Conditional survival and interpretable prognostic nomograms for primary chondrosarcoma based on the SEER database

Diagnosis-based survival estimates become progressively less representative as patients survive beyond the early high-risk period. We characterized conditional overall survival (OS) and cancer-specific survival (CSS) in primary chondrosarcoma and developed interpretable models for conventional and time-updated prediction. Eligible patients diagnosed during 2004–2023 were identified in the Surveillance, Epidemiology, and End Results (SEER) 17 Registries. Conditional survival was calculated as S(x + y)/S(x), and the cohort was divided 7:3 into training and internal testing cohorts. Cox regression, random survival forest (RSF), and CoxBoost were evaluated as feature-selection and survival-modeling methods in nine combinations. The nine combinations were assessed using the C-index, IBS, and time-dependent AUC. The selected Cox models were evaluated by calibration, decision-curve analysis, and comparison with a TNM-only Cox model. Among 2,442 patients, 669 died from any cause and 509 died from chondrosarcoma. Five-year OS and CSS at diagnosis were 76.0% and 79.5%, respectively. After survival to 1, 3, 5, and 10 years, 5-year conditional OS was 81.6%, 86.8%, 89.9%, and 92.4%, while conditional CSS was 85.1%, 90.7%, 94.6%, and 97.5%. Across nine modeling pipelines, CoxBoost feature selection followed by Cox regression provided the best balance between predictive performance, consistency across endpoints, and interpretability. In the internal testing cohort, the final OS model achieved a C-index of 0.8198, an IBS of 0.1105, and a mean AUC of 0.8440; the corresponding CSS values were 0.8304, 0.0949, and 0.8485. Both nomograms showed higher 3-, 5-, and 10-year AUCs than the TNM-only models, with close calibration and generally favorable decision-curve patterns. The nomograms estimated 3-, 5-, and 10-year survival and the probability of reaching 10 years after survival to 1, 3, or 5 years. Conditional survival in primary chondrosarcoma generally increased with accrued survival time. CoxBoost-selected Cox models provided interpretable diagnosis-time and conditional predictions and showed better descriptive internal performance than TNM-only models. Independent external validation is required before clinical implementation.