifying high-risk patients (e.g., those with a total nomogram score ≥1.84) who may benefit from more intensive treatment regimens, closer follow-up schedules, or inclusion in clinical trials for novel therapies. Conversely, low-risk patients might be candidates for de-escalation strategies to avoid overtreatment and reduce toxicity; (II) patient counseling: offering a visual and numerical estimate of personalized survival probability can aid in shared decision-making and setting realistic expectations for patients and their families; (III) resource allocation: the model can help institutions prioritize resources and monitoring efforts towards patients with the poorest predicted outcomes.

The innovations of this study include: (I) integrating three common pathological types (osteosarcoma, Ewing sarcoma, chondrosarcoma, accounting for 69.1%, 19.9%, and 11.0%), comprehensively reflecting the disease spectrum; (II) separating independent factors of CSM and OCM through competing risk models, excluding interference from non-tumor deaths, and providing a quantitative basis for personalized treatment.

The limitations of this study should be acknowledged. First, as a retrospective analysis based on the SEER database, inherent confounding biases exist (e.g., unrecorded confounding factors such as comorbidities and treatment adherence), which may affect the causal inference of risk factors. Second, the SEER database lacks detailed information on molecular markers, inflammatory indicators, and surgical margins (e.g., R0 resection status), limiting the fine-grained evaluation of individual prognosis. Third, the lack of external cohort validation from non-SEER populations may reduce the generalizability of the model to regions with different ethnicities or medical practices. In the future, prospective multicenter studies integrating molecular biomarkers and dynamic clinical indicators are needed to further optimize the model.

Conclusions

The competing risk model for malignant bone tumors constructed in this study can accurately predict patients’ CSS probability, outperforming traditional staging systems, and is helpful for the development of individualized treatment plans, identification of high-risk patients, and clinical research on new therapies. In the future, the model needs to be improved through multicenter data and biomarkers to promote the development of precision medicine.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-1886/rc

Peer Review File: Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-1886/prf

Funding: None.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-1886/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Choi JH, Ro JY. The 2020 WHO Classification of Tumors of Bone: An Updated Review. Adv Anat Pathol 2021;28:119-38. [Crossref] [PubMed]
2. Xu Y, Shi F, Zhang Y, et al. Twenty-year outcome of prevalence, incidence, mortality and survival rate in patients with malignant bone tumors. Int J Cancer 2024;154:226-40. [Crossref] [PubMed]
3. Hiraga H, Machida R, Kawai A, et al. Methotrexate, Doxorubicin, and Cisplatin Versus Methotrexate, Doxorubicin, and Cisplatin + Ifosfamide in Poor Responders to Preoperative Chemotherapy for Newly Diagnosed High-Grade Osteosarcoma (JCOG0905): A Multicenter, Open-Label, Randomized Trial. J Clin Oncol 2025;43:1886-97. [Crossref] [PubMed]
4. Ogura K, Morizane C, Satake T, et al. Statistics of bone sarcoma in Japan: report from the population-based cancer registry in Japan. Int J Clin Oncol 2024;29:1209-19. [Crossref] [PubMed]
5. Zheng S, He Y, Chen Y, et al. A population-based study using nomograms to predict overall and cancer-specific survival in HPV-associated CSCC. Cancer Sci 2025;116:470-87. [Crossref] [PubMed]
6. Wang J, Zhanghuang C, Tan X, et al. A Nomogram for Predicting Cancer-Specific Survival of Osteosarcoma and Ewing's Sarcoma in Children: A SEER Database Analysis. Front Public Health 2022;10:837506. [Crossref] [PubMed]
7. Chen B, Zeng Y, Liu B, et al. Risk Factors, Prognostic Factors, and Nomograms for Distant Metastasis in Patients With Newly Diagnosed Osteosarcoma: A Population-Based Study. Front Endocrinol (Lausanne) 2021;12:672024. [Crossref] [PubMed]
8. Austin PC, Lee DS, Fine JP. Introduction to the Analysis of Survival Data in the Presence of Competing Risks. Circulation 2016;133:601-9. [Crossref] [PubMed]
9. Wolbers M, Koller MT, Stel VS, et al. Competing risks analyses: objectives and approaches. Eur Heart J 2014;35:2936-41. [Crossref] [PubMed]
10. Wu X, Wang J, He D. Establishment and validation of a competitive risk model for predicting cancer-specific survival in patients with osteosarcoma: a population-based study. J Cancer Res Clin Oncol 2023;149:15383-94. [Crossref] [PubMed]
11. Beauchamp C. Chondrosarcoma in the United States (1973 to 2003): An Analysis of 2890 Cases from the SEER Database. Yearbook of Orthopedics 2010;2010:212-3. [Crossref]
12. Feng Z, Hu W, Teresky AK, et al. Declining p53 function in the aging process: a possible mechanism for the increased tumor incidence in older populations. Proc Natl Acad Sci U S A 2007;104:16633-8. [Crossref] [PubMed]
13. Hientz K, Mohr A, Bhakta-Guha D, et al. The role of p53 in cancer drug resistance and targeted chemotherapy. Oncotarget 2017;8:8921-46. [Crossref] [PubMed]
14. Raj N, Attardi LD. Tumor suppression: p53 alters immune surveillance to restrain liver cancer. Curr Biol 2013;23:R527-30. [Crossref] [PubMed]
15. Yang ZM, Yang MF, Yu W, et al. Molecular mechanisms of estrogen receptor β-induced apoptosis and autophagy in tumors: implication for treating osteosarcoma. J Int Med Res 2019;47:4644-55. [Crossref] [PubMed]
16. Zhang C, Mao M, Guo X, et al. Nomogram based on homogeneous and heterogeneous associated factors for predicting bone metastases in patients with different histological types of lung cancer. BMC Cancer 2019;19:238. [Crossref] [PubMed]
17. Chantrain CF, Feron O, Marbaix E, et al. Bone marrow microenvironment and tumor progression. Cancer Microenviron 2008;1:23-35. [Crossref] [PubMed]
18. Ortiz MA, Anguera G, Cantó E, et al. Pilot Investigation on Markers of Bone Metabolism, Angiogenesis, and Neuroendocrine Activity as Potential Predictors of Survival of Metastatic Prostate Cancer Patients with Bone Metastases. Int J Mol Sci 2025;26:4669. [Crossref] [PubMed]
19. Weis SM, Cheresh DA. Tumor angiogenesis: molecular pathways and therapeutic targets. Nat Med 2011;17:1359-70. [Crossref] [PubMed]
20. Roy S, Kumaravel S, Sharma A, et al. Hypoxic tumor microenvironment: Implications for cancer therapy. Exp Biol Med (Maywood) 2020;245:1073-86. [Crossref] [PubMed]
21. Godfrey LC, Regalado B, Schweber SR, et al. p300/CBP is an essential driver of pathogenic enhancer activity and gene expression in Ewing sarcoma. EMBO Rep 2025;26:4766-93. [Crossref] [PubMed]
22. Elzi DJ, Song M, Houghton PJ, et al. The role of FLI-1-EWS, a fusion gene reciprocal to EWS-FLI-1, in Ewing sarcoma. Genes Cancer 2015;6:452-61. [Crossref] [PubMed]